[HN Gopher] New 2GB Raspberry Pi 5 has 33% smaller die, 30% idle...
___________________________________________________________________
New 2GB Raspberry Pi 5 has 33% smaller die, 30% idle power savings
Author : ingve
Score : 188 points
Date : 2024-08-29 15:54 UTC (7 hours ago)
(HTM) web link (www.jeffgeerling.com)
(TXT) w3m dump (www.jeffgeerling.com)
| GGO wrote:
| is there going to be D0 stepping for the 8GB version?
| jsheard wrote:
| I don't see why there wouldn't be, it's cheaper to manufacture
| with seemingly no downsides. They probably won't revise the 4GB
| and 8GB versions until their stocks of the original stepping
| are used up though, and once they do introduce revised versions
| it may be a lottery which version you get for a while.
| AstroJetson wrote:
| I'll bet they just slipstream them in. There was a huge
| backlog of the 8GB, that now looks pretty much cleared out.
| So it could be awhile before the D0 show up.
| hajile wrote:
| TSMC charged just a hair under $4000 per 16nm wafer in 2020.
|
| Wafer calculators at 0.2 defect/cm2 on a 300mm wafer gives 950
| fully-good dies out of 1061 for the old die (~89% good) and
| 1469 fully-good dies out of 1584 (~93%) for the new dies.
|
| Dividing that out gives $4.21/chip for the old chip and
| $2.72/chip for the new chip. At $80 for an 8gb board, that
| represents a ~1.9% increase in profit per board. For the $60
| 4gb version, it's more like 2.5% increase in profit per board.
|
| In real-world terms, if they sell 10M Pi5 units with the new
| chip, they'll have an extra $15M in the bank in saved
| production costs alone (minus whatever costs to strip
| everything out and tape out again). Furthermore, the new chip
| gets cheaper with every chip they make as the R&D costs get
| more and more diluted.
| rjsw wrote:
| You had speculated in a previous thread that the new die could
| result in power savings, good to see some real measurements.
| toastau wrote:
| Nice work!
| 0cf8612b2e1e wrote:
| Some of the power savings could be chalked up to less RAM,
| because more RAM requires more power. But that doesn't explain
| all the results.
|
| Is there a calculation to estimate RAM power consumption? I keep
| wanting to get a low powered N100, and have been wondering if I
| use say 8 vs 16GB RAM, would that make a measurable power
| difference?
| Mistletoe wrote:
| A small amount.
|
| > As a rule of thumb, you can expect to allocate around 3 watts
| of power for every 8GB of DDR3 or DDR4 memory.
| 0cf8612b2e1e wrote:
| Similar for DDR5?
|
| In the context of a N100, 3W is actually pretty high. Many of
| these systems idle around 10W.
| flyingpenguin wrote:
| Yeah but unless you are deploying a bunch of them, its only
| like $3/year each
| snovv_crash wrote:
| That's why many laptops run LPDDR these days.
| spockz wrote:
| My N100 from minisforum runs on 3.6W idle with 4-8W for
| medium to high cpu usage to 20W on full load with SSD, CPU,
| and integrated encoder loaded.
|
| It is way more powerful than a pi and in low power mode as
| efficient.
| fswd wrote:
| I probably have the same N100 miniforums and can only get
| it down to 7w. Is it possible you could share how you got
| it down to 3.6W and 4-8W?
|
| I just got this N100 a few weeks ago, have it on a
| killawatt next to my desk.
| spockz wrote:
| There are many tips here: https://gathering.tweakers.net/
| forum/list_messages/2096876 but I think the thing that
| did the trick was simply `powertop ---auto-tune`. I also
| disabled the wifi in the bios. I ran for a bit with turbo
| disabled which lowered the consumption a tiny bit and
| severely limited the peak usage, but it also made the
| system slow overall. So I enabled it again and now
| sometimes usage spikes but feels snappy. I'm running
| Fedora.
| mallets wrote:
| That's a ridiculous amount of power, is that answer from a
| LLM? It's in the range of ~150mW for LPDDR4. Around half that
| in idle.
| hnuser123456 wrote:
| You specified LP, they didn't.
| mallets wrote:
| It doesn't matter too much whether it's LP or not for
| idle, around 20% more at worst. It's at full BW util
| where it's close to double power consumption. That would
| still put it at less than half watt for full load.
|
| Look for "AM62x Power Consumption" app note, page 5.
| adgjlsfhk1 wrote:
| https://www.servethehome.com/ddr4-dimms-system-power-
| consump... which is the only source I've seen that looks even
| vaguely trustworthy gives 2 watts per 8 gb of ddr4 (and ddr5
| should be about 10% lower since it drops to 1.1 from 1.2
| volts assuming similar amounts of current).
| hajile wrote:
| Their numbers seem to indicate that Pi disabled parts of the
| chip in firmware instead of physically lasering them out. This
| leaves those parts still leaking some current and explains the
| power difference.
