Earlier quoted context omitted.
> If they make that chip I have no doubt that it will revolutionize the entire field. Ha, ha. Yeah… no. For one thing, there's no mention of its power consumption, no CUDA support, questionable memory design (sure, you can get a million TFlops without cache, now try to get that chip to do anything useful), etc. Intel probably bought 'em to work on integrated GPUs or Xeon Phi or something.
> Intel probably bought 'em to work on integrated GPUs or Xeon Phi or something. Ha, ha. Yeah… no. > there's no mention of its power consumption Can't be tremendously higher than Pascal for reasons of physics. > sure, you can get a million TFlops without cache, now try to get that chip to do anything useful "Without cache" is certainly an exaggeration. It won't have a globally coherent cache hierarchy in the style of…
> Ha, ha. Yeah… no.
Oh? You don't think they'd acquihire a machine learning startup to work on their compute/machine learning offering in Xeon Phi?
I would be very, very surprised if Intel introduces a new machine-learning-oriented processor—though that would be quite interesting, so let's hope you're right.
> Can't be tremendously higher than Pascal for reasons of physics.
I wouldn't be so quick to assume that. Nvidia spends a lot of resources optimizing power efficiency. To think a much smaller company could match their power efficiency while scaling performace¹ is plausible, but not likely.
If anything, physics would say that it can't be much lower than Pascal, but nothing about higher. For example, the large gap in power efficiency between Nvidia and AMD's 28nm GPUs last generation. It's not correct that similar size process nodes always have similar power consumption—that comes down a lot to architecture.
> It won't have a globally coherent cache hierarchy in the style of CPUs. It certainly will have various on chip memories to hold intermediate results.
You're right, the link² posted in another comment clears that up—and you're right that it could be very good for performance, but also very difficult to optimize for.
> You're just being silly now.
You're spouting stupid shit (like that bit about physics and power consumption… not to mention the ridiculous hyperbole in your original comment) and call me silly. I'm not sure if you're a troll, an overly zealous brogrammer buying into deep learning buzzwords, or retarded, but maybe you should cut it out.
> CUDA isn't a standard, it's proprietary to NVIDIA and this isn't a general purpose processor anyway.
Which is precisely why it won't revolutionize the field at all. Deep learning is relatively locked in to CUDA.
Essentially, extraordinary claims (such as revolutionizing deep learning) require extraordinary evidence—evidence which neither you nor Nervana have shown the slightest signs of producing.
--
¹ I just noticed the nextplatform link specified teraops instead of teraflops. I don't know if that's referring to integer operations, would be… unusual, to say the least, for a deep learning chip. If it is, then we don't really know what its single-precision float performance is.
² https://www.nervanasys.com/nervana-engine-delivers-deep-lear...