HN Explorer

Ask HN: How come my Ryzen 3750H is faster than my RTX 2060 in TensorFlow?

7 points by s-p-n ↗ HN
I thought the RTX would do better here, but the Ryzen 3750h performs better on training a model in tensorflow. I'm using this guide to get things going, and I'm usng tf-nightly with Ubuntu 20.04.

https://www.tensorflow.org/tutorials/quickstart/beginner

I'm wondering if using nvidia 440 drivers is the reason? I also have 3 versions of CUDA installed (10.0, 10.1, and 10.2). The tensorflow-js test I did uses CUDA 10.0, and the Ryzen performs roughly 50% faster. The Python test I did uses CUDA 10.1, and the CPU is less than 10% faster.

Even so, shouldn't the GPU perform an order of magnitude better here? I feel like I must be doing something wrong :/

6 comments

[ 2.9 ms ] story [ 22.4 ms ] thread
navjack27
I think my calculations are correct here

  Model       Zen+  2060
  Cores       4     1920
  Threads     8     1920
  Cycles      2.3   1.365
  FP64 GFLOPS 2944  163
  FP32 GFLOPS 5888  5242
  FP16 GFLOPS 11776 10483
So it would kind of make sense yeah. 2060 isn't that strong.

But note: I'm not totally sure on the Zen+ theoretical speeds

s-p-n
Yeah, I just finished installing the CUDA/tensorflow/python set on windows 10, and have the exact same results. So, it's not my configuration that's the problem.

Wow, this really brings to light how well Ryzen really performs. My nvidia 950m put shame to whatever intel i7 I had on my last laptop lol. Now I'm not even sure why I got a GPU on this rig

navjack27
Kinda shows how gimped gaming GPUs are for doing training really.
Ragib_Zaman
The network that you train in that tutorial is quite small. Your networks have to get quite large before the overhead of transferring the network onto the GPU is negligible to the cost of training it. You'll see the benefit of the GPU with larger models.