[HN Gopher] AMD's Chiplet APU: An Overview of Strix Halo
       ___________________________________________________________________
        
       AMD's Chiplet APU: An Overview of Strix Halo
        
       Author : zdw
       Score  : 155 points
       Date   : 2025-10-18 04:26 UTC (18 hours ago)
        
 (HTM) web link (chipsandcheese.com)
 (TXT) w3m dump (chipsandcheese.com)
        
       | oDot wrote:
       | I read somewhere, but can't remember where, that a major reason
       | those APUs aren't as efficient as the Apple ones is a conscious
       | decision to share the architecture with Epyc and therefore accept
       | worse efficiency at lower wattage as a tradeoff.
       | 
       | Can someone confirm/refute that?
        
         | christkv wrote:
         | They are ok but yeah they do not have anything like the memory
         | bandwidth of an m3 ultra. But they also cost a lot less. I'm
         | primarily looking to replace my older desktop but just have to
         | make sure i can run an external gpu like the A6000 that i can
         | borrow from work without having to spend a week fiddling with
         | settings or parameters
        
         | jeswin wrote:
         | In this review, Hardware Canucks tested [1] the M4 Pro (3nm 2nd
         | gen) and the 395+ (4nm) at 50w and found the performance being
         | somewhat comparable. The differences can be explained away by
         | 3nm vs 4nm.
         | 
         | [1]: https://www.youtube.com/watch?v=v7HUud7IvAo
        
           | aurareturn wrote:
           | It isn't comparable at all. In MT, maybe it is comparable
           | with the M4 Pro still winning. In ST, it is 3-4x ahead of
           | Strix Halo in efficiency.
        
           | jamiek88 wrote:
           | How the heck is a 3.6x faster single thread M4 Pro
           | 'comparable'? Which by the way you can buy in a $600 prebuilt
           | not $2500 if you can even find this unobtanium chip.
        
       | christkv wrote:
       | I love the concept of it and have been thinking about getting one
       | the only problem I see right now is no ability as far as I can
       | see to get an external dock to run an additional external gpu in
       | the future.
        
         | izacus wrote:
         | I'm not sure what do you mean? I'm running eGPU with my Strix
         | Point laptop via Thunderbolt.
         | 
         | I've also seen quite a few mini PCs with Oculink port and Strix
         | Halo CPUs.
        
           | christkv wrote:
           | I was reading people having problems getting the external
           | cards working if they had a lot of memory?
        
             | Maxious wrote:
             | > In a Linux context I got some GPUs working and I can add
             | some [external] GPU devices. Minis forum when I reached out
             | to them said they don't officially support either via the
             | Thunderbolt compatibility USB 4, USB4 v2 or even the built-
             | in PCIe slot. Yeah, not technically officially supported
             | and it's because of the resource allocation and the BAR
             | space and they need somebody on the BIOS team to understand
             | that to fix it.
             | 
             | https://www.youtube.com/watch?v=TvNYpyA1ZGk
             | 
             | There are ways to manage BAR better in linux or with UEFI
             | preboot environments for windows as hobbyists have been
             | doing for ages due to bad BIOS support
             | https://github.com/xCuri0/ReBarUEFI
        
               | InTheArena wrote:
               | Thanks for the pointer. I have been struggling to get
               | either a oculink or USB4 PCIe tunnel to work with the
               | framework desktop. HOpefully some clues here.
        
         | rcarmo wrote:
         | There are plenty of mini-PCs with USB4 and Oculink, and you can
         | get an M.2 adapter (might be tricky to retrofit into a laptop
         | though).
        
       | TulliusCicero wrote:
       | So potentially competitive with a 5070M for graphics? Sounds very
       | nice, as long as price and power draw are reasonable.
        
         | makeitdouble wrote:
         | Power draw is around 75W. It can be manual boosted, but will
         | stay below 100W under all circumstances (from memory, as I was
         | researching the Z13)
         | 
         | The chip itself should accept higher power draws, and ASUS
         | usually isn't shy on feeding 130+W to a laptop, so the 75W
         | figure was quite a surprise to me.
        
       | ivape wrote:
       | I was just thinking the other day that AMD can match Nvidia pound
       | for pound on the raw hardware specs, and if they don't just yet,
       | they get pretty close. If AI is a bubble, then AMD should not
       | catch up. If there isn't a bubble, then there is no choice but to
       | learn to use whatever is out there and AMD is truly set to be
       | another trillion dollar company. The 10% stake OpenAI took is
       | going to look like a Google buying YouTube moment in the long
       | run.
       | 
       | And it's worth noting, AMD has always matched up with Nvidia
       | hardware wise for decades, plus or minus. They are an interesting
       | company in that they took on both Nvidia and Intel, and is still
       | continuing to do so.
        
