[HN Gopher] AMD's Ryzen 9 9950X3D2 Dual Edition crams 208MB of c...
       ___________________________________________________________________
        
       AMD's Ryzen 9 9950X3D2 Dual Edition crams 208MB of cache into a
       single chip
        
       Author : zdw
       Score  : 291 points
       Date   : 2026-03-28 02:17 UTC (20 hours ago)
        
 (HTM) web link (arstechnica.com)
 (TXT) w3m dump (arstechnica.com)
        
       | Readerium wrote:
       | Can someone explain if the 3D Vcache are stacked on top of each
       | other or side by side.
       | 
       | If they are stacked then why not 9800X3D2?
        
         | zdw wrote:
         | The 99xx chips have two CPU dies, and one cache die is on each
         | CPU die.
        
           | modeswitch wrote:
           | The 3D V-Cache sits underneath only one of the CCDs. See
           | https://en.wikipedia.org/wiki/Ryzen#Ryzen_9000.
        
             | Tostino wrote:
             | Did you forget which thread we are on?
        
             | anonymars wrote:
             | That's what's different about this one. "Enter the Ryzen 9
             | 9950X3D2 Dual Edition, a mouthful of a chip that includes
             | 64MB of 3D V-Cache on both processor dies, without the
             | hybrid arrangement that has defined the other chips up
             | until now."
        
       | chao- wrote:
       | Crazy to think that my first personal computer's entire storage
       | (was 160MB IIRC?) could fit into the L3 of a single consumer CPU!
       | 
       | It's probably not possible architecturally, but it would be
       | amusing to see an entire early 90's OS running entirely in the
       | CPU's cache.
        
         | bombcar wrote:
         | IIRC some relatively strange CPUs could run with unbacked
         | cache.
        
           | twbarr wrote:
           | Intel's platform, at the very least, use cache-as-ram during
           | the boot phase before the DDR interface can be trained and
           | started up. https://github.com/coreboot/coreboot/blob/main/sr
           | c/soc/intel...
        
         | cwzwarich wrote:
         | https://github.com/coreboot/coreboot/blob/main/src/soc/intel...
        
           | wmf wrote:
           | Context: Early in the firmware boot process the memory
           | controller isn't configured yet so the firmware uses the
           | cache as RAM. In this mode cache lines are never evicted
           | since there's no memory to evict them to.
        
             | lathiat wrote:
             | I remember the talk about the Wii/WiiU hacking they
             | intentionally kept the early boot code in cache so that the
             | memory couldn't be sniffed or modified on the ram bus which
             | was external to the CPU and thus glitchable.
        
             | coppsilgold wrote:
             | There may be server workloads for which the L3 cache is
             | sufficient, would be interesting if it made sense to create
             | boards for just the CPU and no memory at scale.
             | 
             | I imagine for such a workload you can always solder a small
             | memory chip to avoid having to waste L3 on unused memory
             | and a non-standard booting process so probably not.
        
               | stingraycharles wrote:
               | Most definitely, I work in finance and optimizing
               | workloads to fit entirely in cache (and not use any
               | memory allocations after initialization) is the de-facto
               | standard of writing high perf / low latency code.
               | 
               | Lots of optimizations happening to make a trading model
               | as small as possible.
        
         | pwg wrote:
         | In my case it began with 16K (yes, 16 _1024 bytes) and 90K
         | (yes, 90_ 1024 bytes) 5.25" floppy disks (although the floppies
         | were a few months after the computer). Eventually upgraded to
         | 48K RAM and 180K double density floppy disks. The computer:
         | Atari 800.
        
           | MegaDeKay wrote:
           | I'll see your Atari 800 and raise you my Atari 2600 with its
           | whopping _128 bytes_ of RAM. Bytes with a B. I can kinda
           | sorta call it a computer because you _could_ buy a BASIC
           | cartridge for it (I didn 't and stand by that decision - it
           | was pretty bad).
        
             | acomjean wrote:
             | I thought the timex Sinclair 1000 win 2 Kbytes of ram was
             | bad.
             | 
             | The membrane keyboard wasn't great (the lack of a space bar
             | was a wierd choice) but it did work. We had programs on
             | casette and did get the 16Kbyte memory expansion.
             | 
             | https://en.wikipedia.org/wiki/Timex_Sinclair_1000
             | 
             | I didn't realize the Atari 2600 had basic, always thought
             | of it as a game console.
        
               | makapuf wrote:
               | You can buy this bad boy [attiny11] with _no_ ram, only
               | registers.
               | 
               | https://ww1.microchip.com/downloads/en/DeviceDoc/1006S.pd
               | f
        
         | basilikum wrote:
         | KolibriOS would fit in there, even with the data in memory. You
         | cannot load it into the cache directly, but when the cache
         | capacity is larger than all the data you read there should be
         | no cache eviction and the OS and all data should end up in the
         | cache more or less entirely. In other words it should be
         | really, really fast, which KolibriOS already is to begin with.
        
           | vlovich123 wrote:
           | Unless you lay everything out continuously in memory, you'll
           | still get cache eviction due to associativty and depending on
           | the eviction strategy of the CPU. But certainly DOS or even
           | early Windows 95 could conceivably just run out of the cache
        
             | basilikum wrote:
             | Well, yeah, reality strikes again. All you need is an
             | exploit in the microcode to gain access to AMD's equivalent
             | to the ME and now you can just map the cache as memory
             | directly. Maybe. Can microcode do this or is there still
             | hardware that cannot be overcome by the black magic of CPU
             | microcode?
        
             | chao- wrote:
             | Yeah, cache eviction is the reason I was assuming it is
             | "probably not possible architecturally", but I also figured
             | there could be features beyond my knowledge that _might_
             | make it possible.
             | 
             | Edit: Also this 192MB of L3 is spread across two Zen CCDs,
             | so it's not as simple as "throw it all in L3" either,
             | because any given core would only have access to half of
             | that.
        
             | tadfisher wrote:
             | Windows 95 only needed 4MB RAM and 50 MB disk, so that's
             | certainly doable. The trick is to have a hypervisor spread
             | that allocation across cache lines.
        
