3 ms·
You'd also need extreme hyperthreading. A GPU can cycle between several warps in the same execution unit (barrel-processor-style), padding out the time per inst
by immibis 2y ago
You'd also need extreme hyperthreading. A GPU can cycle between several warps in the same execution unit (barrel-processor-style), padding out the time per instruction to hide memory latency, while still getting the same throughput. That's counter to the fundamental design of CPUs.