3 ms·

I may be showing ignorance of some important use case here, but I'd put AVX512 in the "junkware feature" category. It can run AI/HPC code faster than the main p

I may be showing ignorance of some important use case here, but I'd put AVX512 in the "junkware feature" category. It can run AI/HPC code faster than the main processor, but still so much slower than a GPU as to be pointless; when it does run, it lowers the frequency and grinds the rest of the processor to a halt until it eventually overheats anyway. Maybe, maybe it has a place in their big desktop cores, but why would anyone want it in a laptop? You can't even use it in a modern ultrabook-style casing without a thermal shutdown and probably first-degree burns, without even mentioning the battery drain.

PaulHoule 4y ago
Yes and no. If you have all the cores running hard the power consumption goes up a lot but if it is just one core it won't go up too much. If worse come to worse you can throttle the clock. In general it's a big problem with SIMD instructions that they aren't compatible across generations of microprocessors. It's not a problem for a company like Facebook that buys 20,000 of the same server but normal firms avoid using SIMD entirely or they use SIMD that is many years out of date. You see strange things like Safari not supporting WebP images on a 2013 Mac while Firefox supports them just fine because Apple wants to use SIMD acceleration and they'd be happier if you replaced you 2013 Mac with a new one. I worked on a semantic search engine that used an autoencoder neural network that was made just before GPU neural networks hit it big and we wrote the core of our implementation in assembly language using one particular version of AVX. We had to do all the derivatives by hand and code them up in assembly language. By the time the product shipped we bought new servers that supported a new version of AVX that might have run twice as fast but we had no intention of rewriting that code and testing it. Most organizations don't want to go through the hassle of keeping up with the latest SIMD flavor of the month so a lot of performance is just left on the table. Intel is happy because their marketing materials can tell you how awesome the processor is but people in real life don't experience that performance.
TinkersW 4y ago
Slowdown/overheating are not inherit to AVX512, they are specific to older implementations. And no the GPU is not a replacement, the latency difference between GPU and AVX512 is off the charts.
- Miraste 4y ago
  Okay, there's a use case I didn't know about. Do you know any SIMD projects that require the lower latency? I'm glad to hear some newer AVX512 implementations are better. I haven't used an Intel one for years, but I have used the current-gen Ryzen version. Unfortunately it's still suffering the same problems-last time I stress-tested it my CPU hit 100C and some absurd power number before I was too scared to continue. Granted, that was with Prime95, which I believe is close to peak utilization.

PaulHoule 4y ago

Yes and no. If you have all the cores running hard the power consumption goes up a lot but if it is just one core it won't go up too much. If worse come to worse you can throttle the clock. In general it's a big problem with SIMD instructions that they aren't compatible across generations of microprocessors. It's not a problem for a company like Facebook that buys 20,000 of the same server but normal firms avoid using SIMD entirely or they use SIMD that is many years out of date. You see strange things like Safari not supporting WebP images on a 2013 Mac while Firefox supports them just fine because Apple wants to use SIMD acceleration and they'd be happier if you replaced you 2013 Mac with a new one. I worked on a semantic search engine that used an autoencoder neural network that was made just before GPU neural networks hit it big and we wrote the core of our implementation in assembly language using one particular version of AVX. We had to do all the derivatives by hand and code them up in assembly language. By the time the product shipped we bought new servers that supported a new version of AVX that might have run twice as fast but we had no intention of rewriting that code and testing it. Most organizations don't want to go through the hassle of keeping up with the latest SIMD flavor of the month so a lot of performance is just left on the table. Intel is happy because their marketing materials can tell you how awesome the processor is but people in real life don't experience that performance.

TinkersW 4y ago

Slowdown/overheating are not inherit to AVX512, they are specific to older implementations. And no the GPU is not a replacement, the latency difference between GPU and AVX512 is off the charts.

Miraste 4y ago
Okay, there's a use case I didn't know about. Do you know any SIMD projects that require the lower latency? I'm glad to hear some newer AVX512 implementations are better. I haven't used an Intel one for years, but I have used the current-gen Ryzen version. Unfortunately it's still suffering the same problems-last time I stress-tested it my CPU hit 100C and some absurd power number before I was too scared to continue. Granted, that was with Prime95, which I believe is close to peak utilization.