           | hrmtst93837 wrote:
           | That assumes KolibriOS or any major component is pinned to
           | one core and one cache slice instead of getting dragged
           | between CCDs or losing memory affinity. Throw actual users,
           | IO, and interrupts at it and you get traffic across chiplets,
           | or at least across L3 groups, so the nice 'everything lives
           | in cache' story falls apart fast.
           | 
           | Nice demo, bad model. The funny part is that an entire OS can
           | fit in cache now, the hard part is making the rest of the
           | system act like that matters.
        
           | RiverCrochet wrote:
           | I thought there was an MSR buried deep somewhere that enables
           | "Cache as RAM" mode and basically maps the cache into the
           | memory address space or something like that.
           | 
           | Lol a quick Google search leads me to a Linked in post with
           | all the gory technical details?
           | 
           | https://www.linkedin.com/pulse/understanding-x86-cpu-
           | cache-m...
        
         | m463 wrote:
         | I wonder how much faster dos would boot, especially with floppy
         | seek times...
        
           | userbinator wrote:
           | Instantly.
           | 
           | If you run a VM on a CPU like this, using a baremetal
           | hypervisor, you can get very close to "everything in cache".
        
           | RulerOf wrote:
           | You can get close with a VM, but there's overhead in device
           | emulation that slows things down.
           | 
           | Consider a VM where that kind of stuff has been removed, like
           | the firecracker hypervisor used for AWS Lambda. You're
           | talking milliseconds.
        
         | compounding_it wrote:
         | Maybe in 50 years the cache of CPUs and GPUs will be 1TB.
         | Enough to run multiple LLMs (a model entirely run for each
         | task). Having robots like in the movies would need LLMs much
         | much faster than what we see today.
        
           | nextaccountic wrote:
           | doubtful that we will still have this computer architecture
           | by then
        
         | shric wrote:
         | You had ~160,000 times more storage than I did for my first
         | personal computer.
        
         | HerbManic wrote:
         | My first PC had a 20MB HDD with 512Kb of RAM. So yeah that
         | could fit into cache 10 times now.
        
         | alfiedotwtf wrote:
         | > it would be amusing to see an entire early 90's OS running
         | entirely in the CPU's cache.
         | 
         | There's actually already two running (MINIX and UEFI), and it's
         | the opposite OS amusing - https://www.zdnet.com/article/minix-
         | intels-hidden-in-chip-op...
        
         | Zardoz84 wrote:
         | My first computer whole RAM could fit in L1 of a single core
         | (128k)
        
         | defrost wrote:
         | Commodore PET for me - 8 KB of RAM _and_ all the data you could
         | store and read back from a TDK 120 cassette tape . . .
         | 
         | * https://en.wikipedia.org/wiki/Commodore_PET
         | 
         | Same time as the Trash-80 and BBC micro were making inroads.
        
         | amelius wrote:
         | 640K ought to be enough for anybody.
        
         | tumdum_ wrote:
         | My first pc had 40MB hrs and 8MB ram :D
        
       | renewiltord wrote:
       | I have a gigabyte of cache on my 9684x at home!
        
       | nexle wrote:
       | Breakdown of the (semi-clickbait) 208MB cache: 16MB L2 (8MB per
       | die?) + 32MB L3 * 2 dies + 64MB L3 Stacked 3D V-cache * 2
       | 
       | For comparison, 9950X3D have a total cache of 144MB.
        
         | trynumber9 wrote:
         | > 16MB L2 (8MB per die?)
         | 
         | It is indeed 8MB per compute die but really 1MB per core. Not
         | shared among the entire CCD.
        
         | teaearlgraycold wrote:
         | I wouldn't be caught dead with less than 200MB of cache in my
         | desktop in 2026.
        
       | magicalhippo wrote:
       | Probably fun for those who already bought DDR5 memory... still
       | kicking myself for not just pulling the trigger on that 128GB
       | dual stick kit I looked at for $600 back in September. Now it's
       | listed at $4k...
       | 
       | Meanwhile I hope my AM4 will chug along a few more years.
        
         | jofzar wrote:
         | I really want a x3d because a game I play is heavily single
         | threaded, I have the income and the financial stability but I
         | can't in any good conscious upgrade to am5 with the ram prices.
         | It's insane
        
           | magicalhippo wrote:
           | Yep exactly the same situation.
           | 
           | I would not be surprised if we see casualties in adjacent
           | markets, such as motherboards, coolers and whatnot.
        
           | fakwandi_priv wrote:
           | AMD had an upgrade path with the 5700x3d, assuming you're on
           | AM4.
           | 
           | Just reading now that they went out of production half a year
           | ago which is a shame. I was very impressed being able to
           | upgrade with the same motherboard 6 years down the line.
        
             | timschmidt wrote:
             | I'm the mythical customer who went from a 1700X in a B350
             | motherboard near launch day to a 5800X3D in the same board
             | (after a dozen BIOS updates). Felt amazing. Like the old
             | 486DX2 days.
        
               | slightlygrilled wrote:
               | Same! Kept checking back for bios updates and even years
               | later they kept announcing more support! Truly crazy.
               | 
               | Other than the speed it's a very good reason to go with
               | amd, the upgrade scope is massive, on am5 you can go from
               | a 6 core and soon all the way to a 24 core with the new
               | zen6
        
           | Panzer04 wrote:
           | What game, if you don't mind my asking?
        
             | jofzar wrote:
             | World of Warcraft
        
           | tyjen wrote:
           | I was waiting too, but the one game I play often that
           | requires FPS performance decided to ruin their game with poor
           | development direction. Now, I'm planning to buy for local llm
           | hosting.
           | 
           | Here's hoping to more developments like TurboQuant to improve
           | LLM memory efficiency.
        
         | Aurornis wrote:
         | > Now it's listed at $4k...
         | 
         | You can buy 128GB of DDR5-6000 with a 9950X3D (not this newest
         | X2 version, but still a $699 CPU) and a motherboard and a case
         | for $2800 right now:
         | https://www.newegg.com/Product/ComboDealDetails?ItemList=Com...
         | 
         | If you don't need 128GB, there are quality 64GB kits for under
         | $700 on Newegg right now, which is cheaper than this CPU.
         | 
         | If someone needs to build something now and can wait to upgrade
         | RAM in a year or two, 32GB kits are in the $370 range.
         | 
         | I don't like this RAM price spike either, but in the context of
         | building a high-end system with a 16-core flagship CPU like
         | this and probably an expensive GPU, it's still reasonable to
         | build a system. If you _must_ have 128GB of RAM it can be done
         | with bundles like the one I linked above but I 'd recommend
         | waiting at least 6 months if you can. There are signs that
         | prices are falling now that panic-buying has started to trail
         | off.
         | 
         | 128GB of RAM should not cost $4K even in this market.
        