       | chao- wrote:
       | Comparing this against mobile dGPUs and the (finally real) DGX
       | Spark, this feels like a latent market segment that has not
       | arrived at its final form. I don't know what delayed the DGX
       | Spark so long, but it granted AMD a huge boon by allowing them
       | capture some market mindshare first.
       | 
       | Compared to discrete GPUs (mobile or not), the advantage of a
       | dGPU is memory bandwidth. The disadvantage of a dGPU is power
       | draw and memory capacity--if we set aside CUDA, which I grant is
       | a HUGE thing to just "set aside".
       | 
       | If we mix in the small DGX Spark desktops, then those have an
       | additional advantage in the dual 200Gb network ports that allow
       | for RDMA across multiple boxes. One could get more from of a
       | small stack (2, 3 or 4) of those than from the same number of
       | Strix Halo 395 boxes. However, as sexy as my homelab-brain finds
       | a small stack of DGX Spark boxes with RDMA, I would think that
       | for professional use, I would rather have a GPU server (or
       | Threadripper GPU workstation) than four DGX Spark boxes?
       | 
       | Because the DGX Spark isn't being sold in a laptop (AFAIK,
       | CMIIW), that is another differentiator in favor of the Strix
       | Halo. Once again, it points to this being a weird, emerging
       | market segment, and I expect the next generation or two will
       | iterate towards how these capabilities really ought to be
       | packaged.
        
         | wffurr wrote:
         | "dGPU" usually means "discrete GPU". Do you mean "iGPU" for
         | "integrated GPU" instead?
         | 
         | Strix Halo is also being marketed for gaming but the
         | performance profile is all wrong for that. The CPU is too fast
         | and the iGPU still not strong enough.
         | 
         | I am sure it's amazing at matmul though.
        
           | chao- wrote:
           | Yes, I intended to use the term "discrete GPU" before using
           | "dGPU" as a shorthand for that exact reason (in the second
           | paragraph). I now see that I edited the first paragraph to
           | use "dGPU" without first defining it as such.
           | 
           | I also agree that they aren't for gaming (something I know
           | little about). My comment was with respect to compute
           | workloads, but I never specified that. Apologies.
        
           | speed_spread wrote:
           | As a casual gamer I'm already okay with the RTX 3050 dGPU on
           | my laptop. Reports put Strix Halo at RTX 4070 level which is
           | massive for an iGPU and certainly allows for 2k single screen
           | gaming. Hardcore gaming will always require a desktop with
           | PCIe boards.
        
             | lostmsu wrote:
             | Strix Halo is nowhere near RTX 4070 (desktop at least, not
             | familiar with laptop GPUs).
        
               | speed_spread wrote:
               | Maybe there's been some selective optimization and
               | careful marketing but to even be in that ballpark for
               | some games now means that more is coming.
               | 
               | https://www.techspot.com/news/106835-amd-ryzen-strix-
               | halo-la...
        
               | lostmsu wrote:
               | This link is a terrible source. In one of the graphs 4060
               | is faster than 4070. This speaks to the quality of
               | testing.
        
               | kimixa wrote:
               | In some power constrained scenarios that sort of thing is
               | often petty reproducible.
               | 
               | Especially if the different SKUs have different power
               | budgets. Laptop GPU naming and performance is a bit of a
               | mess, as in the example shown (the 4060 on the Asus TUF
               | Gaming A16 has a limit of 140w GPU+CPU, while the 4070 on
               | the Asus Proart PX13 has 115w GPU+CPU - and even that is
               | a "custom" non-default mode with 95w being the actual
               | out-of-the-box limit).
               | 
               | With wildly varying power profiles laptop graphics need
               | to be compared by chassis (and the cooling/power supply
               | that implies) as much as by GPU SKU.
        
               | lostmsu wrote:
               | That just proves the point about the source, right?
        
           | AmVess wrote:
           | I have one. Framework Desktop mainboard that I put into a
           | larger ITX chassis and regular power supply.
           | 
           | It's fine for 1440p gaming. I don't use it for that, but it
           | would not be a bother if that was all I had.
        
           | dismalaf wrote:
           | From what I've seen the gaming benchmarks are fantastic.
           | Beats the mobile 5070 for some games and settings, or
           | slightly behind on others. While being very far ahead of
           | every other iGPU.
           | 
           | I have a laptop with an Nvidia GPU. Ruins battery life and
           | makes it run very hot. I'd pay a lot for a powerful iGPU.
        
         | justincormack wrote:
         | Fyi its not dual 200Gb its 1x 200 or 2x 100Gb
        
           | justinclift wrote:
           | How sure are you of that? :)
           | 
           | Everything I've seen says it's 2x 200GbE.
           | 
           | One of many examples:
           | https://www.storagereview.com/review/nvidia-dgx-spark-
           | review...
        
