Any good examples of historical rendering dead-ends, e.g. Ecstatica’s ellipsoid technology?
-
-
I haven't worked on LRB but I remember at the time thinking there was no way for it to succeed as a discrete GPU. That said I'm grumpy and skeptical about most things nowadays. Also usually right, but that's easy, most stuff fails.
-
Funnily enough the architecture of modern GPUs is not that distant from that original grand vision.
-
Certainly LRBNI/AVX512 turned out spookily similar to GCN ISA. Some good "convergent evolution" going on. The micro-architectures are quite different though.
-
I mean at a glance, I guess you are right but it’s not that these GPU concepts didn’t exist before lrb or gcn. Unless you are talking about some more specific things, you are too vague.
-
But I mean even just thinking of the isa, and not the uarch, there are several key differences. I’d say I see more differences than similarities but I am very curious to know what you mean
-
ISA the same: predication, flow control, vector/scalar split (regs, pipeline and ISA). GCN had more free negate/abs/scale options. Broad concept of taking to TexUnits similar - all state explicit packed into request, rather than being implicit inside TXU.
-
I have been always wondering whether GCN has 64 wide waves because it needs to send resource descriptor inside each sample/load instruction. Wider wave makes this additional data less significant (as UVs = 64 bits per lane = 4096 bits in total). 256 bit desc = 6.25%.
-
SIMD width choice is always about efficiency (wider) vs occupancy (narrower). I suspect there's LOTS of places that went into the decision. Intel's GPUs hedge their bets and support multiple SIMD widths.
- 6 more replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