           | magicalhippo wrote:
           | No such bundle deals where I am. Absolute cheapest DDR5 128GB
           | kit around is 2 sticks of 5600 64GB for $2k.
           | 
           | Cheapest 64GB kit is $930.
           | 
           | The kit I was oh-so-close to buying was two 6400 64GB sticks.
           | 
           | Not gonna buy now, not that desperate. I have a spare AM4
           | board, DDR4 memory and heck even CPU, I'll ride this one out.
           | Likely skip AM5 entirely if something doesn't drastically
           | change.
        
             | Aurornis wrote:
             | > Absolute cheapest DDR5 128GB kit around is 2 sticks of
             | 5600 64GB for $2k.
             | 
             | That's not far from the bundle deal above, once you
             | subtract the $700 CPU.
             | 
             | If you really need 128GB the 5600 kit is fine. Having 208MB
             | of total cache on the CPU means the real world difference
             | between a 5600 kit and a slightly faster kit is negligible
             | in most use cases.
             | 
             | If you don't need to upgrade then clearly don't force an
             | upgrade right now. I just wanted to comment that $4K for
             | 128GB of RAM is a very bad price right now, even with the
             | current situation.
        
               | magicalhippo wrote:
               | > I just wanted to comment that $4K for 128GB of RAM is a
               | very bad price right now
               | 
               | Oh absolutely. Just mentioned it since I was very close
               | to buying it back then, and now it's completely bonkers.
               | 
               | That bundle deal is quite well priced all things
               | considered, it basically prices the memory where it was.
               | Again, sadly no great bundle deals here.
        
               | throwup238 wrote:
               | _> a slightly faster kit is negligible in most use cases_
               | 
               | Does that "most use cases" caveat really apply to someone
               | buying 128G of RAM? If I'm buying that much, it means I'm
               | actually going to put it through its paces, unless it's
               | just there for huge reserved guest VM overhead.
        
               | Aurornis wrote:
               | The 208MB of total cache on the CPU we're discussing does
               | a good job of reducing sensitivity to RAM speed
               | differences on this platform.
               | 
               | If you're trying to run LLMs off of the CPU instead of
               | the GPU then the RAM speed dictates a lot. It's going to
               | be slow mo matter what, though. Dual channel DDR5 just
               | isn't enough to run large LLMs that start to fill 128GB
               | of RAM and the difference between 5600 and 6400 isn't
               | going to make it usable.
               | 
               | If you're just running a lot of VMs or doing a lot of
               | mixed tasks that keep a lot of RAM occupied then you'd
               | probably have a hard time measuring a difference between
               | 5600 and 6400 if you tried with one of these X3D CPUs
               | with a lot of cache.
               | 
               | This is a frequent topic of discussion for gamers because
               | some people obsess over optimizing their RAM speed and
               | timings and pay large premiums for RAM with CAS latency
               | of 28 instead of 36. Then they see benchmarks showing
               | 1-2% differences in games or even most productivity apps
               | and realize they would have been better spending that
               | extra money on the next faster GPU or CPU or other part.
        
           | sspiff wrote:
           | I bought 192GB (4x 48GB) of DDR5-6400 for 299 euro in
           | September but returned it because I couldn't get 4 DIMMS to
           | run at decent speeds in the system.
           | 
           | 6 or so weeks after I returned it the kit was listed at 1499.
        
             | 2001zhaozhao wrote:
             | Yeah the only way to run 4 sticks of DDR5 decently is with
             | Intel. It's a bit of a shame that you can't cram enough RAM
             | to run big models.
             | 
             | The most I could get running on 10GB VRAM + 96GB RAM was a
             | REAP'd + quantized version of MiniMax-M2.5
        
               | jodleif wrote:
               | Threadripper is a good alternative. No point having a lot
               | of dual channel ram for LLMs, too slow
        
               | HauntingPin wrote:
               | I had the same issue with Intel. It's not guaranteed
               | there either.
        
               | WD-42 wrote:
               | I'm running 128gb on a 9550x now with 4x32gb sticks and
               | it's terrible. It's unstsable, post time is about 2
               | minutes (not exaggerating)and I'm stuck at a lower speed.
               | I'm considering just taking 2 of the sticks out and
               | working with 64gb and increasing my swap partition. The
               | nvme drive is fast at least.
               | 
               | This is my first time off intel and I have to say I don't
               | understand the hype.
        
               | magicalhippo wrote:
               | > It's unstsable, post time is about 2 minutes (not
               | exaggerating)
               | 
               | The long POST times must mean it's retraining the memory
               | each time, which is not normal. Just in case you
               | haven'ttried it yet, I'd start by reseating them, I've
               | had weird issues with marginally seated RAM before.
               | 
               | Also you definitely have to go much slower with 4 sticks
               | compared to two, so lower speed as much as you can. If
               | that doesn't help, I'd verify them in pairs.
               | 
               | If they work in pairs but not in quad at the slowest
               | speed, something is surely wrong.
               | 
               | Once you get them working in quad, you can start bumping
               | up the speed, might need voltage boost as well.
        
               | hxorr wrote:
               | What ddr5 speed are you running? 6000 is technically an
               | over clock, AMD only guarantees being able to run at
               | something like 4800 or 5200.
               | 
               | You may need to bump up voltages slightly for your CPU's
               | IMC (I needed to on my ryzen 8700F to run 6000 stable).
               | Its CPU sample dependant.
               | 
               | Also as other commenter pointed out, typically 4 sticks
               | will achieve lower stable clocks
        
               | mort96 wrote:
               | Got it running with 4800MT/s and literally 30 minute boot
               | times in an AM5 machine. The 30 minute boot time could be
               | worked around by enabling the (off-by-default) memory
               | context restore option in BIOS, but it really made me
               | think something was broken and it wasn't until I found
               | other people talking about 30 minute boot times that I
               | stopped debugging and just let it sit for an eternity.
               | 
               | It's so bad. I don't get why they sell AM5 motherboards
               | with 4 RAM slots.
               | 
               | At least that system has been running well for like two
               | years. But had I known that the situation is so much more
               | dire than with DDR4, I would've just gotten the same
               | amount of RAM in two sticks rather than four.
        
