7 ms·

New 2GB Raspberry Pi 5 has 33% smaller die, 30% idle power savings

214 points by ingve 2y ago

GGO 2y ago
is there going to be D0 stepping for the 8GB version?
- jsheard 2y ago
  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 2y ago
    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 2y ago
  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 2y ago
You had speculated in a previous thread that the new die could result in power savings, good to see some real measurements.
toastau 2y ago
Nice work!
0cf8612b2e1e 2y ago
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 2y ago
  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 2y ago
    Similar for DDR5? In the context of a N100, 3W is actually pretty high. Many of these systems idle around 10W.
    - flyingpenguin 2y ago
      Yeah but unless you are deploying a bunch of them, its only like $3/year each
    - snovv_crash 2y ago
      That's why many laptops run LPDDR these days.
    - spockz 2y ago
      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 2y ago
        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.
quaintdev 2y ago
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 2y ago
  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 2y ago
    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 2y ago
      I think you are applying your usecase and thinking that is everyone's use case. Think battery power.
      - mlyle 2y ago
        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).
PedroBatista 2y ago
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 2y ago
Is the firmware still closed source?
- written-beyond 2y ago
  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 2y ago
    In other words: yes, the firmware is still closed source.
    - moffkalast 2y ago
      They do seem to be experimenting with RISC-V now at least, the Pi 6 might be completely open when it launches in probably a decade or something like that.
- autoexec 2y ago
  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 2y ago
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 2y ago
  Sure, if you need a machine to do "normal" PC/server things, buying a computer would be the natural path.
- geerlingguy 2y ago
  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 2y ago
    > 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.
  - wpm 2y ago
    I think your thermal pad issue was why yours ran so hot, I got one recently in the second batch and I so far haven’t been able to get it over 65C. The pad I got was about 1-1.25mm thick, not crumbly like yours was on video. I assume they figured out what the issue was with their supplier. Still think the entire thermal design is whack, a single fan perpendicular to the board on the end could draw air through the heat sink as well as over the SSD on the top. Working on whipping up some 3D printable stuff once I get some time next week.
pseudosavant 2y ago
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 2y ago
Nice comparison. Looking forward to the 4nm version
spintin 2y ago
[dead]
ksec 2y ago
Again. The 33% smaller die and cost savings ( or at least room for Broadcom to charge them a little less ) is minuscule compared to the cost of DRAM. While LPDDR are slowly approaching DDR5's level, we still need to find way to lower cost. It doesn't seems there are any breakthrough or idea to push production cost of DRAM down to $1/GB.
rldjbpin 2y ago
from the reading, my understanding is that the "dark silicon" was nothing but properly functional hardware which was redundant for use with rpi (e.g. ethernet controller). if the makers knew what would not be required, why did they wait until the revision to make this change? the supply/demand argument does not make much sense to me especially given the surge in demand they've seen over the year, especially from the commercial arm.

214 points by ingve 2y ago

GGO 2y ago

is there going to be D0 stepping for the 8GB version?

jsheard 2y ago
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 2y ago
  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 2y ago
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 2y ago

You had speculated in a previous thread that the new die could result in power savings, good to see some real measurements.

toastau 2y ago

Nice work!

0cf8612b2e1e 2y ago

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 2y ago
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 2y ago
  Similar for DDR5? In the context of a N100, 3W is actually pretty high. Many of these systems idle around 10W.
  - flyingpenguin 2y ago
    Yeah but unless you are deploying a bunch of them, its only like $3/year each
  - snovv_crash 2y ago
    That's why many laptops run LPDDR these days.
  - spockz 2y ago
    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 2y ago
      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.

quaintdev 2y ago

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 2y ago
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 2y ago
  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 2y ago
    I think you are applying your usecase and thinking that is everyone's use case. Think battery power.
    - mlyle 2y ago
      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).

PedroBatista 2y ago

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 2y ago

Is the firmware still closed source?

written-beyond 2y ago
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 2y ago
  In other words: yes, the firmware is still closed source.
  - moffkalast 2y ago
    They do seem to be experimenting with RISC-V now at least, the Pi 6 might be completely open when it launches in probably a decade or something like that.
autoexec 2y ago
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 2y ago

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 2y ago
Sure, if you need a machine to do "normal" PC/server things, buying a computer would be the natural path.
geerlingguy 2y ago
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 2y ago
  > 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.
- wpm 2y ago
  I think your thermal pad issue was why yours ran so hot, I got one recently in the second batch and I so far haven’t been able to get it over 65C. The pad I got was about 1-1.25mm thick, not crumbly like yours was on video. I assume they figured out what the issue was with their supplier. Still think the entire thermal design is whack, a single fan perpendicular to the board on the end could draw air through the heat sink as well as over the SSD on the top. Working on whipping up some 3D printable stuff once I get some time next week.

pseudosavant 2y ago

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 2y ago

Nice comparison. Looking forward to the 4nm version

spintin 2y ago

[dead]

ksec 2y ago

Again. The 33% smaller die and cost savings ( or at least room for Broadcom to charge them a little less ) is minuscule compared to the cost of DRAM. While LPDDR are slowly approaching DDR5's level, we still need to find way to lower cost. It doesn't seems there are any breakthrough or idea to push production cost of DRAM down to $1/GB.

rldjbpin 2y ago

from the reading, my understanding is that the "dark silicon" was nothing but properly functional hardware which was redundant for use with rpi (e.g. ethernet controller). if the makers knew what would not be required, why did they wait until the revision to make this change? the supply/demand argument does not make much sense to me especially given the surge in demand they've seen over the year, especially from the commercial arm.