Benchmark result for an AMD 6800

[ffleader1]

I see the readme in this repo does provide the time/epoch for several GPU. For future reference, here is my result on a Rx 6800, Rocm v5.2.3, Pytorch-Rocm v5.1.1 Time/Epoch:

• Without mixed percision: 2m1s
• With mixed percision: 1m7s

So apparently using mixed percision on Navi 21 does bring up certain improving in training time too.

[suvojit-0x55aa]

Thanks for this @ffleader1, would you happen to have the accuracy scores too. I'll add it to the README with acknowledgement of course.

[ffleader1]

Thanks for this @ffleader1, would you happen to have the accuracy scores too. I'll add it to the README with acknowledgement of course.

Hmm I did not run for 200 epochs. That would take quite a few hours, and I do need my PC for work. However, I do want to add some benchmark result. I will try to utilize my Pc when I am not actively using it. Will get back to you in a few days.

[ffleader1]

Did you use any special param tuning related to the learning rate? Because after 200 epoch, without mixed percision, I got a 37% accuracy, which is horrendous, with the running command as you specified in the Read.me

[suvojit-0x55aa]

Just the script I've put here, Maybe it need some changes due to how training in RX6800 libs are handled ?