               | secondcoming wrote:
               | Your machine takes 30 minutes to boot because of the RAM?
               | Or it takes 30 minutes to load a model?
        
               | WD-42 wrote:
               | It's the RAM. It needs to "trained" which takes some time
               | but for for some reason these boards seem to randomly
               | forget their training, requiring it to happen again.
        
               | jazzyjackson wrote:
               | huh, its been a decade since i built a PC, whats changed?
        
               | WD-42 wrote:
               | My guess is bigger numbers, higher voltages, tighter
               | timings.
        
               | mort96 wrote:
               | DDR5 is much, _much_ more fickle than DDR4 and earlier
               | standards. I think it 's primarily due to pushing clock
               | speeds (6000 MT/s would be insanely fast for DDR4, but
               | kinda slow for DDR5).
               | 
               | Memory training has always been a thing: during boot,
               | your PC runs tests to work out what slight changes
               | between signals and stuff it needs to adapt to the
               | specific requirements of your particular hardware. With
               | DDR4 and earlier, that was really fast because the
               | timings were so relatively loose. With DDR5, it can be
               | really slow because the timings are so tight.
               | 
               | That's my best understanding of it at least.
        
               | roboror wrote:
               | It's an AMD thing
        
               | magicalhippo wrote:
               | I've never had memory training be forgotten with my AM4
               | nor LPDDR5-based laptops and NUCs. Is this a new thing
               | with AM5 or something? Or just a certain brand of BIOSes?
        
               | WD-42 wrote:
               | I'm in the same situation! My machine will take 2-5
               | minute to post every few reboots, it seems random. The
               | messed up part is the marketing material says this things
               | can handle 256gb of ram or whatever absurd number, f me
               | for thinking then 128gb should be no problem. Honestly
               | this whole thing has soured me on AMD. Yea they have
               | bigger numbers than intel but at what cost, stability?
        
               | noir_lord wrote:
               | Check you have MCR (Memory Context Restore) _enabled_ ,
               | otherwise you train the RAM way more often than you need
               | to (every boot).
        
               | noir_lord wrote:
               | You need to enable MCR (which trains the memory once and
               | caches the result for (iirc) 30 days) otherwise yeah,
               | booting is horribly slow, even the 64GB I have can take
               | several minutes but with MCR it boots basically
               | instantly.
               | 
               | Some motherboards have it off by default.
        
               | kenhwang wrote:
               | Memory training seems to be getting faster with each bios
               | update. In 2024 when I upgraded to AM5, 64GB memory
               | training took like 15 minutes. Now the same setup takes
               | about a minute when it needs to retrain, then near
               | instant with MCR (Windows 11 takes significantly longer
               | to load than the POST process).
        
               | mort96 wrote:
               | From my comment:
               | 
               | > The 30 minute boot time could be worked around by
               | enabling the (off-by-default) memory context restore
               | option in BIOS
        
           | adrian_b wrote:
           | $2800 is still a huge price in comparison with the last year.
           | 
           | Last summer, a 9950X3D + motherboard + cooler + 128 GB DRAM +
           | VAT sales taxes was the equivalent of $1400 in Europe, where
           | I live.
           | 
           | That's half of your quoted price. That was without case and
           | PSU, but adding e.g. $200 for those would not change much.
        
             | alias_neo wrote:
             | In January I upgraded my desktop, 9950X3D PS600, 64GB
             | DDR5-6000 PS600, MSI MAG Tomahawk X870E PS300, Samsung 990
             | Pro 4TB PS350, Asus Prime 9070XT PS580. I spent a another
             | PS250 on PSU and cooler and reused my case (Phanteks Evolv
             | Enthoo TG, beautiful case but horrible cooling. Will cut
             | some holes in it and if it doesnt work out look for
             | something with more airflow).
             | 
             | The RAM price was already inflated at that time, and the
             | same kit is now PS800, but in October or earlier last year
             | I'd have saved possibly the cost of the CPU/GPU on the
             | whole thing, but now it's be about the cost of a CPU/GPU
             | more expensive.
             | 
             | On a side note for anyone not aware, 9950X3D isn't the best
             | choice for pure gaming, 9850X3D is cheaper and marginally
             | better, also I went with 2 sticks of RAM kit, 4 sticks is
             | much harder to run at the advertised speed (6000) which is
             | actually an overclock.
             | 
             | Im a dev and a linux user/gamer hence my choice of CPU/GPU.
        
               | sqquima wrote:
               | Very similar config, but I bought a second pair of ram.
               | Running 4 sticks at 3600. Also, the LAN port of the
               | motherboard stopped working after a week, so I had to buy
               | an Ethernet card
        
               | alias_neo wrote:
               | Ouch, were you not willing to RMA for that ethernet port?
               | I wouldn't be too pleased after only a week if parts of
               | the board stopped working.
               | 
               | I don't really want to run my RAM that slow which is why
               | I'll probably stick with two sticks.
        
             | Aurornis wrote:
             | Yes of course. We all know prices are up.
             | 
             | I commented because someone thought that $4K was the going
             | price for 128GB of RAM, which is way too much even with the
             | demand crunch.
        
               | adrian_b wrote:
               | Due to the high prices of DRAM and SSDs they now are the
               | greatest fractions of the total price of a computer.
               | 
               | In January I was forced to upgrade an ancient Intel NUC,
               | by replacing it with an Arrow Lake H based ASUS NUC. The
               | complete system with 32 GB DRAM and 3 TB SSDs has cost
               | EUR 1200, including VAT sales tax.
               | 
               | The distribution of the price was like this:
               | Barebone mini-PC:   41%       32 GB DDR5 SODIMMs: 26%
               | 2 TB PCIe 5.0 SSD:  24%       1 TB PCIe 4.0 SSD:   9%
               | 
               | Since then, the prices of DDR5 and SSDs have continued to
               | increase, so now the fraction spent for memory would be
               | even higher than 59%.
               | 
               | Before 2026, for so small amounts of memory its cost
               | would have been much less than the rest of the system.
        
