2 ms·
A fun quick of GPU pricing economics is that on Google Cloud platform, the relatively recent T4 (on Turing and has FP16 support) is cheaper than the ancient K80
by minimaxir 7y ago
A fun quick of GPU pricing economics is that on Google Cloud platform, the relatively recent T4 (on Turing and has FP16 support) is cheaper than the ancient K80s. https://cloud.google.com/compute/gpus-pricing https://cloud.google.com/compute/gpus-pricing
- tedivm 7y agoThat's not a great comparison at all though. The K80 is a general purpose chip while the T4 is explicitly marketed as an inference chip. The K80 has more ram (super important for batch sizes during training), can access that ram faster (480 GB/s versus 320 GB/s), and is an overall more powerful chip than the T4 is.
- minimaxir 7y agoThose metrics are for both GPUs on a board; the GCP K80 only uses one GPU, so those performance metrics in theory would be halved (notably, the GCP K80 has 12 GB VRAM vs. T4's 16 GB VRAM), and it's still more expensive than a T4.
- philsnow 7y agoI don't know any specifics about GCP but in general for datacenter design, heat dissipation and energy costs are really important. When you shrink the process width, you get incredible savings in power (saves $$) -> less heat dissipation per flop (saves $$ because you don't have to remove that heat). That and I suppose there could be a premium for older setups of any kind if for whatever there are licensing agreements that are negotiated for certain hardware (or for certain numbers of cores).