GPU recommendation needed

[cma6]

I have a dual-Xeon system with 2 Intel Xeon E5-2686-V3 cores or 2 X 18 = 36 cores total. I would like to add a decent GPU card so as to test lc0 but not reach for the top (GTX 2080). What would be recommended in terms of a value purchase: GTX 2070; GTX 2070S; or is there an AMD GPU I should look at (NAVI or Vega 64). Will there be any further GPU releases in 2019?

I read, probably on this board, that there is a huge disparity in performance as the Nvidia GPUs benefit from the CUDA coding, while the AMD cards are restricted to the OpenCL model. This can lead to differences of 10x the performance when upper end cards of similar prices are compared.

Thanks to the experts.

[Nordlandia]

RTX 2080 Super is brand new.

[dragontamer5788]

I read, probably on this board, that there is a huge disparity in performance as the Nvidia GPUs benefit from the CUDA coding, while the AMD cards are restricted to the OpenCL model. This can lead to differences of 10x the performance when upper end cards of similar prices are compared.

It goes beyond just CUDA vs OpenCL. NVidia's chips have "tensor units", which are dedicated 16-bit multiply-and-accumulate units to especially accelerate deep-learning / neural network code. AMD has "packed floats", but that will never be as fast as NVidia's tensor units.

AMD's "general compute" Radeon VII and Vega64 chips are very impressive on a pure specs perspective. Radeon VII has 16GB of HBM2 delivering 1TBps (!!!) to the GPU. But ultimately, they don't have those neural-network instructions, so they won't be as fast as NVidia Turing / NVidia Volta. If you were doing general purpose SIMD-compute, I would argue that Radeon VII / Vega64 are best bang/buck.

But if you're going to be running LeelaZero over Tensorflow, you should get an NVidia Volta / NVidia Turing instead. 2080 Super is probably your best bet.

[mwyoung]

I have a dual-Xeon system with 2 Intel Xeon E5-2686-V3 cores or 2 X 18 = 36 cores total. I would like to add a decent GPU card so as to test lc0 but not reach for the top (GTX 2080). What would be recommended in terms of a value purchase: GTX 2070; GTX 2070S; or is there an AMD GPU I should look at (NAVI or Vega 64). Will there be any further GPU releases in 2019?

I read, probably on this board, that there is a huge disparity in performance as the Nvidia GPUs benefit from the CUDA coding, while the AMD cards are restricted to the OpenCL model. This can lead to differences of 10x the performance when upper end cards of similar prices are compared.

Thanks to the experts.

If you want to test Lc0. Then you need a RTX card. What RTX card depends on what you want to spend.

RTX 2080 ti
RTX 2080 S
RTX 2080
RTX 2070s
RTX 2070
RTX 2060s
RTX 2060

Basicly the Super version will be replacing the standard version. With a 4 to 10 percent speed increase.

I would recommend the RTX 2070s. Depending on price. This gets you really close to a RTX 2080 in performance.

RTX 2070 SUPER $499

I would recommend this card.

MSI GeForce RTX 2070 SUPER VENTUS OC Graphics Card $509.99

MSI has the best coolers & fan noise. And you will need this when running Lc0.

[Nordlandia]

Check here.

https://gpu.userbenchmark.com/Compare/N ... 4050vs4026

[Hai]

I have a dual-Xeon system with 2 Intel Xeon E5-2686-V3 cores or 2 X 18 = 36 cores total. I would like to add a decent GPU card so as to test lc0 but not reach for the top (GTX 2080). What would be recommended in terms of a value purchase: GTX 2070; GTX 2070S; or is there an AMD GPU I should look at (NAVI or Vega 64). Will there be any further GPU releases in 2019?

I read, probably on this board, that there is a huge disparity in performance as the Nvidia GPUs benefit from the CUDA coding, while the AMD cards are restricted to the OpenCL model. This can lead to differences of 10x the performance when upper end cards of similar prices are compared.

Thanks to the experts.