           | nicman23 wrote:
           | that bs of you don't need 128 are toxic. what if you want to
           | upgrade from ddr4 and you already have 128?
        
         | snvzz wrote:
         | I am glad I decisively ordered 96GB (2x48) DDR5 ECC back in
         | June, alongside the 9800x3d.
         | 
         | I hope this is still enough for the planned upgrade to Zen7 in
         | 2028.
        
           | Panzer04 wrote:
           | You're basically me. I was mulling 48 vs 96, decided 200$
           | wasn't worth quibbling too much over and bought 96GB in
           | August.
           | 
           | Feeling pretty chuffed now XD (though still sad because
           | building a new PC is dumb when RAM costs more than a 24 core
           | monster CPU)
        
             | snvzz wrote:
             | This is the good side.
             | 
             | The not so good side is that getting a RVA23 development
             | board this year with an usable size of RAM (for e.g.
             | compiling and linking large code bases) is not going to be
             | cheap.
        
           | disillusioned wrote:
           | Same... got 2x48 DDR5 for $304 back in February of 2025.
           | Equivalent kits are going for $900-$1,100. Madness.
        
           | mroche wrote:
           | I'm looking at building a new system, and was waiting to see
           | what happens with this chip and Intel's Arc Pro B70 card. I
           | can't find ECC UDIMMs of 64GB per-stick to make 128GB, but I
           | can put together two solo UDIMMs of 32GB or 48GB for $800 and
           | $1000 per stick respectively.
           | 
           | I really want to see what enabling the L3 cache options in
           | the BIOS do from a NUMA standpoint. I have some projects I
           | want to work on where being able to even just simulate NUMA
           | subdivisions would be highly useful.
        
             | snvzz wrote:
             | I was surprised to find that ECC modules available were 24
             | or 48, so 128GB with 2 sticks was impossible.
             | 
             | While I was aiming at 128, I settled for 96GB, because any
             | more than 2 sticks means a sharp drop in RAM clocks this
             | generation.
        
         | jmyeet wrote:
         | After randomly breaking the AM4 CPU and motherboard in my 4
         | year old PC last year and seeing that at the time I'd spent
         | almost a new PC to get new parts and rebuild it. Less if I
         | wanted to do a complete rebuild myself but I'm over building
         | PCs. I've done that for years.
         | 
         | It was an expensive mistake as I bought a few options to
         | experiment including a NUC and an M4 Mac Mini but eventually
         | bought a 9800X3D 5070Ti PC for <$2 and for no reason in
         | particular I bought a 64GB DDR5-6000 kit for $200 in August or
         | so. I checked recently and that kit is pushing $1000. I also
         | bought a 4080 laptop and bought a 64GB kit and an extra SSD for
         | it too last year.
         | 
         | That's pretty lucky given what's happened since. I don't claim
         | any kind of foresight about what would happen.
         | 
         | I do kind of want to take the parts I have and build another
         | AM4 PC. The 5900XT is not a bad option with 16 cores for ~$300
         | but my DDR4 RAM is almost useless because the best deals now
         | are for combos of CPU + motherboard + RAM at steep discounts.
         | 
         | You can get some good deals on prebuilts still. Not as good as
         | 6+ months ago but still not bad. Costco has a 5080 PC for
         | $2300. There's no way I'm going overboard and building a 128GB+
         | PC right now.
         | 
         | I've seen multiple RAM spikes. We had one at the height of the
         | crypto hysteria IIRC but this is significantly worse and is
         | also impacting SSDs. I kinda wish I'd bought 1-2 4TB+ SSDs last
         | year but oh well.
         | 
         | We're really waiting for the AI bubble to pop. Part of me think
         | sthat'll be in the next year but it could stay irrational
         | substantially longer than that.
        
           | sundvor wrote:
           | The C30 64GB kits are nearly impossible to buy now, so, well
           | done. Got one in September '23 for ~$380 AUD, on the rare
           | occasions it's available today it's been over $1600 AUD.
           | 
           | I upgraded my UPS to a sine interactive unit to minimise the
           | risk of it dying to bad power while the market is so crazy...
        
         | DeathArrow wrote:
         | >Meanwhile I hope my AM4 will chug along a few more years.
         | 
         | I am fine with my 2 year old 128GB DDR4 for now. I will just
         | upgrade the 14700K to 14900KS CPU and wait 2 more years.
         | 
         | Judging by the benchmarks newer CPUs aren't much better for
         | multithreading workloads than 14900KS anyway, so it doesn't
         | make a lot of sense to upgrade to newer CPUs, DDR5 and a new
         | mobo.
        
         | tom_alexander wrote:
         | > Probably fun for those who already bought DDR5 memory
         | 
         | Nah, those of us who already bought DDR5 memory also already
         | bought decent CPUs. Dropping another $1k for these incremental
         | gains would be silly. It'd make a lot more sense if DDR5 had
         | been around longer so that people had the option to make
         | generational upgrades to this CPU but DDR5 on AMD has only been
         | around for Zen4 and Zen5.
        
         | throawayonthe wrote:
         | oh wow you weren't joking:
         | https://pcpartpicker.com/products/memory/#xcx=0&b=ddr5&Z=131...
         | 
         | (cheapest at $1240 USD)
        
           | MrDOS wrote:
           | PCPartPicker are also publishing charts showing the
           | astronomic rise in DDR5 prices over time:
           | https://pcpartpicker.com/trends/price/memory/. Those charts
           | don't cover any kits with 64 GB sticks, but they're a good
           | demonstration of the general scale.
        
         | tarangsutariya wrote:
         | Wonder how much sales amd and intel are losing because of tight
         | DDR5 supply
        
           | magicalhippo wrote:
           | I can't imagine it's looking good in the consumer space, but
           | server space seems to be lit[1]:
           | 
           |  _Su said that typically, the first quarter (Q1) is slower
           | due to seasonal patterns, but AMD has seen its data center
           | business expand from Q4 into Q1, demonstrating ongoing
           | strength across both CPUs and GPUs. This growth underscores
           | the company's ability to capitalize on rising demand for AI
           | compute and enterprise workloads, even during traditionally
           | quieter periods._
           | 
           |  _"We are going into a big inflection year here in 2026. The
           | CPU business is absolutely on fire."_
           | 
           | [1]: https://stocktwits.com/news-articles/markets/equity/amd-
           | ceo-...
        
