I would like to set up a new Linux test system/server to evaluate HSA and Boltzmann (hipify) software with regards to multi-GPU computing. Ideally, I would like it to be an APU + multi-dGPU system (as many GPU resources as possible!).
Thanks!
I don't think ROCm allows mixing iGPU and dGPU yet. (Setting Up The Radeon Open Compute Platform On Linux - Phoronix Forums ) The safest way seem to go with what they are currently testing (GitHub - RadeonOpenCompute/ROCK-Kernel-Driver: Boltzmann Initiative: Kernel Driver ) They are currently testing multilple R9 Nano's or APU with no dGPUs.
I would like to densely pack R9 Nano GPUs into a server if possible, but I'm unsure if the chassis cooling would be sufficient to mitigate the no-headroom issue with tightly spaced open-air GPUs.
Are there any AMD FirePros that support HSA and Boltzmann?
We also just released a 1.1 version of the kernel component which includes improved thermal management, so Nano's should now ramp up fan speeds if the cooling isn't good enough. In the 1.0 release we manually forced a higher fan setting to make sure tightly-packed configurations didn't overheat, but with 1.1 that should no longer be required.
Thanks bridgman. Have you seen issues with GPU fan performance on the Nano's when they are packed so closely?
I was wondering how well ROCm is tested on APUs. I know APUs are not the focus for AMD at the moment. But ROCK-Kernel-Driver still lists them as supported platform. I'm asking because I ran into issues while hipifying PARBOIL and am trying to figure out whether the cause is my hardware or ROCm software. See this and this.
I'd appreciate any info.
Thanks.