             | wmf wrote:
             | That review says "Allows for a maximum of 200G bandwidth"
             | between the two ports.
        
         | linuxftw wrote:
         | The DGX Spark seems to have one intended usecase: local AI
         | model development and testing. The Strix Halo is an amd64 with
         | iGPU, it can be used for any traditional PC workload, and is a
         | reasonable local-ai target device.
         | 
         | For me, the Strix Halo is the first nail in the coffin of
         | discrete GPUs inside laptops for amd64. I think Nvidia knows
         | this, which is why they're partnering with Intel to make an
         | iGPU setup.
        
           | InTheArena wrote:
           | I think it's beyond that even - it's for local AI toolchain
           | model development and testing or those people who have a ore-
           | exisitng nvidia deployment infrastructure
           | 
           | It feels like nVidia spent a ton of money here on a piece of
           | infrastructure (the big network pipes) that very few people
           | will ever leverage, and that the rest of the infrastructure
           | constrains somewhat.
        
         | Tuna-Fish wrote:
         | Next gen, AMD has the Medusa Halo with (reportedly) a 384bit
         | LPDDR6 bus. This should get you twice the memory of what Strix
         | Halo has with 1.7 times the throughput when using memory that's
         | already announced, with even better modules coming later.
         | 
         | I think with the success of Strix Halo as an inference
         | platform, this market segment is here to stay.
        
           | karmakaze wrote:
           | I'm really excited and looking forward to this refresh. The
           | APU spec leaks for the upcoming PS6 and XBox have some clues
           | as well. My wishlist: more memory bandwidth, more GPU/NPU
           | cores, actual unified memory rather than designating, more
           | PCIe lanes. Of course there could be more/new AMD packaging
           | magic sprinkled in too.
        
       | makeitdouble wrote:
       | The saddest part of this is the lack of availability: at this
       | point there's 2 standard laptops using this chip, the Z13 being
       | the only high perf one. There's the Framework lines as well, but
       | they aren't available in many countries, and it's a very specific
       | public.
       | 
       | And that's after half a year after the first machines to come to
       | the market.
       | 
       | I love the Z13, but it's clearly a niche machine, so I'm assuming
       | they are having a really hard time manufacturing the chips ? All
       | the capacity is getting eaten by Apple ?
        
         | voidmain0001 wrote:
         | HP ZBook Ultra G1a is a great option and can be bought with up
         | to 128GB RAM.
        
           | makeitdouble wrote:
           | Yes. It looks to be the more standard option form factor
           | wise, which is a blessing and a curse.
           | 
           | For instance they went for the standard lower resolution
           | display (1920x1440 for 14" vs 2560x1600 for 13" on the Z13).
           | The thermals also looked better on the Z13, which comes
           | partly with the form factor, partly with Asus optimizing for
           | that for so many years.
           | 
           | Of cours the Z13 keyboard is meh, I expect most owners to
           | have it detached 90% of the time and handle the machine more
           | like a standalone screen/touch/pen input.
        
         | ThreatSystems wrote:
         | Cognisant US pricing for the HP Z Book Ultra was astronomical,
         | within the EU it's on par with standard laptops and to good
         | effect. The only regret I have is ordering on release day and
         | not wanting to wait for the 128gb version; but battery life and
         | power has remain unmatched to any of the pretty large workloads
         | I have thrown at it!
         | 
         | Outside of laptops, Beelink and co. are making NUCs with them
         | which are relatively affordable!
         | 
         | I do agree, the scarcity has limited their opportunity to
         | assess the growth opportunity.
        
           | green7ea wrote:
           | I also have one with 64gb -- best laptop I've ever used :-).
           | I have the same regret of not waiting for the 128gb version
           | to be available before buying.
        
         | wmf wrote:
         | Beelink, GMKtec, Minisforum, Corsair...
        
         | dontlaugh wrote:
         | You can't even buy the Z13 with more than 32 GB in most of
         | Europe and certainly not with the 2-3 years of warranty most
         | employers require for hardware they purchase.
         | 
         | I'm annoyed that I'll probably have to pick a Framework 13 with
         | less CPU and much less GPU merely because of availability.
        
       | mumber_typhoon wrote:
       | I wonder if higher TDP is possible with framework desktop. That
       | one probably has much better cooling than these laptops with the
       | same chip and if numbers are different.
        
         | AmVess wrote:
         | I haven't tested the power draw, but I have the mainboard from
         | Framework that I put into a larger ITX case for better cooling.
         | 
         | My main PC is a 7950X3D which has the same core count/threads
         | as the Strix unit, and the Strix benches within margin of error
         | as the 7950X3D. Which is to say the performance is the same.
         | 
         | That you can get the same computer power in a laptop is crazy.
        