           | aetimmes wrote:
           | None. Every component is seeing huge demand.
        
       | fc417fc802 wrote:
       | Given that the dies still have L3 on them does this count as L4
       | or does the hardware treat it as a single pool of L3?
       | 
       | Would be neat to have an additional cache layer of ~1 GB of HBM
       | on the package but I guess there's no way that happens in the
       | consumer space any time soon.
        
         | trynumber9 wrote:
         | Per compute die it functions as one 96M L3 with uniform
         | latency. It is 4 cycles more latency than the configuration
         | with smaller 32M L3. But there are two compute dies, each with
         | their own L3. And like the 9950X coherency between these two L3
         | is maintained over global memory interconnect to the third (IO)
         | die.
        
       | monster_truck wrote:
       | The extra cache doesn't do a damn thing (maybe +2%)
       | 
       | The lower leakage currents at lower voltages allowed them to
       | implement a far more aggressive clock curve from the factory.
       | That's where the higher allcore clock comes from (+30W TDP)
       | 
       | I'm not complaining at all, I think this is an excellent way to
       | leverage binning to sell leftover cache.
       | 
       | Though if I may complain, Ars used to actually write about such
       | things in their articles instead of speculate in a way that
       | suspiciously resembles what an AI would write.
        
         | EnPissant wrote:
         | It's very workload dependent. It certainly does more than 2% on
         | many workloads.
         | 
         | See https://www.phoronix.com/review/amd-
         | ryzen-9-9950x3d-linux/10
         | 
         | > Here is the side-by-side of the Ryzen 9 9950X vs. 9950X3D for
         | showing the areas where 3D V-Cache really is helpful:
         | 
         | Coincidentally, it looks they filtered to all benchmarks with
         | differences greater than 2%. The biggest speedup is 58.1%, and
         | that's just 3d vcache on half the chip.
        
           | spockz wrote:
           | I think GP was saying that the _additional_ 3D cache on this
           | chip compared to the standard x3d isn't going to do much.
           | 
           | I'm curious to see whether the same benchmarks benefit again
           | so greatly.
        
             | bell-cot wrote:
             | > I think GP was saying...
             | 
             | Agree. The article's 2nd para notes "AMD relies on its
             | driver software to make sure that software that benefits
             | from the extra cache is run on the V-Cache-enabled CPU
             | cores, which usually works well but is occasionally error-
             | prone." - in regard to the older, mixed-cache-size chips.
             | 
             | > I'm curious to see...
             | 
             | Yeah - though I don't expect current-day Ars Technica will
             | bother digging that deep. It could take some very
             | specialized benchmarks to show such large gains.
        
               | monster_truck wrote:
               | Some of their writers, who are quite excellent, still do.
               | Others just seem to regurgitate press releases with very
               | little useful investigation.
               | 
               | How critical of the lazy writers I am may seem outsized,
               | but I grew up reading and learning from the much better
               | version of Ars -one I used to subscribe to.
        
               | spockz wrote:
               | I hoping that phoronix will be able to redo the benchmark
               | of the 9950x3D with this new X3D2 variant.
               | 
               | I might even shell out for an upgrade to AM5 and DDR5. On
               | the other hand, my 5900X is still blazing fast.
        
             | adrian_b wrote:
             | On AMD the L3 cache is partitioned between the 2 chiplets.
             | 
             | So for 9950X3D half of the cores use a small L3 cache.
             | 
             | For applications that use all 16 cores, the cases where
             | X3D2 provides a great benefit will be much more frequent
             | than for a hypothetical CPU where the same cache increase
             | would have been applied to a unified L3 cache.
             | 
             | The threads that happen to be scheduled on the 2nd chiplet
             | will have a 3 times bigger L3 cache, which can enhance
             | their performance a lot and many applications may have
             | synchronization points where they wait for the slowest
             | thread to finish a task, so the speed of the slowest thread
             | may have a lot of influence on the performance.
        
         | Aurornis wrote:
         | > The extra cache doesn't do a damn thing (maybe +2%)
         | 
         | It depends on the task. For some memory-bound tasks the extra
         | cache is very helpful. For CFD and other simulation workloads
         | the benefits are huge.
         | 
         | For other tasks it doesn't help at all.
         | 
         | If someone wants a simple gaming CPU or general purpose CPU
         | they don't need to spend the money for this. They don't need
         | the 16-core CPU at all. The 9850X3D is a better buy for most
         | users who aren't frequently doing a lot of highly parallel work
        
           | YoumuChan wrote:
           | But consumer product does not support SDCI (only Epyc Turin
           | supports it), so it does not benefit too much if an
           | accelerator is involved.
        
             | monster_truck wrote:
             | It's also useful to point out that the use cases and
             | workloads where SDCI are most beneficial are far, far
             | beyond the scope of what anyone will have installed in a
             | Zen rig. Dual 100G networking cards? The cost of both of
             | those damn near buys all of a 9950X3D2 setup.
        
               | justincormack wrote:
               | no, dual 100Gb are not that expensive any more, eg
               | https://www.scan.co.uk/products/2-port-
               | intel-e810-cqda2blk-d... UK retail for gbp349.
        
           | monster_truck wrote:
           | It really doesn't. In virtually every case the work is being
           | completed faster than the cache can grow to that size. What
           | little gains are being realized are from not having to wait
           | for cores with access to the cache to become available.
        
             | Numerlor wrote:
             | There are some very specific workloads (say simple object
             | detection) that fit into cache and have crazy performance
             | where the value of the cpu will be unbeatable, as the
             | alternative is one of the cache epycs, everywhere else
             | it'll only be small improvement if the software is not
             | purpose made for it
        
             | Aurornis wrote:
             | > It really doesn't. In virtually every case the work is
             | being completed faster than the cache can grow to that
             | size.
             | 
             | If your tasks don't benefit then don't buy it.
             | 
             | But stop claiming that it doesn't help anywhere because
             | that's simply wrong. I do some FEA work occasionally and
             | the extra cache is a HUGE help.
             | 
             | There are also a lot of non-LLM AI workloads that have
             | models in the size range than fit into this cache.
        
           | zahlman wrote:
           | Sorry, what is "CFD" in this context?
        
