Radeon ROCm 2.0 Officially Out With OpenCL 2.0 Support, TensorFlow 1.12, Vega 48-bit VA

[TemplarGR]

Stop complaining and buy Navi.

We should complain, because we have no reason to buy Navi if AMD acts like older hardware does not exist. Because Navi will become "old" a year after release too... What happens when they refuse to support a new feature for Navi then?

It is especially sad when you realize that Tonga is essentially the same architecture as Fiji, yet it has no support whatsoever. I realize there are manpower concerns, and i am patient enough, but if it ends up never getting support then it would be sad. It inspires no confidence in a company when hardware that was still sold a year ago gets overlooked.

[Michael]

Hey Michael, do you have any plans to benchmark TensorFlow with it? I don't know enough about it to contribute some tests (though I heard there was a container for ROCm + TensorFlow somewhere), but I know some people that would be interested in the results

Plan to look at the ROCm tensorflow support soon as I have one of my Vega cards not busy with other tests, hopefully this weekend~Christmas.

[bridgman]

It is especially sad when you realize that Tonga is essentially the same architecture as Fiji, yet it has no support whatsoever. I realize there are manpower concerns, and i am patient enough, but if it ends up never getting support then it would be sad. It inspires no confidence in a company when hardware that was still sold a year ago gets overlooked.

Tonga was developed and shipped almost a year before Fiji (part of 2xx series) and IIRC had some issues with the hardware since we hadn't developed the ROCm stack for dGPUs at the time and no other driver uses those HW features. The CU's are similar to Fiji but other blocks were updated.

It's not a matter of "overlooking" although I suppose you could argue it if we could throw enough developer effort at the problem we might be able to figure out a way to work around it.

[TemplarGR]

Tonga was developed and shipped almost a year before Fiji (part of 2xx series) and IIRC had some issues with the hardware since we hadn't developed the ROCm stack for dGPUs at the time and no other driver uses those HW features. The CU's are similar to Fiji but other blocks were updated.

It's not a matter of "overlooking" although I suppose you could argue it if we could throw enough developer effort at the problem we might be able to figure out a way to work around it.

Yes, i get that, but GFX7 gpus are older than Tonga, have experimental support (which is better than no support), and don't require pcie atomics to boot! I mean, ok, it is hard to make it work because it lacks some stuff from other newer GFX8 gpus, it lacks stuff from older GFX7 gpus too?

[AndyChow]

Plan to look at the ROCm tensorflow support soon as I have one of my Vega cards not busy with other tests, hopefully this weekend~Christmas.

If you do, please make sure that PCIe Atomics is working, because ROCm tensorflow will work without Atomics on Vega, but much slower.

I'd love to see the results, but I can wait.

[aufkrawall]

TemplarGR I really wouldn't advise buying in good faith companies to provide new feature support in the future. Disappointment is predestined elso.

I also think it's also not unfair to buy new products from a FOSS friendly company like AMD (with limited resources) from time to time.

[TemplarGR]

TemplarGR I really wouldn't advise buying in good faith companies to provide new feature support in the future. Disappointment is predestined elso.

I also think it's also not unfair to buy new products from a FOSS friendly company like AMD (with limited resources) from time to time.

Dude, i only buy from AMD, not "from time to time". For many years now. Still, it sucks that older gpus are supported but not mine, this is not so much to ask... I am not asking from AMD to support HD3000 gpus in ROCm, just Tonga...

[Tomin]

ROCm is the HSA effort and AMD has another OpenGL implementation for hardware that can not be supported on ROCm. The unfortunate thing is that frameworks like Tensorflow do not support OpenGL but use CUDA (cuDNN) and such (ROCm/MIOpen).

[TemplarGR]

ROCm is the HSA effort and AMD has another OpenGL implementation for hardware that can not be supported on ROCm. The unfortunate thing is that frameworks like Tensorflow do not support OpenGL but use CUDA (cuDNN) and such (ROCm/MIOpen).

Tonga can be supported by ROCm though... If GCN 1.0 can be supported, so can Tonga.

[aufkrawall]

Dude, i only buy from AMD, not "from time to time". For many years now. Still, it sucks that older gpus are supported but not mine, this is not so much to ask... I am not asking from AMD to support HD3000 gpus in ROCm, just Tonga...

Well, they can't do it, so stop asking.
There have been very cheap RX 570 cards lately which won't lose much total value when Navi shows up.