ISTR seeing some benchmarks where using AVX512 really hurts your throughput if you're running multithreaded code. Again, vague memory, but the reason was that it winds up heating up the core that's running the AVX512 code and causing all the other cores to thermally throttle.

Yes, exercising AVX512 on intel (very much at all) thermal throttles in most workloads.

Avx512 is very new, especially to the server market. There likely aren't many people absolutely relying on it yet.

Isn't Zen 2 adding support for AVX512?

Yes, but a slow one. They just bump AVX units from 128-bit to 256-bit, making it Haswell-style. Skylake-X has a full 512-bit wide AVX unit.

Doesn't Skylake-X have the whole downclocking issue when running AVX-512 workloads though? Unless you can use almost exclusively AVX-512 then it doesn't make sense.

https://blog.cloudflare.com/on-the-dangers-of-intels-frequen...

AVX and AVX512 both do downclocking. Likely AMD will have to downclock too, pretty fundamental limits at work there.

AMD verified they would be running AVX2 at full clocks. Doing AVX-512 would happen at half-rate, but the processor would continue at full speed. In contrast, skylake downclocks base clocks by 30+%.

Not only does the AVX calculation downclock, but all the other calculations in other ALUs/FPUs also downclock. For pure AVX-512 workloads, Intel would be a little faster, but for mixed loads, AMD should be much faster overall.

AMD says they don't downclock AVX2 (about 2/5 down the page)

https://www.anandtech.com/show/13578/naples-rome-milan-zen-4...

See page 13 for AVX clockspeeds

https://www.intel.com/content/dam/www/public/us/en/documents...

With 7nm they might not need to downclock that much.

AVX-512 isn't just a wider vector, it also fills in some holes in the AVX instruction sets. It's still useful to support it, even if they are only doing it at AVX2 throughput levels.

Citation please. I never heard of Zen2 supporting AVX512.