Ljink96 said:
The thing is, the document says 1024 Flops. So I can't really assume at this point. I kinda doubt switch would be .5TF, we've heard .75 from digital foundry. The document is confusing. It really doesn't specify if it's a Megaflop, Gigaflop or at FP16 or 32 but you are most likely right. It could be 1024Gflops and be 1TF FP16, or 1024Gflops FP32, aka 512TF. But do hardware manufacturers default to FP16 when noting specs? |
I think the idea is to optimize software for FP16 so that you really do have 1TF worth of compute power.
If you look at spec sheets provided by nVidia for Tegra, they list both FP16 and FP32 tflops. We're used to seeing FP16 / 2 = FP32 flops, but that isn't always the case. The Tegra K1 only had ~350 gflops in either FP16 or FP32 modes because the Kepler architecture couldn't even leverage FP16. Starting only with Tegra X1 did they do anything to take advantage of FP16.
People have been reeeaaally quick to write off FP16, but seldom give you valid, scientific reasons why.
I predict NX launches in 2017 - not 2016
