Hey cryoSPARC Team,
I am wondering if there are any plans to implement AMD GPU support similar to RELION? The AMD RX 9070XT or the workstation variant AMD Radeon AI PRO R9700 appear to be quite performant and price-competitive to their Nvidia counterparts. Also, more and more libraries support those GPUs, and for example, AWS has AMD as an option.
Cheers,
Ole
Agreed.
Given the utterly ridiculous prices that nVidia are demanding for their Blackwell GPUs (and, frankly, continuing concerns over the safety of the 12VHPWR connector) diversifying options for compute can only be a good thing.
That said, ROCm is still a joke. I’ve grown tired of AMDs promises, and the contortions necessary to get ROCm working. Maybe with ROCm 7. But I’ve heard that before with 6, 5 and 4.x…
@rbs_sci I’ve never tested ROCm myself, but your experience and that of many other people in forums should “encourage” AMD to improve a lot. Maybe ROCm 7 will be the game changer, perhaps not, but if I recall correctly, the RX 9070 series even launched without ROCm support…
On the same note, Intel GPU support might be desirable as well, having even more diverse GPU supply options. Many cryoSPARC tasks are limited by something else than raw GPU compute and don’t require a RTX4090/RTX5090 to perform well.
The best target might be Vulkan Compute, as that is card (and even CPU) agnostic - get a card/driver and, regardless of manufacturer, provided it supports Vulkan, it should largely work transparently. But from my very limited tinkering, Vulkan is painfully verbose and fussy.
Fully agree with the desire for additional GPU compatibility. The latest release of Relion has AMD GPU support (HIP/ROCm), and I was able to get it running somewhat easily on my personal machine with a 6950XT. I believe they also support Intel GPUs (SYCL).
@gdodge, how does the performance of your 6950XT compare to Nvidia GPUs in Relion? I am just curious. According to the Relion documentation, it should be possible to install it on Nvidia/AMD/Intel GPUs. It would be great if cryoSPARC would support it too, since those are among the two most important processing tools in SPA.
I can’t really comment much on performance, since this was on my home machine that has minimal RAM compared to your average workstation. I mostly ran a few low-RAM jobs to see if things would actually complete. Since my current rig only supports DDR4, I’m waiting for a full mobo/CPU upgrade before adding additional RAM.