NVIDIA GeForce RTX 2070 OpenCL, CUDA, TensorFlow GPU Compute Benchmarks

[phoronix]

Phoronix: NVIDIA GeForce RTX 2070 OpenCL, CUDA, TensorFlow GPU Compute Benchmarks

Here are the first of our benchmarks for the GeForce RTX 2070 graphics card that launched this week. In our inaugural Ubuntu Linux benchmarking with the GeForce RTX 2070 is a look at the OpenCL / CUDA GPU computing performance including with TensorFlow and various models being tested on the GPU. The benchmarks are compared to an assortment of available graphics cards and also include metrics for power consumption, performance-per-Watt, and performance-per-dollar.

http://www.phoronix.com/vr.php?view=26984

[gdiamos]

@Michael, RTX gpus have FP16 2x rate performance like Vega, nothing to do with Tensor cores. Also there is not such a thing as Tensor core, we call them DSPs you know for years now. Also if Nvidia wants they can flash Gtx1000 with the Pro firmware and deploy FP16 2x there to. They can probably backport DLSS to.

This is incorrect. See the following references from NVIDIA's official website and wikipedia:

https://devblogs.nvidia.com/nvidia-t...ture-in-depth/
https://en.wikipedia.org/wiki/List_o...ce_10_series_2

I asked a legitimate question and don't have a clear answer.

[pbarletta]

Cheapest one I can find locally is 660€ with no warranty. Fuck nvidia and its prices.

I've been hearing these rumors that NVIDIA is supposedly prepping up another launch for a 2019Q4 and that this iteration is but an unfinished release to bank some money in. So you might get better values next year. Or maybe not. Who knows

[eydee]

Cheapest one I can find locally is 660€ with no warranty. Fuck nvidia and its prices.

[OMTDesign]

I'm glad that you included the GTX980

Also if Nvidia wants they can flash Gtx1000 with the Pro firmware and deploy FP16 2x there too.

Wait what! Are you telling me I can flash the VBIOS of a Quadro P5000 onto a GTX1080, and effectively use it as a Quadro card. If so can provide some evidence of this?

Awesome, now all of the dozen or so people who really need to do ML for a living, will have a GPU to buy.

[hauberg]

Is my understanding correct if I say that the Tensorflow benchmarks all focus on the the forward pass (at times also "inference") of a neural network? If so, is it possible to measure the backwards pass (backpropagation)? as well

[artivision]

@Michael, RTX gpus have FP16 2x rate performance like Vega, nothing to do with Tensor cores. Also there is not such a thing as Tensor core, we call them DSPs you know for years now. Also if Nvidia wants they can flash Gtx1000 with the Pro firmware and deploy FP16 2x there to. They can probably backport DLSS to.

[gdiamos]

Are the fp16 tests using the TensorCores or normal fp16?

[Michael]

Typos:

Thanks, as always... Long day/week with all the product launches recently and tomorrow...