Notes on new 2 GB Raspberry Pi 5 D0 stepping

[geerlingguy]

The new Pi 5 has a D0 stepping, which supposedly has a smaller die (for better cost per chip) with some of the 'dark silicon' Broadcom used for other purposes removed.

Right now it seems to only be available on the Pi 5 2GB, so I ordered a few samples and am testing them...

I have found a few things already:

• Geekbench 6 won't complete (even at default clocks) as the OOM kills it when it runs the Photo Filter multicore test.
• I can get one of my samples to run somewhat stably up to 3.5 GHz unmodified (with this firmware hack), but 3.6 GHz results in a number of kernel memory errors, and ultimately a kernel panic. (This is with a heatsink underneath and peltier on top).

[geerlingguy]

CPU die dimensions are 6.30mm x 5.98mm (37.674mm²)

Old C0 stepping that I have on hand measured 6.47 x 8.63mm (55.836mm²)

The result is about a 32.5% decrease in die area, quite respectable!

[geerlingguy]

CNX Software posted: Comparison of Raspberry Pi 5 with 2GB and 8GB RAM – Hardware, benchmarks, and power consumption.

	Raspberry Pi 5 2GB	Raspberry Pi 5 8GB
Power off	1.8 Watts	2.7 Watts
Idle (ondemand)	2.7 Watts	3.5 Watts
Idle (performance)	4.4 Watts	5.5 Watts

A few notes:

• I did observe idle 2.4W on the Pi 5 2GB I was testing. Didn't do much comparison on thermals yet. 3.2W on the Pi 5 4GB, need to re-test though with controlled conditions.
• It would make sense for a 2GB Pi 5 to use less power while turned on, as it
• The power off power can be much lower if you use POWER_OFF_ON_HALT=1 like I mention in the blog post Reducing Raspberry Pi 5's power consumption by 140x
• @ThomasKaiser points out the OPP tables are different, with the 2GB Pi D0 SoC getting lower voltages for given frequencies beyond the base clock (so 1.6 GHz and beyond).

[geerlingguy]

Some of my own testing:

	Pi 5 2GB D0	Pi 5 4GB C1	4GB Delta	Pi 5 8GB C1	8GB Delta
Idle power	2.4W	3.3W	0.9W (+32%)	3.2W	0.8W (+29%)
Idle temp	30°C	32°C	2°C (+6%)	32°C	2°C (+6%)
stress-ng power	8.9W	9.8W	0.9W (+10%)	9.8W	0.9W (+10%)
stress-ng temp	59°C	63°C	4°C (+7%)	64°C	5°C (+8%)

All these tests were performed in as identical a physical environment as possible:

• Ambient temp around 74°F/23.3°C (within 1°C)
• Pi Active cooler mounted (but fan not plugged in)
• Externally-powered Noctua fan blowing over the entire setup
• Pi mounted to an acrylic case with standoffs for airflow above and below the board
• Same Pi OS install on same microSD card, running Pi OS 64-bit Lite with only wired Ethernet plugged in
• 1 minute of idle time prior to running my stress/temperature benchmark each run

Sample of the 4GB Pi 5 C1 power consumption over time:

[geerlingguy]

I'm running another HPL benchmark to get energy efficiency: https://github.com/geerlingguy/top500-benchmark/issues/40

Because of the lack of RAM (my main theory), the 2GB Pi 5 gets a much lower score than the 8GB Pi 5, and even with the smaller chip efficiency isn't as great: 24.213 Gflops at 11.7W, for 2.07 Gflops/W

That's compared to 30.249 Gflops / 2.75 Gflops/W with the 8GB model.

Just reinforcing the idea that if your use case is memory-constrained, the 2GB model is a poor choice.

[ThomasKaiser]

0.2W (+10%)

Should read "0.9W (+10%)" both times :)

[geerlingguy]

@ThomasKaiser - Ah, drat, formatting markdown tables is always fun :) Fixed it!

[geerlingguy]

I'm re-running Geekbench 6 with swap size 1024 MB instead of the default 200 MB, just to see if it completes...

Got 893 / 1557, which is slightly faster single core, a good deal slower multi-core (probably because of swap + memory limitations): https://browser.geekbench.com/v6/cpu/7519453

[ThomasKaiser]

probably because of swap

sbc-bench -G would have shown the full story since monitoring for swapping and the other stuff ruining benchmark scores.

Or firing up sbc-bench -R and then once it's ready to monitor other benchmark executions running Geekbench in another Terminal. When finished benchmarking and stopping sbc-bench via [ctrl]-[c] it will tell whether swapping, throttling, frequency capping (and so forth) happened while the other benchmark has been executed :)

[geerlingguy]

I think my preliminary exploration is done, will be posting this on my blog/YouTube soon (hopefully Thurs/Fri, as long as baby doesn't arrive before I can post it lol).