3 ms·
Each compute unit handles 2560 hardware threads (64 per clock tick per compute unit). AMDs GCN architecture runs 4 threads over 4 clock ticks natively. I forge
by dragontamer 7y ago
Each compute unit handles 2560 hardware threads (64 per clock tick per compute unit).
AMDs GCN architecture runs 4 threads over 4 clock ticks natively. I forget if it was vertical SIMD or horizontal SIMD or whatever the proper terminology was. But this 4 threads over 4 clock ticks avoids the register stall issue. (RDNA performs 1 thread per clock tick and schedules around the stalls)
The other threads come from an SMT-like / hyperthread-like feature, allowing 10 hardware threads per slot.
All together, a compute unit supports 16 way SIMD per vALU x 4 vALUs per CU x 4 threads over 4 clock ticks x 10 'hyperthread-like' = 2560 hardware threads (over 4 clock ticks)
16 SIMD x 4 clock ticks gives your 64 sized wavefronts / warps that the programmer works with. The majority of the 'threads' are for latency hiding: work the GPU can do while waiting for RAM to respond. (In contrast to a CPU, which will search ~200 instructions out of order for work while waiting on RAM)