GPU rumors - AMD Navi and Nvidia GTX 1160

[smatovic]

AMD confirms launch of AMD Navi (TSMC 7nm) in 2019 (possible Q2):

https://www.fudzilla.com/news/graphics/ ... i-for-2019

And there are rumors about Nvidia GTX 1150, 1160 and 1160 TI with TensorCores but no RayTracing for February/March 2019:

https://www.fudzilla.com/news/graphics/ ... -in-retail

So, for patient buyers it could pay off to wait.

It is unknown if AMD Navi has a counterpart to Nvidia TensorCores onboard to boost LC0,
but even if not, the release could lower gpu prices overall...

And for really patient people - Intel still plans to enter discrete gpu market in 2020.

--
Srdja

[Ozymandias]

there are rumors about Nvidia GTX 1150, 1160 and 1160 TI with TensorCores but no RayTracing for February/March 2019

The 1660Ti is coming out this week, hardly any patience required to get one, at most you'll have to wait a little for your favourite dealer to list it.

[M ANSARI]

Getting rid of Ray Tracing is certainly a great thing for Chess as the RT part is not utilized. This would make the card much cheaper and hopefully not as large and maybe with less thermal problems. But looking at the specs of the GTX 1160 it seems weaker than RTX 2060. Hopefully the idea of removing RT will catch on and will also happen on the higher performing cards as well.

[Ozymandias]

Getting rid of Ray Tracing is certainly a great thing for Chess as the RT part is not utilized. This would make the card much cheaper

They've taken a 20% of the cores for a 20% price reduction. They've also increased core clocks about 10% to offset missing features (RT included), so they've been rather on the greedy side of things with this particular model. Maybe next month they'll prove to be a little more generous with the vanilla 1660.

[smatovic]

Just for the record, GTX 1160 TI dropped RTX and TensorCores,

https://www.anandtech.com/show/13973/nv ... c-gaming/2

--
Srdja

[Ozymandias]

This part still gives out hope:

with the original TU102 GPU and the RTX 2080 Ti [...] NVIDIA changed how standard FP16 operations were handled. Rather than processing it through their FP32 CUDA cores, as was the case for GP100 Pascal and GV100 Volta, NVIDIA instead started routing FP16 operations through their tensor cores.

The tensor cores are of course FP16 specialists, and while sending standard (non-tensor) FP16 operations through them is major overkill, it’s certainly a valid route to take with the architecture. In the case of the Turing architecture, this route offers a very specific perk: it means that NVIDIA can dual-issue FP16 operations with either FP32 operations or INT32 operations, essentially giving the warp scheduler a third option for keeping the SM partition busy. Note that this doesn’t really do anything extra for FP16 performance – it’s still 2x FP32 performance – but it gives NVIDIA some additional flexibility.

Of course, as we just discussed, the Turing Minor does away with the tensor cores in order to allow for a learner GPU. So what happens to FP16 operations? As it turns out, NVIDIA has introduced dedicated FP16 cores!

These FP16 cores are brand new to Turing Minor, and have not appeared in any past NVIDIA GPU architecture. Their purpose is functionally the same as running FP16 operations through the tensor cores on Turing Major: to allow NVIDIA to dual-issue FP16 operations alongside FP32 or INT32 operations within each SM partition. And because they are just FP16 cores, they are quite small. NVIDIA isn’t giving specifics, but going by throughput alone they should be a fraction of the size of the tensor cores they replace.

To users and developers this shouldn’t make a difference – CUDA and other APIs abstract this and FP16 operations are simply executed wherever the GPU architecture intends for them to go – so this is all very transparent. But it’s a neat insight into how NVIDiA has optimised Turing Minor for die size while retaining the basic execution flow of the architecture.

As long as Lc0 isn't using Tensor Cores for anything else than FP16 operation, these cards could still be a good option.

[brianr]

As long as Lc0 isn't using Tensor Cores for anything else than FP16 operation, these cards could still be a good option.

It was mentioned on the Leela Discord that cudnn backend does not support the type of fp16 on this card yet:

"ankan Today at 11:43 AM
TU116 GPU doesn't have tensor cores, but it has double throughput for fp16 math - similar to P100 GPU (it's also available on GV100, and TU10x GPUs, but tensor cores are preferred as they have 4x higher throughput).
right now I think the cudnn backend of lc0 can't make use of the non-tensor core fp16 math throughput because it needs NCHW data layout (and right now we use NHWC with fp16 which is needed for tensor cores)
I can work on a PR to try to support that (for GPUs like 1660Ti that have higher precision fp16 math throughput but no tensor cores)."

[smatovic]

AMD unveiled their 7nm Navi based Radeon RX 5700 GPUs to be released on 07/07,

https://www.anandtech.com/show/14412/am ... roved-perf

https://www.anandtech.com/show/14528/am ... 700-series

https://en.wikipedia.org/wiki/AMD_Radeon_5000_series

Aimed for the middle-class with new RDNA architecture, noticeable higher clock rates, GDDR6 and PCIe 4.0,
but it seems there are no TensorCores on board...

--
Srdja