| moffkalast wrote:
| > 8 vs 16GB RAM, would that make a measurable power difference?
|
| I think the short answer is.. it depends. Current draw depends
| on usage, supposedly from 10% at idle up to full when doing
| massive read/writes. But it should be in the two to three digit
| miliwatt range which isn't much compared to the N100 itself
| which pulls 6W at idle, being an inefficient x86 space heater.
|
| The Pi 4 4GB can idle at 1W using LPDDR4, the Pi 5 8GB using
| 50% more efficient LPDDR4x idles at... 3W. Meanwhile the
| average 12GB LPDDR5X Android phone can idle a whole week on a
| tiny 1 cell lipo (with power saving mode on), making this look
| so bad it's actually funny.
|
| The long standing problem with Pi Foundation products is a
| complete disregard for any low power states, sleep or
| hibernation, so they probably don't do any RAM related power
| optimization either. It's only now with the RP2350 that they've
| finally implemented some kind of working sleep mode for the
| very first time in anything at all.
| Marsymars wrote:
| > N100 itself which pulls 6W at idle
|
| That's not right - my entire N305 system takes 4-5W at idle.
| moffkalast wrote:
| Fair point, N100 systems do differ quite a bit generally,
| the numbers I've got in my head are what I've seen people
| quote for the Radxa X4 recently. Some might be more, some
| less.
| quaintdev wrote:
| Genuine question what is the use case for such configuration I
| mean less RAM and powerful CPU?
|
| I use my Pi for self hosting so I need more memory and more CPU
| is always better for my case. If they need less power consumption
| then they could have used Pi 4 or other lower version.
|
| Besides low cost I don't see other advantage of such
| configuration. Please enlighten me.
| mlyle wrote:
| Pi 5's peak consumption is higher, but if you are doing the
| same amount of computational work (but a bit above idle) it is
| less.
|
| Now this uses even less, and even the idle powers are
| comparable to 4. This wins in every way over Pi4, except cost.
| tonymet wrote:
| Given people typically use these for a home server or
| embedded controller, who cares about saving a couple watts?
|
| 1 watt savings idle? .72 kwh a month? in my area that's about
| 5[?] , even in california that's 40[?] / month / device ?
|
| Are 1000-instance Pi arrays common? i can think of 100 better
| ways to get that amount of computing resources.
| byteknight wrote:
| I think you are applying your usecase and thinking that is
| everyone's use case. Think battery power.
| mlyle wrote:
| Yup. I had students running a machine learning model on a
| Pi attached to a scooter. The power draw of the computing
| system was significant, and the performance limitations
| were too.
|
| The existence of commodity, well-documented, well-
| supported SBCs offering more computing in less power is
| exciting.
|
| (Also, using less power means you can put it in more
| places without worrying about getting rid of the heat, or
| can go longer before throttling).
| tonymet wrote:
| as in portable or embedded? Embedded, Pi has too much
| consumption. saving a watt isn't your problem
| Aachen wrote:
| If your needs are greater than a Pi but not at the level where
| you want to be buying professional server hardware, I can
| recommend checking if you or a family member still has a laptop
| they're not using anymore. Free hardware that would otherwise
| end up going to waste, and anything that had good performance
| in 2012 is still faster than a new Pi today. The older, the
| more likely it is to have replaceable RAM so you can stick in
| 16GB easily (older RAM types are also cheaper)
| Max-q wrote:
| Maybe real time image processing from a camera. 2GB would
| probably be more than enough, and less power draw is nice when
| running continuously.
| justin66 wrote:
| For an embedded application involving signage or a kiosk, 2GB
| is probably plenty, and if you're doing enough of them the
| money saved would be worth it.
|
| For a home user with one, or only a handful of, machines I
| don't think you'd necessarily need to save a few bucks this
| way. Although _people obsessing over pennies_ is a theme in
| most Raspberry Pi threads, so there is clearly appeal for some.
| georgeecollins wrote:
| I don't know if this configuration works for me, but unlike a
| lot of people here I have used Raspberry Pi's in robots. The
| 5's power consumption is such that it is a problem. The jensen
| nano is no longer supported. So maybe there is a "thingaverse"
| need for something more powerful then a pico /arduino but not
| so power hungry.
| suprjami wrote:
| Broadcom pushed the old Raspberry Pi SoC architecture as far as
| it would go, but it's hard limited to 1Gb RAM.
|
| Raspberry Pi has become a development platform for Broadcom
| SoC, so they can sell the same chip in their corporate products
| and the cost of development is partially covered by Pi sales.
|
| They also have an educational non-profit organisation, but
| don't mistake the Raspberry Pi for a 100% selfless endeavour.
