anon.1 (abc.delete@this.def.com) on October 31, 2020 8:06 am wrote:
> >
> >
> >
>
> Are you just seeing the effect of an effectively larger cache
> or do you know that these benchmarks are 100% compute bound?


We do not know the details of the benchmarks used by SiSoftware, but they appear to be microbenchmarks, which should fit inside the caches of any of the tested CPUs.

Therefore I do not believe that the benchmarks were significantly affected by cache differences.


When compared to Zen 2, a part of the speed increase is due to clock frequency increases, but that part is only of about 4% or 5% at most.

However, about the same speed increase is seen versus Intel CFL/CML, which have higher clock frequencies.

The benchmark "Vector SIMD" is multithreaded and in its floating-point tests the 16-core Zen 3 has the same speed as the 14-core Cascade Lake i9-10940x, which uses AVX-512 with dual 512-bit FMA units and which has much larger memory throughput. Zen 2 would have required 20 cores to match the Intel AVX-512 speed (down from 14 * 2 = 28 cores, due to the large drop of frequency for Cascade Lake when using AVX-512).

In the integer SIMD tests the 16 core Zen 3 exceeds by a large difference the 14-core AVX-512 CPU.

So I think that the only explanation is an increase in the number of SIMD execution units, maybe from 2 to 3 for FMA and from 3 to 4 for integer.

Zen already had 4 SIMD pipelines, but there were various restrictions on what kinds of operations could be executed by each of them. In Zen 3 they might have become more complex, so that some operations can now be executed in more of them.


https://www.sisoftware.co.uk/2020/10/30/amd-ryzen-9-5950x-zen3-review-benchmarks-cpu-16-core-32-thread-performance/