3 ms·
Everyone has expected the price per Tflop to decrease at every generation. If the price per Tflop and the power consumption per Tflop do not decrease in a new
by adrian_b 4y ago
Everyone has expected the price per Tflop to decrease at every generation.
If the price per Tflop and the power consumption per Tflop do not decrease in a new generation of GPUs, then there is no reason for a new generation to exist.
The power consumption per Tflop has continued to decrease reasonably well, but the price has almost stagnated, for FP32. On the other hand for FP64 computations, the price per Tflop has greatly increased during the recent years, due to the fact that both NVIDIA and AMD have restricted this feature to their "datacenter" GPUs, which have prices that are acceptable only for those who buy them in very large numbers and for whom the costs for the expensive GPUs are compensated by savings in rack space, power consumption and cooling.
It seems that AMD has removed completely the FP64 execution units in RDNA 3, while NVIDIA provides only a 1:64 throughput ratio for FP64:FP32 in RTX 4000, so they should have better also removed completely the FP64 hardware, because a software implementation would be faster than that.
Now a CPU like Ryzen 9 7950X has a lower price per FP64 Tflop than any GPU, which is opposite from a few years ago, when e.g. a Radeon VII had a much lower price per FP64 Tflop than any CPU.