Hacker News Reader

Comment by hyperpallium | Hacker News Reader

Apple has a 600 GFLOPS NPU in the iPhone X. Nvidia's $3000 Titan V has a 110 TFLOPS TPU. After a couple of decades of researchers using GPU's for compute, dedicated hardware is available, with a better FLOPS/$ ratio. Looks like the future of compute.

We'll see Apple releasing more uses of their NPU for AI.

View on HN

Flops is a useless metric in general, there are benchmarks for machine learning that make valid comparisons between CPU/GPU/NPU/etc.

I wasn't confining it to machine learning, but compute in general. (I'm interested in simulation.) Are machine learning benchmarks applicable to non-machine learning computation?

I'm not understanding your comment correctly, is it sarcastic? Nvidia doesn't ship a TPU on the Titan V. Also, 0.6 TFLOPS/$1000 vs 36.7 TFLOPS/$1000 is nearly a 2 orders of magnitude difference favoring Nvidia.

It's not sarcastic. It's not contrasting the two, but pointing out a general shift towards NPU/TPU (as the comment states). Of course a mobile device isn't going to have the same power ratio as a Titan V (It's amazing that the iphone X has 500GF, in addition to its GPU and CPU. I expect it will go much higher; this is just an initial foray).

I don't understand the objection to describing the tensor cores of the Titan V as a TPU. As a GPU, it has only around 10TF. How would you describe the tensor cores?

Perhaps a more apt comparison is to the Tegra processor, and still, Nvidia crushes apple in performance/Watt.

I wasn't comparing them, but pointing out a trend common to both.

Reply on news.ycombinator.com