Silicon lottery.
Now that the Raspberry Pi 5s been readily available (at least in most regions) for a few months, more people started messing with clocks, trying to get the most speed possible out of their Pi 5s.
Unlike the Pi 4, the Pi 5 is typically comfortable at 2.6 or even 2.8 GHz, and some Pi 5s can hit 3.0 GHz (but no higher—more on why tomorrow well... this limit may be able to be lifted).
After some testing, I found the default 2.4 GHz clock on the Pi 5 is pretty much the efficiency sweet spot, and after a lot more testing recently, I can confirm that's still the case, testing a number of Pi 5 samples.
Also unlike the Pi 4, the Pi 5's not very picky about cooling. While the chip runs a little hotter, it doesn't need exotic cooling like spraying ice, water cooling, or LN2, to run at its highest possible clock speeds. It does need some cooling, but I've found the biggest factor affecting how fast your Pi 5 will go is the chip itself.
The silicon lottery basically states that among samples of the exact same silicon chip—in this case the BCM2712 at the heart of the Pi 5—there are variations affecting performance, thermals, etc.
These variations are almost imperceptible at the 2.4 GHz default clock Raspberry Pi chose for the board, but they can affect how fast your Pi 5 can go, no matter what cooling solution you use.
To prove that, over the past 5 months, I've slowly acquired 10 Pi 5s from vendors around the world (1 at a time), to make sure they come from different batches, and I've been testing overclocking performance on each. I also had another project slated for a number of these Pis (which I'll reveal tomorrow), but one thing I wanted to quantify is how many of the Pi 5s would hit 3.0 GHz.
And the answer? Out of the ten I bought, only one! It was an 8GB model (I did buy a few 4GB Pi 5s as well, since the performance characteristics can be slightly different, and it runs at 3 GHz reliably with any form of active cooling.
I've tested it with:
- Raspberry Pi's own Active Cooler
- Argon THRML (a giant tower cooler for the Pi 5)
- Argon Neo (a case with a larger heatsink and PWM fan for the Pi 5)
- EDATEC Pi 5 Fanless Case (a top + bottom heatsink cover for the Pi 5)
And in all cases, it worked fine.
I tested my other Pis with all those cooling solutions, and even tried a couple on my massively-overkill water cooling setup, and no matter what, none of them reached 3.0 GHz. The closest I got was 2.9 GHz on a couple of them. The rest maxed out at 2.8 GHz.
The result of that 3.0 GHz overclock? A marginally-improved Geekbench 6 score of 1662, versus 1507 with no OC. To achieve that 10% speedup, it ate up about 20% more power, so efficiency-wise, it's not worth it.
Also, if you're only concerned about raw performance and efficiency, an RK3588-based SBC might be the better option regardless—it gives almost double the maximum performance, using about the same power! (notwithstanding software issues and vendor / community support.)
Bottom line: not every Pi 5 can achieve a 3.0 GHz overclock—in fact, in my testing, most can't. 2.4 GHz is the most efficient clock for the C0 BCM2712 silicon running on the Pi 5, at least based on my testing.
More on the Pi 5's silicon coming tomorrow—which happens to be Pi Day :)
Comments
Regarding your recommendation for the RK3588's: basically, I'd be with you there - if there wasn't that pesky "software" and "support" thing you mentioned: I still remember reading about that fun game being a thing, a while back, of "distribution for board variant X doesn't run on board variant Y, despite both (supposedly) running on the same chip and chipset" - not worth the hassle, IMHO...
Plus, that "double the performance" side comes at about triple the price (at least here in Europe), and for way less than that you can get a bunch of X86 mini PCs, that don't draw any more power either, but eat the RK's for breakfast, performance wise (okay, granted, that NPU is a thing on the 3588, but is there any useful support for that, on stuff mere mortals would want to use? For example, Ollama sounds kinda interesting ("run a ChatGPT-like, but locally"), but as CPU-only, it feels quite slow 🐌 on a Pi5 (check Kevin McAleer's video 😉), and as I checked a few days ago, not even a Google Coral module can help you with that, due to "no support", probably much less that NPU on the RK, and CPU-wise the RK isn't _that_ much faster to get it to run at more practicable speeds...)
Using an RK3588 NPU for Frigate is about the best use case I can think of—something I'm going to try out soon! But you're right, most of the RK3588 boards were a better value before the glut of N100 mini PCs came out in similar price ranges, and before the Pi shortage was over.
Coral TPU is still great for that, and finally in stock some places again, but it would be cool if there were more options for 'drop in inference' on SBCs that weren't too expensive, and supported decently.
It would be interesting to know where the TSMC process was running the day the produced that magic 1 Pi 5. My company always has TSMC run special lots which are intentionally shift toward the edge of their process just to verify our chip meets its specs over the entire range of TSMC processes.
Did you try downclocking to confirm that a lower GHz might not be even more efficient?
Yes; in the article I linked in this post on under/overclocking, I clocked down from 2.2, to 2.4, to 2.6 GHz, and separately have run tests at 2.8 and 3.0 GHz.
The term "Silicon Lottery" is apt, but hasn't the Silicon Lottery always been what determines the success of out-of-the-box overclocking? (Out-of-the-box meaning foregoing any additional cooling solutions.) I'm curious though what the overclocking results may be with using superceded water.
It is strange, because 3 of 3 I have, one is reaching 2.9GHz, the other two can do 3GHz without problem. All without over_voltage_delta.
All 4GB versions. Bought them last December in Germany.
Sold the first one, the second one can do stable 3.15GHz/1.15GHz GPU with over_voltage_delta=5000, the other one - 3225MHz/1.15GHz GPU stable with the same over_voltage (1.0V)