             | detaro wrote:
             | https://en.wikipedia.org/wiki/Computational_fluid_dynamics
        
           | addaon wrote:
           | CFD benefits from cache, but it benefits even more from
           | sustained memory bandwidth, no? A small(ish) chunk of L3 +
           | two channels of DRAM is not going to compete with a quarter
           | as much L3 plus eight channels of DRAM when typical working
           | set sizes (in my experience) are in the tens of gigabytes, is
           | it?
        
       | erulabs wrote:
       | 9950X3D2? AMD, who is making you name your products like this? At
       | some point just give up and name the chip a UUID already.
        
         | jofzar wrote:
         | I actually don't mind this one, 9950 is the actual chip, x3d is
         | the cache (where it's larger) and the 2 stands for it being on
         | both chiplets.
        
         | sidkshatriya wrote:
         | Like your UUID joke but agree with sibling comment that
         | 9950X3D2 is actually a good name.
        
         | hu3 wrote:
         | can't agree. this name has logical meaning
        
       | throwaway85825 wrote:
       | It's disappointing that they had this for years but didn't
       | release it until now.
        
         | stingraycharles wrote:
         | I think it's mostly that they had leftover cache.
        
           | magicalhippo wrote:
           | Makes sense. RAM pricing surely has lead to a fall of AM5
           | high-end CPU purchases, might as well try to get some extra
           | cash from those who still buy. Bin the remaining now non-X3D
           | chips as something else.
        
             | Ekaros wrote:
             | Bad time to move entirely new platform. Perfect time to
             | sell to upgrade junkies just CPU.
        
           | neRok wrote:
           | This video made the argument that AMD released it to not give
           | Intel a look-in: [AMD KILLED Intel's 290K Dreams w/ R9
           | 9950X3D2](https://www.youtube.com/watch?v=u7SyrDPbKls)
        
             | stingraycharles wrote:
             | I like this theory more, perhaps it's both.
        
       | tw1984 wrote:
       | that is larger than the HDD of my first PC.
        
       | jaimex2 wrote:
       | Can someone like... boot Windows 98 on these on a system with no
       | ram?!
        
         | bell-cot wrote:
         | _Conceptually_ - yes, easily.
         | 
         | But to do it literally - I'm not a low-level motherboard EE,
         | but I'd bet you're looking at 5 to 7 figures (US $) of
         | engineering work, to get around all the ways in which that
         | would violate assumptions baked into the designs of the CPU,
         | support chips, firmwares, etc.
        
           | ggm wrote:
           | Make a fake ram which offers write through guarantee and
           | returns bus no matter what address is referenced. You could
           | possibly short circuit any "is ram there" test if it just
           | says yes for whatever size and stride got configured.
        
           | anticensor wrote:
           | The CPU literally initialises itself without DDR then
           | initialises the DDR PHY, there must be a way of keeping the
           | CPU in that "cache as RAM" mode.
        
         | brandnewideas wrote:
         | Theoretically anything is possible with enough thought and
         | work.
        
       | DeathArrow wrote:
       | My first computer had 64KB of RAM. My first PC had 8MB of RAM.
        
       | 2001zhaozhao wrote:
       | I don't really see a huge reason to buy this other than it being
       | a top-tier halo product.
       | 
       | For gaming, AMD already pins the game threads to the CCD with the
       | extra cache pretty well.
       | 
       | For multi-threaded workloads the gain from having cache on both
       | CCDs is quite small.
        
         | adrian_b wrote:
         | The gain is very workload dependent, so there are no generally-
         | applicable rules.
         | 
         | There are many applications which need synchronization between
         | threads, so the speed of the slowest thread has a
         | disproportionate influence on the performance.
         | 
         | In such applications, on X3D2 the slowest thread has a 3 times
         | bigger cache on an X3D2 vs. X3D. That can make a lot of
         | difference.
         | 
         | So there will be applications with no difference in
         | performance, but also applications with a very large difference
         | in performance, equal to the best performance differences shown
         | by X3D vs. plain 9950X.
        
         | pixl97 wrote:
         | It really comes down to how much more this CPU is over the next
         | one down if you're building a new rid for a long period of
         | time. I'm running on a 5950X which is coming up on it's 6 years
         | in November. I could have spend a little less on the next model
         | down, but I expect this rig will last me for a few more years
         | (especially with how much memory is). The per year extra
         | expense for that CPU was almost nothing over its lifetime.
         | 
         | Now, would I upgrade an existing computer with a slightly
         | slower processor with it, probably not.
        
       | sylware wrote:
       | With the best silicon tech, in R&D, what would be the maxium
       | static RAM(L1 cache) you could really slap to a 8 core CPU? (Zero
       | DRAM).
        
       | pwr22 wrote:
       | I'm interested to know if the L3 cache all behaves as a single
       | pool for any core on either CCD, whether there's a penalty in
       | access time depending on locality or whether they are just
       | entirely localised.
        
         | undersuit wrote:
         | AMD didn't have to introduce a special driver for the Ryzen 9
         | 5950x to keep threads resident to the "gaming" CCD. There was
         | only a small difference between the 5950x and the non-X3d Ryzen
         | 7 5800x in workloads that didn't use more than 8 cores unlike
         | the observed slowdowns in the Ryzen 9s 7950X3D and 7900X3D when
         | they were released compared to the Ryzen 7 7800X3D .
         | 
         | When the L3 sizes are different across CCDs the special AMD
         | driver is needed to keep threads pinned to the larger L3 CCD
         | and prevent them from being placed on the small L3 CCD where
         | their memory requests can exploit the other CCD's L3 as an L4.
         | The AMD driver reduces CCD to CCD data requests by keeping
         | programs contained in one CCD.
         | 
         | With equal L3 caches when a process spills onto the second CCD
         | it will still use the first's L3 cache as "L4" but it no longer
         | has to evict that data at the same rate as the lopsided models.
         | Additionally the first CCD can use the second CCD's L3 in kind
         | reducing the number of requests that need to go to main memory.
         | 
         | The same sized L3s reduce contention to the IO die and the
         | larger sized L3s reduce memory contention, it's a win-win.
         | 
         | https://www.phoronix.com/review/amd-3d-vcache-optimizer-9950...
        
