At 27W Apples M1 scores 286/W in CBR23. The 3600X scores 281/W at 28.5W. Let's talk

Apple has chosen to use performance per watt as their benchmark for comparing their processors in laptops and small desktops. So, lets take a close look at how their processors compare with other desktop chips not intended to compete in low-power environments.

As we've seen with AMD and overclocking, you can get a few percent increase in performance for massive increases in heat. That's because P = I² x R, power is proportional to current squared x resistance (think voltage).

Current, at constant voltage, is roughly proportionally to frequency, so that your score goes up roughly with the square of power. But it's actually worse. As you increase frequency you need to increase voltage for stability, which makes the Performance vs power even worse.

And this is where Apple decided to "make magic". By limiting the performance of their computers to around a Cinebench score of 7500 with processors around 3.3Ghz, they are maximizing their performance per Watt. If they are using modern manufacturing they should simply be able to dial back the score until they get a specific Score/W. With a 5nm manufacturing processes they should even be able to BEAT other companies.

So can they?

Without a 5000 series Ryzen, or even a 4000 series mobile processor, I decided to play Apples game. I wouldn't go for the highest score, but the highest performance per Watt. Well, I would try to match their score at a similar power using only a 6-core machine. (Performance per watt goes as sqrt(Processor Count) so expect a 15% increase for an 8-core.)

To insure an ... Apples to apples comparison, I followed the exact same method found at https://www.anandtech.com/show/16252/mac-mini-apple-m1-tested I measured the idle processor+chipset power, and subtracted that from the full load. They seem to have constrained the M1 chipset to 27W.

At that point I began to adjust my power settings using the Ryzen Master to limit clock speed and voltage, targeting both a score of 7700 and a wattage of 27.

At first I disabled mult-threading, but found it significantly increased performance per WATT. I didn't want to spend more than 20 minutes on this, so I quickly found a frequency that got me 7500 points, about 3.3GHz, then I decreased the voltage. I never actually crashed the computer or found it unstable, reaching 0.925V. At this point I locked in the voltage and began increasing clock speeds, settling on a combination of 3.5 and 3.6GHz.

Final scores.

3600X @ 3.5/3.6Ghz @0.925V @ 3.2GHz DDR4 = 28.5W and 8035 = 281/W
AppleM1 @ 3.2/2.4GHz @ Firestorm memory = 27.0W and 7780 = 288/W

You may ask: Why did Apple stop there? Well, a good score is nice, and there is a minimum voltage for a chip to operate. As long as that voltage is met, you get good performance/Watt.

Notes: Architectures between Ryzen 2 and M1 are 7 and 5nm, but I expect most difference to be due to memory. Ryzen 3 have shown around 25% increase in performance with a significant decrease in power. So it is all but assured that a 5600X would CRUSH the M1 in performance power Watt. Also, remember we expect a 15% improvement going from 6 to 8 cores.

I would say that the differences are in the noise and that Apple has made a SOC that is commensurate with, not exceeding, current standards. The fact that a desktop processor easily matches Apples low power performance-watt ratio is interesting. It will be interesting if Apple tries to compete in the higher powered areas, or just adds more processors (which is great for computer benchmark scores, but not always great for real applications)

That is all.

Submitted December 05, 2020 at 06:08PM by WSL_subreddit_mod https://ift.tt/36Ht9JQ via TikTokTikk