4 ms·
There's some great information about how GPUs are organized here. It's interesting that GPUs and CPUs aren't that different, even though maximizing performance
by olafalo 8y ago
There's some great information about how GPUs are organized here. It's interesting that GPUs and CPUs aren't that different, even though maximizing performance requires different considerations. SIMD lanes are just wider on the GPU, and you have to treat memory differently (registers are bigger, main memory has higher latency but better bandwidth, etc).
It's also funny that Nvidia lists the total number of SIMD lanes as "CUDA cores", meaning that a single compute unit with 64 SIMD lanes counts as 64 CUDA cores. That's cheating, if you ask me :P
- twtw 8y agoIn Volta, each "SIMD lane" has its own PC, if that makes it more legit.
- monocasa 8y agoCitation? SXM2 didn't look that different, and adding that many ports to the new instruction cache they added would be absolutely crazy.
- twtw 8y agohttps://devblogs.nvidia.com/inside-volta/ https://devblogs.nvidia.com/inside-volta/ Ctrl-F "volta SIMT model"
- MaxBarraclough 8y agoI admit I'm not really keeping up here, but if it has its own PC (rather than doing SIMT lane-masking), how is it a SIMD lane? That sounds like manycore, not SIMD/SIMT, no?