Ryzen 3000 leaked benchmarks

[ShogoXT]

https://m.imgur.com/a/YkoOCgM

If you scroll down you will see one of the charts running on handbrake using hevc 10bit. It's faster than the Ryzen 2950x and almost as fast as the 2990wx. Does this point to better avx performance?

I hope to get a 3950x if it comes down in price a bit. Twitch will be enabling vp9 support soon and hopefully it would provide decent single PC encoding detail.

[nevcairiel]

Originally Posted by ShogoXT

Does this point to better avx performance?

You don't need leaked benchmarks to know that. Ryzen 1000/2000 only had 128-bit execution units for AVX - which is natively 256-bit, as such it only performed at roughly half the speed that AVX was designed to run at, and how Intel would run it.

Ryzen 3000 extends those units to full 256-bit, so it can now run AVX1/2 at full speed, which greatly benefits x265, for example. This was fully confirmed at various Ryzen 3000 press events.

[soresu]

I always get confused by the whole thing about Zen 2 doubling the FP/FMAC backend.

I get that it would heavily benefit AVX2 FP SIMD, but I thought that most encoders use mostly integer based calculations, or have I got things completely backwards?

It would make it alot easier if they would stop mixing nomenclature about SIMD and FP as if they are one and the same.

[NikosD]

Originally Posted by soresu

I always get confused by the whole thing about Zen 2 doubling the FP/FMAC backend.



I get that it would heavily benefit AVX2 FP SIMD, but I thought that most encoders use mostly integer based calculations, or have I got things completely backwards?



It would make it alot easier if they would stop mixing nomenclature about SIMD and FP as if they are one and the same.

Video encoding like x265 is definitely an integer calculating process just like AVX2.

AVX2 extended the 256 bit calculations of AVX1 from FP to integers.

So, both AVX1/2 involve 256 bit calculations which is mainly FP for AVX1 and integer for AVX2.

The confusion still exists because AVX2 introduced also FMA3 which doubled the speed of specific (FMAC) calculations of FP AVX1 256 bit processing.