         | phire wrote:
         | The short answer is that L3 is local to each CCD.
         | 
         | And that answer is good enough for most workloads. You should
         | stop reading now.
         | 
         | _______________________
         | 
         | The complex answer is that there is some ability one CCD to
         | pull cachelines from the other CCD. But I've never been able to
         | find a solid answer for the limitations on this. I know it can
         | pull a dirty cache line from the L1/L2 of another CCDs (this is
         | the core-to-core latency test you often see in benchmarks, and
         | there is an obvious cross-die latency hit).
         | 
         | But I'm not sure it can pull a clean cacheline from another CCD
         | at all, or if those just get redirected to main memory _(as the
         | latency to main memory isn 't that much higher than between
         | CCDs)._ And even if it can pull a clean cacheline, I'm not sure
         | it can pull them from another CCD's L3 (which is an eviction
         | cache, so only holds clean cachelines).
         | 
         | The only way for a cacheline to get into a CCD's L3 is to be
         | evicted from an L2 on that core, so if a dataset is active
         | across both CCDs, it will end up duplicated across both L3s.
         | Cachelines evicted from one L3 do NOT end up in another L3, so
         | an idle CCD can't act as a pseudo L4.
         | 
         | I haven't seen anyone make a benchmark which would show the
         | effect, if it exists.
        
         | trynumber9 wrote:
         | It does not. For any of the dual CCD parts AMD has ever
         | released for consumers. Even Strix Halo which has higher
         | bandwidth, lower latency interconnect doesn't make a single L3
         | across CCDs.
         | 
         | It'll probably only happen when they have a singular, large die
         | filled with cache upon which both CCDs are stacked.
         | 
         | Run this test if you're curious:
         | https://github.com/ChipsandCheese/MemoryLatencyTest
         | 
         | On a regular CCD:
         | 
         | 32768,46.115
         | 
         | 65536,74.243
         | 
         | 98304,85.699
         | 
         | 131072,91.42
         | 
         | 262144,99.402
         | 
         | On a 3D cache CCD:
         | 
         | 32768,11.992
         | 
         | 65536,12.712
         | 
         | 98304,29.921
         | 
         | 131072,49.91
         | 
         | 262144,86.059
        
       | swarnie wrote:
       | Factorio mega basing just found a new ceiling.
        
         | Lightkey wrote:
         | I'm curious to see if that is true. The maximum amount of cache
         | addressable per core didn't increase after all.
        
       | senfiaj wrote:
       | Back in 2004 my PC RAM was 256. My relative's laptop had 128.
       | That's crazy when a modern CPU cache can theoretically host an OS
       | (or even multiple OSes) from early 2000s.
        
         | egeozcan wrote:
         | The RAM prices are so high and the storage is also getting more
         | expensive every day, so we're forced to fit everything inside
         | the CPU cache as a solution! /s
        
           | sqquima wrote:
           | It would be interesting if it allowed to use the cache as ram
           | and could boot without any sticks on the motherboard.
        
             | 0-_-0 wrote:
             | Instead of a cache you could put down an SRAM buffer, it
             | would be more efficient than a cache and just as fast. And
             | addressable. Interesting idea.
        
             | addaon wrote:
             | Several processors support this by effectively locking
             | cache lines. At the low end, it allows a handful of fast
             | interrupt routines without dedicated TCM. At the high end,
             | it allows boot ROMs to negotiate DRAM links in software,
             | avoiding both the catch 22 and complex hardware
             | negotiation.
        
         | addaon wrote:
         | The Power4 MCM had 128 MB cache in 2001. The G4 TiBook sold the
         | same year came with 128 MB of system RAM base, and OS X
         | supported 64 MB configurations for a few years after this.
        
       | rietta wrote:
       | I am so grateful that I bought my 128 GB ram kit in January of
       | last year for my own 9950 upgrade. We just built my dad a 7000
       | series to replace his old AM4 (2017 build) and 32 gigs DDR five
       | was nearly the same price at Micro Center that I paid last year.
       | I was able to gift him an Nvidia 1060 discreet graphics card so
       | that he could continue to run his two monitors. The newer
       | motherboards have much less on board capability for that.
        
         | hu3 wrote:
         | 1060 is a sweet card for multi monitor. good on you for gifting
         | him.
        
           | rietta wrote:
           | I upgraded to a 4070 super last year. I ran both cards at the
           | same time for a little bit, but it got really frustrating to
           | keep the wrong card from being assigned to a particular task
           | with llama. I really should've taken an R&D tax credit on my
           | AI research but I'm still able to expense it for the
           | business.
        
       | Jotalea wrote:
       | so you're telling me I can (theoretically) have a full Alpine
       | Linux installation in _just_ the CPU? I 'm impressed
        
       | varispeed wrote:
       | I know the prices of RAM are high, but 256GB RAM limit seems like
       | omission. If they supported at least 512GB in quad or eight
       | channel that would be something worth looking at for me. I know
       | there is Threadripper but ECC memory is out of reach.
        
       | electronsoup wrote:
       | Whenever I see a chip like this, I think "why wont my company let
       | me use a decent computer"
        
       | SubiculumCode wrote:
       | Oh man. I am running computations on my server that involve
       | computing geodesic distances with the heat method. The job turns
       | out to be a L3 cache thrasher, leaving my cpus underutilized for
       | multi worker jobs .... 208mb instead of my 25 per socket sounds
       | amazing
        
         | AnthonyMouse wrote:
         | They sell essentially the same chips with more CCDs as Epyc
         | instead of Ryzen. 9684X has more than 1GB of L3 per socket (but
         | it's not cheap).
        
       | MaximilianEmel wrote:
       | They should allow it to function without any external RAM.
        
       | kristianp wrote:
       | Nobody adds L1+L2+L3 like that, because L1 stores a subset of L2
       | and L2 stores a subset of L3. Just say 192MB of L3.
        
         | Retr0id wrote:
         | It depends on the implementation, it is possible for a cache
         | line to be in L1 but not L2, etc.
        
       | abcde666777 wrote:
       | A year ago I swapped out a 5800x for a 5800x3d to get more stable
       | frame rates in Counterstrike 2. Made a sizable difference,
       | especially to 1% lows, so these large caches can clearly be a big
       | boon. Granted it's also obvious the game is poorly optimized, the
       | gains look less significant for most other titles.
        
       ___________________________________________________________________
       (page generated 2026-03-28 23:00 UTC)