07:43 colo: superkuh: I think that poll is very strange
07:43 colo: the answer, of course, should be: all of them
12:34 Venemo: superkuh: it would be interesting to know what stops you from supporting all of them. it's not like there are that many differences between the ISA. especially if you support some chips from a generation, what stops you from supporting the rest? especially if they use the exact same ISA.
12:36 Venemo: superkuh: for example, some RDNA2 chips are already supported. so what's the difficulty in supporting the rest of RDNA2?
12:47 glehmann: QA time?
13:00 FLHerne: ditto. From a user perspective AMD's support lifecycle for compute is incredibly offputting
13:01 FLHerne: You can buy any Nvidia GPU from the last decade and compute will just work -- or, more to the point, you can write CUDA code and it'll just work for all your users/customers on whatever Nvidia GPU they have
13:04 FLHerne: this constant nickel-and-diming about "but which GPUs *really* matter to you" is inane and makes it imposible to trust any AMD device whether it happens to currently be supported or not
14:42 agd5f: there is ROCm support for all GPUs already, what is considered "supported" is what QA validates
16:00 superkuh: Venemo, I mostly want ROCm back for the late GCN4.0 refresh cards like the RX 580. They have the memory bandwidth of a RX 6600, just a bit less compute, a bit more power, and cost a tiny fraction of it.
16:01 superkuh: Trying to do compute with an RX 580 and ROCm is sort of feasible currently if you bend over backwards using old drivers and pray that the software you're using can also use old stuff (it almost never can).
16:01 superkuh: So I end up having to use Vulkan which is slower.
16:03 superkuh: So, no, not "all GPUs already".