|
| Broadcom have been very hostile at taking Pi clones of the
| market. Look at that happened to ODroid's Zero clone with the
| same SoC as the Zero.
| PedroBatista wrote:
| More memory also needs more power.
|
| Not sure what percentage of those gains can be attributed to less
| memory and what can is due to better die/process.
| kobieps wrote:
| Is the firmware still closed source?
| written-beyond wrote:
| They've improved a lot, it's mostly the broadcom graphics
| firmware that still requires blobs. Probably because they've
| licensed the IP from somewhere else and the effort to reward
| ratio is too low.
| mort96 wrote:
| In other words: yes, the firmware is still closed source.
| autoexec wrote:
| I wouldn't mind them not using open firmware as much if we had
| unofficial alternatives that worked without sacrificing
| functionality, but I don't think anyone has managed that even
| after all this time.
| alxjsn wrote:
| I rather just buy used Lenovo Thinkcentere PCs on eBay. Way more
| power, cheaper and relatively small. There's a lot of different
| CPU/RAM/DISK configurations you can find.
|
| I've been buying these, throwing Fedora IoT, docker, and
| Tailscale on them and running them from different locations for
| personal projects.
| PedroBatista wrote:
| Sure, if you need a machine to do "normal" PC/server things,
| buying a computer would be the natural path.
| geerlingguy wrote:
| There's also the Radxa X4 -- Pi 4/5 form factor (tiny bit
| larger but not substantial), has a PoE HAT available, has
| 40-pin GPIO (with caveats, but it's basically a Pico strapped
| on the same board), and the N100's built-in GPU can run circles
| around the Pi or even RK3588 boards.
|
| The efficiency isn't there, and you'll need to figure out a
| better cooling/case solution than the one Radxa ships, but I'm
| impressed by this little board.
|
| If you can stretch your budget past $100 you can get a good
| brand new N100 or N305 system that will go further. Used gear
| is fine, but the power efficiency for anything in the $50-80
| range used is pretty rough. Some people don't worry much about
| that, but in some parts of the world it can be $5+/month more
| to run older machines!
| throwup238 wrote:
| _> Some people don 't worry much about that, but in some
| parts of the world it can be $5+/month more to run older
| machines!_
|
| California being one of the worst offenders, ironically. In
| SoCal my family and I pay (across several households) between
| $0.50/kWh and $0.99/kWh, so even a 15W idle can cost us at
| least $5 a month.
| bri3d wrote:
| A few notes (as someone who is pretty staunchly anti-Pi-as-a-
| server, I end up having this debate often, and I do think there
| are reasons to do both):
|
| * A Pi will sit much lower in total power consumption than
| almost any used PCs if both are doing effectively nothing (ie -
| simple, spiky tasks like filtering DNS, serving static content
| from RAM, etc.). You need to be doing something with the system
| before a PC server comes out ahead, and most people using a Pi
| as a home server... aren't.
|
| Compared to a modern low-power x86 PC system, the difference
| isn't meaningful, but if you're buying used stuff 3 generations
| back, the difference becomes somewhat meaningful in terms of
| electric cost (on the order of tens of dollars per year, which
| is significant for hardware which cost tens of dollars to start
| with).
|
| * The Pi of course has GPIO, SPI, etc. exposed, so you can use
| it as a nice "hybrid-IoT" device where it's a home server _and_
| a sensor aggregator, for example. And the hat ecosystem, while
| generally insanely overpriced, is convenient.
|
| Now, the moment you're running K8s/Docker or a real compute
| workload (security camera image recognition, etc.) you should
| probably move off of the Pi and onto something nicer, indeed. I
| absolutely never understood people running clusters of Pis or
| those goofy multi-Pi carrier boards. Just buy a real PC.
| Aachen wrote:
| > A Pi will sit much lower in total power consumption than
| almost any used PC
|
| For electricity consumption, beyond the wallet, it actually
| seems that hardware should have a lifespan on the order of
| decades before electricity consumption savings offset the
| environmental impact:
|
| "For laptops and similar computers, manufacturing,
| distribution and disposal account for 52% of their Global
| Warming Potential (i.e. the amount of CO2-equivalent
| emissions caused). For mobile phones, this is 72%. The report
| calculates that the lifetime of these devices should be at
| least 25 years to limit their Global Warming Potential."
| --https://wimvanderbauwhede.codeberg.page/articles/frugal-
| comp...
|
| Rather than buying a new Pi, repurposing a 5-year-old laptop
| has advantages if this something one cares about. Desktops
| are quite a bit more hungry (I've heard this got better in
| recent years), but I can attest that a 2012 laptop still
| functions very well as a server, easily better than a
| 2024-era Pi. Probably I'll replace it in the next 2-4 years
| (so at ~15yo) when my current laptop finally will have given
| me enough grief (my inner grandpa complains they don't make
| 'em like they used to), and I'm not saying others must
| optimise for climate alone either, but it's something to
| consider when deciding on a good balance
| k_bx wrote:
| Environmental impact is not the main concern when it comes
| to power consumption. Main concern is how long will it work
| on a battery before a maintenance person can come by and
| switch the battery or if the sun will come out and start
| charging it again.
