!!PSA: Major AMD Species reshuffle Part 2 underway!!
Posted: Thu Nov 16, 2023 9:20 pm
HI,
certain things have forced our hand to reshuffle AMD performance bracket (aka Species) yet again. This time we are not only splitting GPUs per performance, but also per GPU arch (GCN, RDNA 1, 2, 3, CDNA). This should allow us higher flexibility when assigning work or disabling certain architectures from certain projects either due to catastrophic performance or instability bugs.
While HIP is coming (hopefully), all the issues which plagued OpenCL, might be gone, but seeing how AMD is EOL'ing their support for older GPUs left and right, I think species reshuffle will still be extremely relevant.
Also, starting from RDNA3, there will no longer be performance brackets, as everything with RDNA3 arch on desktop and dGPU mobile is stupidly fast. iGPUs will still be pushed to lower tiers per their theoretical performance (hopefully HIP will makes those far more relevant than they are now).
Please, be mindful of this change in next week or so, until we get all the researchers informed and acting on changing their constraints. During this period your AMD GPUs might fail receiving any work, or might receive work which either performs very badly, or outright crashes. Hopefully this will be kept to the minimum, and everyone will come out happier
certain things have forced our hand to reshuffle AMD performance bracket (aka Species) yet again. This time we are not only splitting GPUs per performance, but also per GPU arch (GCN, RDNA 1, 2, 3, CDNA). This should allow us higher flexibility when assigning work or disabling certain architectures from certain projects either due to catastrophic performance or instability bugs.
While HIP is coming (hopefully), all the issues which plagued OpenCL, might be gone, but seeing how AMD is EOL'ing their support for older GPUs left and right, I think species reshuffle will still be extremely relevant.
Also, starting from RDNA3, there will no longer be performance brackets, as everything with RDNA3 arch on desktop and dGPU mobile is stupidly fast. iGPUs will still be pushed to lower tiers per their theoretical performance (hopefully HIP will makes those far more relevant than they are now).
Please, be mindful of this change in next week or so, until we get all the researchers informed and acting on changing their constraints. During this period your AMD GPUs might fail receiving any work, or might receive work which either performs very badly, or outright crashes. Hopefully this will be kept to the minimum, and everyone will come out happier
