Hacker News Reader

Comment by mvanaltvorst | Hacker News Reader

If you take this CPU with its 280W TDP and scale it down to 18W like an Apple M1, you get a score of ~7200 while the Apple M1 gets 7700. Actually pretty amazing how close the performance per watt is to be honest; totally different form factors, different CPU architecture, different R&D (both having spent billions of dollars), different fabs, but still within 10% performance/watt. Makes you wonder how close we are to the physical limits.

View on HN

Yeah, but don't forget that creating systems with massive core count is much more than just scaling up. It is a whole different science and art. Just handling the heat generated by 280W requires teams of hundreds of engineers. And IPC for 8 cores is child play compared to IPC for 64 cores.

You can already see this with M1: the 10W M1 reaches 7.5K while the 18W (15?) M1 is IIRC around 8.7K on cinebench

Anandtech says during their testing that pretty much all the chips were in the 220w range.

At 18w and 7,833 points, that's 435 per watt for the M1

At 440w for 2 chips and 113,631 points, that's 258 per watt for the Milan

At around 140w and 10,096, that's about 74 per watt for the 5950x

At around 75w and 4,517, that's about 61 per watt for the 5600x

At around 140w and 5,994, that's about 43 per watt for the 5800x

Frequency scaling is TERRIBLE for PPW (performance per watt). The screaming almost 5GHz 5800 cores lose efficiency to the downclocking 5950 cores by almost 2x. Vs the 2.5GHz Milan, it's right at 6x difference.

Apple is on the right track pushing wider rather than faster, but that seems like a path that is hard for x86 to take due to decoder power and complexity (decode units literally eat up more die space than the actual integer math units). It isn't all roses though. A huge portion (30-50%) of the Milan power budget is the interconnect. If you double the TDP on the M1 to simulate the interconnect, PPW drops to around 217 which is about 16% worse than AMD.

We also haven't factored in how AMD gets a huge performance boost with a quarter-gig of cache. Apple likely has to scale up their cache too. 16mb and 8 cores is 2mb/core. 256mb/64 is 4mb/core. But because it's multicore and all cores don't necessarily do the same job at the same time on a server, M1's deal with the devil starts to fade. It will need even more cache per core to keep up with the ultra-wide execution path (else lose PPW) which will increase total power and further reduce max clockspeeds. M1 is a great chip in laptops, but if scaling up were easy, everyone would be doing it.

https://www.anandtech.com/show/16529/amd-epyc-milan-review/5

https://www.anandtech.com/show/16252/mac-mini-apple-m1-teste...

https://www.anandtech.com/show/16214/amd-zen-3-ryzen-deep-di...

Keep in mind that every vendor defines TDP differently, linustechtips have highlighted this many times and even has a video on it [1].

So comparing TDP across manufactures is not a fair comparison.

[1] https://www.youtube.com/watch?v=yDWO177BjZY

Probably large chunk of that difference is consumed by interconnect (infinity fabric)

Reply on news.ycombinator.com