       | rcarmo wrote:
       | I would love to try out one of the mini-PCs that ship this, but
       | they seem to be made of either platinum (hugely overpriced in EU)
       | or unobtainium (no retailers carry them here, and getting
       | something direct from China is dicey warranty-wise). ROCm 7 looks
       | to be working already under most Linux distros and having this as
       | a workstation with a local LLM or a "home inference server" with
       | Ollama and a few services seems like a great solution.
        
         | dangus wrote:
         | ROCm is making great progress but I've had enough hiccups
         | (desktop with RX9070XT) that I'd still recommend those looking
         | for AI capability to continue using an Nvidia or Apple solution
         | for the time being.
         | 
         | Still, I think it'll be quite equivalent soon.
         | 
         | I think one of the best AI systems in terms of
         | price/performance is still just to build a desktop with dual
         | RTX 3090's (of course you'll need an board that supports dual
         | cards) and toss it in a closet.
        
           | Tuna-Fish wrote:
           | It depends on what you are doing. A lot of people who want to
           | do local inference want to do it using much larger models
           | than what can be fit onto a RTX3090, and Strix Halo is such a
           | hit because it gives you reasonable (not great, but good
           | enough to not be outright painful) performance with 128GB of
           | memory.
        
             | geerlingguy wrote:
             | Also, Vulkan is great, and much more stable. Plus tends to
             | work great for new, and even very old, graphics cards.
        
           | dismalaf wrote:
           | At this point Vulkan will just take over. AMD and Intel are
           | fumbling ROCm and SYCL, whereas Vulkan already ships nearly
           | everywhere.
        
           | almostgotcaught wrote:
           | > ROCm is making great progress
           | 
           | is the progress in the room with us?
        
         | cpburns2009 wrote:
         | Have you looked at Corsair's AI Workstation 300 Desktop PC? [1]
         | It's 2000-2700 EURO depending on model, and taking VAT into
         | consideration it's comparable to the 1700-2300 USD pretax
         | prices.
         | 
         | [1]: https://www.corsair.com/eu/en/c/ai-workstations
        
           | rcarmo wrote:
           | No, but it falls into the platinum side of the equation. I
           | can rent a cloud GPU for a few hours a month and come out
           | ahead.
        
             | overfeed wrote:
             | If the economics don't work out, perhaps this product is
             | not for you and you're better off renting.
        
             | adgjlsfhk1 wrote:
             | I don't think there's any computer hardware that is now
             | economical to buy and use a couple hours a month than to
             | rent
        
           | hau wrote:
           | I generally agree with gp. Checkout with your link says "This
           | item is currently on pre-order" btw. Retail mini-pcs are
           | somehow harder to obtain than general purpose ones.
        
         | mandelken wrote:
         | I ordered the framework desktop 395 - 128Gb edition for just
         | under 1900 eur. With some extras I paid just over 2000 incl
         | shipping to EU. Didn't feel overpriced to me.
        
           | erinnh wrote:
           | I looked just now and it cost 2500 euro without any storage.
           | 
           | Was it on sale or something?
        
             | mandelken wrote:
             | Huh, indeed, above 2300 eur now. I made a deposit earlier
             | this year and it shipped in August, didn't see the price
             | increased.
        
       | nexle wrote:
       | High Yield has a video that deep dive into the 395 chip on the
       | silicon level: https://youtu.be/maH6KZ0YkXU
        
       | InTheArena wrote:
       | I picked up a framework desktop and am running it through it's
       | paces right now. So far, it's a impressive little box. I'm really
       | hopeful that this continues to drive more and more enthusiast
       | support and engagement. Getting strong vulcan or rocm supported
       | infrastructure would be great for everyone.
        
       | runjake wrote:
       | Related question: Can I buy a desktop Zen 5 CPU and something
       | like an RX 7600 XT and some RAM and have a high shared memory
       | bandwidth situation between the system memory and the GPU ala
       | Strix Halo and Apple Silicon without spending a ton of money?
       | 
       | And get pretty reasonable local LLM performance on some of the
       | larger models for hobbyist use?
       | 
       | Edit: I don't have a good grasp on this but I'm thinking I can
       | only do shared memory when I'm using an APU and not a discrete
       | GPU. Is this correct?
        
         | Rohansi wrote:
         | No, memory is not "unified" when you have a physically separate
         | GPU. In that case memory is accessed through the PCIe bus which
         | will be a significant bandwidth bottleneck. PCIe tops out at
         | 64GB/s for 16 lanes of PCIe 5 and not all GPUs support that.
        
       | dzonga wrote:
       | how does the gpu compare though to the ones in m-series macs ?
        
       ___________________________________________________________________
       (page generated 2025-10-18 23:01 UTC)