Originally posted by AlB80
View Post
Now, I'm just speculating, but if pal666 was talking about GPUs with a 2:1 32:64 ratio, then he's right that because a fp32 is half as big as a fp64, it works out to the same data rate. However, that feels a bit like restating the obvious, so I suspect I'm misunderstanding either the statement or the intent (or both).
If we just back up to my earlier point that fp64 multipliers aren't free (a point which @AIB80 seems to have acknowledged), then is it so hard to consider that maybe the reason they never previously went for 1:1 32:64 is that they were always including a market (i.e. interactive graphics) that prized fp32 performance? And with CDNA, that's no longer the case. So, it seems natural that they would revisit their silicon apportioning and perhaps even their ABI to target the fp64-heavy workloads that are so important to HPC customers.
As I suggested before, maybe the new chip is just a natively 64-bit architecture (i.e. each vector lane is 64-bits, instead of 64-bit ops having to use pairs of 32-bit registers). In that case, perhaps they decided it wasn't worth bothering to support packed 2-in-1 fp32, they way they previously offered fp16 support (i.e. since vega10).
Anyway, we can just wait and see. I'm not about to buy a CDNA card, either way. By the time I'm ready to replace my Radeon VII, perhaps even RDNA cards will be outside my price range. Inflation in both CPU and GPU pricing has really gotten out of hand!
Comment