Will AMD RDNA2 based Radeon RX 6000 series kick butt with Lc0?

[smatovic]

AMD teased its RX 6000 series with new RDNA2 architecture (already in use in
PlayStation/Xbox gamer consoles) and released as high-end consumer series on
Nov 18th. Take a look at the FP16, half precision, performance, pretty!

https://en.wikipedia.org/wiki/Radeon_RX_6000_series

AFAIK RDNA2 supports RayTracing (not of interest for Lc0) but will not have
Nvidia like TensorCores on board. So, question, will the FP16 performance of
RDNA2 be able to take on the boost of Nvidia's RTX TensorCores? How much
speedup do Lc0's CNNs gain from the TensorCores? Maybe we have here a game
changer?

--
Srdja

[smatovic]

Anandtech's dive into RDNA2 Navi 21 chip:

https://www.anandtech.com/show/16202/am ... ber-18th/2

Looks like AMD added 128 MB of L3 cache on its high-end GPUs, another candidate for speedups despite FP16 throughput, or maybe batch-size for Lc0 could decrease, depending on net-size, or alike.

--
Srdja

[Laskos]

AMD teased its RX 6000 series with new RDNA2 architecture (already in use in
PlayStation/Xbox gamer consoles) and released as high-end consumer series on
Nov 18th. Take a look at the FP16, half precision, performance, pretty!

https://en.wikipedia.org/wiki/Radeon_RX_6000_series

AFAIK RDNA2 supports RayTracing (not of interest for Lc0) but will not have
Nvidia like TensorCores on board. So, question, will the FP16 performance of
RDNA2 be able to take on the boost of Nvidia's RTX TensorCores? How much
speedup do Lc0's CNNs gain from the TensorCores? Maybe we have here a game
changer?

--
Srdja

Not sure, I am skeptical. Don't know DX12, but OpenCL is pretty much abandoned, and the maintainers for the biggest ML frameworks (TensorFlow, Caffe, PyTorch) have not shown enough interest. AMD does not provide enough developers to build and maintain their own ML libraries. They should spend time contributing to open-source projects like TensorFlow and PyTorch so that those libraries eventually support AMD GPUs. Nor sure about Lc0, but generally the ML market is basically open pickings for NVIDIA with or without AMD 6000 series. As for FP16 performance, it was high with RX 5700 XT too, as high as RTX 2070, but it meant little to Lc0 as far as I know.

[smatovic]

Hmm, yea, backend, driver and frameworks make a difference, dunno how good DX12 backend on AMD gpus performs.

When I compare RTX 2060 with TenscoreCores with GTX 1650 Super without from below sheet I get roughly a 2x speedup by TensorCores for Lc0:



not sure if backend params and driver sets are comparable.

So AMD RX 6000 series with about 1.3x more FP16 throughput in same price segment would still lack 1.5x in NPS behind TensorCore based Nvidia RTX GPUs...roughly estimated, possible boost with new L3 cache not considered.

--
Srdja

[Laskos]

Hmm, yea, backend, driver and frameworks make a difference, dunno how good DX12 backend on AMD gpus performs.

When I compare RTX 2060 with TenscoreCores with GTX 1650 Super without from below sheet I get roughly a 2x speedup by TensorCores for Lc0:



not sure if backend params and driver sets are comparable.

So AMD RX 6000 series with about 1.3x more FP16 throughput in same price segment would still lack 1.5x in NPS behind TensorCore based Nvidia RTX GPUs...roughly estimated.

--
Srdja

Good estimate, I am getting something on these lines from your link and searching on discord. It looks like RX 5700XT DX12 is something like twice slower compared to RTX 2070. That would mean a RX 6800XT DX12 will probably be the level of RTX 2070, while pricey. Much cheaper RTX 3070 is at least 50% faster.