| sofixa wrote:
| > I absolutely never understood people running clusters of
| Pis or those goofy multi-Pi carrier boards. Just buy a real
| PC.
|
| A bunch of Pis allow you to run multi-node clusters on the
| cheap. If you're just experimenting with
| Kubernetes/Nomad/whatever, you don't need a lot of resources,
| just multiple nodes. It's easier and depending on config
| potentially cheaper than getting a beefier mini PC, throwing
| lots of RAM, and running VMs.
| AshamedCaptain wrote:
| > * A Pi will sit much lower in total power consumption than
| almost any used PCs if both are doing effectively nothing (ie
| - simple, spiky tasks like filtering DNS, serving static
| content from RAM, etc.). You need to be doing something with
| the system before a PC server comes out ahead, and most
| people using a Pi as a home server... aren't.
|
| That's not necessarily true since the Pis are particularly
| terrible at idle power consumption. E.g. the "power off"
| state consumption shown in the article is actually higher
| than the idle consumption of some low-power Atom/Celeron x86
| chips. The Pi is just terrible at power management.
| suprjami wrote:
| Citation needed?
|
| Pi 2 and 3 typically sit at 200 mAh and 230 mAh, Pi 4 is
| not far away. Zero 2W can go down to 96 mAh.
|
| https://www.jeffgeerling.com/blogs/jeff-
| geerling/raspberry-p...
|
| https://www.jeffgeerling.com/blog/2021/disabling-cores-
| reduc...
|
| I don't see any x86 system approaching those numbers.
| supertrope wrote:
| To illustrate a Lenovo Thinkcentre m720s idles at 9W! Intel
| i5-8500 (6C/6T), 8GB DDR4, 256GB NVMe SSD. It cost $120.
| ssl-3 wrote:
| In order to begin to make a valid price comparison to this
| used (ie, not new) PC, we'd also need to know how much a
| used Raspberry Pi system would cost.
| supertrope wrote:
| Go to ebay.com. Search for Raspberry Pi 5 or 4. See
| actual market prices by filtering to sold listings. Add
| cost of case, power supply, storage, maybe even active
| cooler.
| jchw wrote:
| Emphasis on way more _power_. Granted, power draw isn 't a huge
| problem for price reasons, since the cost of electricity is
| _usually_ not that huge of a factor, but if you wind up running
| a lot of these, it can add up. More power draw is also a
| detriment if you want to keep things lean for longer battery,
| which makes me hesitant to put many older stock computers in
| the critical path of my network.
|
| Don't get me wrong, though. Old stock computers are excellent
| for a wide variety of tasks, it's just that they definitely
| don't encroach on a lot of the use cases of modern SBCs. You
| needn't buy a Raspberry Pi 5 either; plenty of use cases like
| Home Assistant will run pretty well on a Pi 4 or even a Pi 3,
| and that's not getting into the many other reasons why a Pi may
| be interesting (like HATs, being able to use PoE power, GPIO,
| or even just the I/O in general.)
| autoexec wrote:
| I can't bring myself to buy Lenovo anything since they've
| repeatedly shipped products infested with malware, sometimes
| doing it in exchange for money. Once a company sinks that low
| and treats users with such disrespect I don't know how or why
| they should be trusted again.
| gosub100 wrote:
| I tried buying a used one on eBay and was disappointed to
| find it used a proprietary SATA power cable or connector,I
| can't recall. I hope they stopped doing that.
| sgt wrote:
| The ThinkCentres kinda look like "evil" IBM corporate
| machines out of a Star Wars movie. I wish I could put an IBM
| logo on mine to complete the picture.
| sgt wrote:
| I just bought one! It's brilliant, I had no idea. Cost almost
| nothing, and I removed the HDD and put an SSD in there, extra
| RAM and it's lightning quick.
|
| Right now I have Windows 10 on it as I needed to run some old
| proprietary software, but it's a proper gem of a machine. I got
| the one without touch screen.
| iwontberude wrote:
| Have 13 of them and I love them
| Rebelgecko wrote:
| What's the break even point for the way more power vs the
| cheaper upfront cost? My raspberry pi has probably saved
| hundreds of dollars over the years in electricity vs what I'd
| have spent running a cheap Intel box
| pseudosavant wrote:
| If they are using a standard 300mm wafer and have similar yields,
| the D0 stepping will allow them to get ~50% more chips per wafer
| than the C1 stepping.
| hcfman wrote:
| Nice comparison. Looking forward to the 4nm version
___________________________________________________________________
(page generated 2024-08-29 23:01 UTC)