RTX Titan
RTX 2080 Ti Super
RTX 2080 Ti

[smatovic]

I have a dual-Xeon system with 2 Intel Xeon E5-2686-V3 cores or 2 X 18 = 36 cores total. I would like to add a decent GPU card so as to test lc0 but not reach for the top (GTX 2080). What would be recommended in terms of a value purchase: GTX 2070; GTX 2070S; or is there an AMD GPU I should look at (NAVI or Vega 64). Will there be any further GPU releases in 2019?

I read, probably on this board, that there is a huge disparity in performance as the Nvidia GPUs benefit from the CUDA coding, while the AMD cards are restricted to the OpenCL model. This can lead to differences of 10x the performance when upper end cards of similar prices are compared.

Thanks to the experts.

The Nvidia RTX series profits from its TensorCores, designed to accelerate deep neural networks.
AMD still has no counterpart to this, and the LC0 OpenCL backend is not as fast as the Nvidia CUDA and cuDNN backends.
AMD released on 7/7 its middle class Navi gpus, this caused Nvidia to drop prices resp. to release the RTX Super series.
It could happen that AMD releases Navi high end gpus this autumn, but up to now there is no official release date.
Intel plans to enter the discrete gpu market in 2020.

--
Srdja

[chrisw]

I have a dual-Xeon system with 2 Intel Xeon E5-2686-V3 cores or 2 X 18 = 36 cores total. I would like to add a decent GPU card so as to test lc0 but not reach for the top (GTX 2080). What would be recommended in terms of a value purchase: GTX 2070; GTX 2070S; or is there an AMD GPU I should look at (NAVI or Vega 64). Will there be any further GPU releases in 2019?

I read, probably on this board, that there is a huge disparity in performance as the Nvidia GPUs benefit from the CUDA coding, while the AMD cards are restricted to the OpenCL model. This can lead to differences of 10x the performance when upper end cards of similar prices are compared.

Thanks to the experts.

The Nvidia RTX series profits from its TensorCores, designed to accelerate deep neural networks.
AMD still has no counterpart to this, and the LC0 OpenCL backend is not as fast as the Nvidia CUDA and cuDNN backends.
AMD released on 7/7 its middle class Navi gpus, this caused Nvidia to drop prices resp. to release the RTX Super series.
It could happen that AMD releases Navi high end gpus this autumn, but up to now there is no official release date.
Intel plans to enter the discrete gpu market in 2020.

--
Srdja

What would be good would be for CPU and GPU to be able to share the same RAM, also to be on the same chip.

[smatovic]

What would be good would be for CPU and GPU to be able to share the same RAM, also to be on the same chip.

Yes, there are different approaches for shared memory and lower latency, e.g.

https://en.wikipedia.org/wiki/Coherent_ ... _Interface

I guess with the upcoming Exa-FLOP super computers by Intel (Aurora in 2020) and
AMD (Frontier in 2021) we will see new kind of interconnections for accelerater
cards, but I am not sure if these will replace PCIe in our personal computers.

--
Srdja

[chrisw]

What would be good would be for CPU and GPU to be able to share the same RAM, also to be on the same chip.

Yes, there are different approaches for shared memory and lower latency, e.g.

https://en.wikipedia.org/wiki/Coherent_ ... _Interface

I guess with the upcoming Exa-FLOP super computers by Intel (Aurora in 2020) and
AMD (Frontier in 2021) we will see new kind of interconnections for accelerater
cards, but I am not sure if these will replace PCIe in our personal computers.

--
Srdja

Fast, realistic games graphics are a big driver of PC sales, with data processing using large nets also. Integrating GPU/CPU would be such a step change for both, that it seems to hardware and processor design non-expert like me, to be inevitable, and on PC, and not to many years off. ML engineers have way overtaken bankers salaries in London, and those are just the engineers, god knows how much the way fewer creative researchers are making. Difference this time, the boom bust cycle got interrupted because the expectations were matched by actual industrial applications working widespread. Hardware will catch up, it’s technically possible and economics will drive it. In my, by now generally out-dated, opinion.