Graphics card

[hoa803]

My RTX 2060 is about 2.5 to 3x faster per playout than my old GTX 1060 6gb on Leela Zero. Furthermore, the power consumption is similar. For neural networks I believe it is worth the price, but ymmv.

[iopq]

Nvidia Tesla V100 has 1650 nps with 10 threads and batchsize of 5

Slightly more with 15 threads, but probably there's diminishing returns, so I don't increase the amount of threads (lower playing strength with too many threads)

[spook]

Before we can talk about relevant numbers we need enough samples.
We should think about a platform to store, index and share this information.

Any ideas ?

[toratujige]

Just ran leelaz benchmark on a RTX 2060.
Network #233 (40x256): 652 n/s
Network #157 (15x192): 1834 n/s

[iopq]

Just ran leelaz benchmark on a RTX 2060.
Network #233 (40x256): 652 n/s
Network #157 (15x192): 1834 n/s

Thanks for this, this is the data point I wanted. This is likely about the best performance per dollar, unless the 2060s is much faster.

[Gomoto]

Any info about the performance of the recent AMD cards available? Are they compatible at all?

[Aram]

As a general rule of thumb:

If you are planning on running software that supports Nvidia RTX series Tensor Cores, then no, AMD does not have anything to compete.
Tensor Cores are purpose built to accelerate neural networks, so in case you are running Leela for example, they will outperform AMD.

AMD simply hasnt added (yet?) these types of specialised cores into their GPU lineup.

In case you are running software which can not use the tensor cores, then the next thing to ask, does the card support faster FP16 execution?

If it does, it can once again cause a 2x speedup. For example the GTX 1660 can do 2:1 FP16:FP32, the GTX 1060 can not, which makes a big difference.

[Gomoto]

Thanks for the info.

I am using a 1080ti right now.

I read some info about Leela Zero running on AMD Vega. But I am interested how the new architecture on the RX 5700 (XT) performs with Leela Zero. I read AMD focused on gaming performance. (I know it will not perform like the 1080ti, I am just curious about the performance.)

[Aram]

I dont have any benchmarks to give you, but i wouldnt be surprised if they performed very similarly, or if the 5700 XT was actually a bit quicker.

The 5700 XT can do 2:1 FP16:FP32, the GTX 1080 Ti (or any cards of that family for that matter) can not.

So on the 5700 XT you get around 18 tflops (theoretically) of fp16 performance

On the GTX 1080 Ti you get 11.5 tflops (theoretically) of fp32 performance (it wont do fp16)

FP16 is fine for neural networks, wont make any difference, so the speed up is "free" in case the card supports it.

From what ive seen the FP16 to FP32 improvement in Leela isnt double, but you still get a speed boost.

Anyway, this all in theory

[explo]

Is the FP16 thing recent? I changed my GTX 1050 for a GTX 1660 Ti in March and got a nice boost at the time, but when using autogtp I now get a lower time per move than I did a month(?) ago.
Sadly the tensor cores support wasn't there back in Marchy, or I would have bought a RTX2060 instead.

[iopq]

So has anyone tried 5700 or 5700xt on lz?

[iopq]

If you rent a K80, it has about 160 n/s or so, similar to an RX 570, so not worth even using trial credits for

[Gomoto]

Friend needs go graphics card. I tend to recommend RTX 2060. RX5700 seems crash or slow according to some posts I have read. Any experiences that differ?

[iopq]

Friend needs go graphics card. I tend to recommend RTX 2060. RX5700 seems crash or slow according to some posts I have read. Any experiences that differ?

it doesn't have tensor cores and worse support (no CUDA, etc.)

I've seen reports of Leela Zero crashing, but you can probably still use KataGo in OpenCL mode

Only get the RX5700 for gaming. I'm actually very interested in its performance for Go-related stuff - so if he gets it please benchmark it

[iopq]

I got an RTX 2060 which gave me 775 nps on current master (40b network) and 1050 nps on KataGo (20b network)