Hacker News Reader

Comment by nnq | Hacker News Reader

I'd bet most successful hardware advances in the ML/AI will come from companies that also push the field's edge in software Bc all the others would be at at disadvantage - their "better" hardware will mismatch the software, and that mismatch will increase development costs A LOT.

Google's TPUs, Tesla's whatever those are etc. Bet on them!

NOT on whatever Intel or IBM or AMD or Arm are doing this field. Heck, even Nvidia will probably start loosing at this game in the long run. The field is too dynamic and volatile, companies need to "eat their own dogfood" by integrating vertically hardware + software, before pushing hardware to sell to others too. This is not the old game of general purpose hardware anymore...

View on HN

I think NVIDIA is in a good spot with its strong software ecosystem, and an army of devtechs. Their architecture is pretty general, so that if the ML algorithms change they adapt better than say TPU. I can't find the video of the talk anymore, but late in the process of TPUv3 IIRC Google had to rebalance their hardware to account for new algorithm(change)s.

Second, most companies that want to deploy AI cannot afford to build their own custom hardware. Even most carmakers partner with Mobileye or NVIDIA.

These big companies also have research teams whose job it is to stay on the ball, research and develop AI techniques and influence how the hardware has to change.

As for eating your own dog food, I think NVIDIA does just that with their autonomous driving software stack, robotics kit, devtechs for customers and library optimization/porting.

With sufficient focus, funding and execution I think AMD and Intel can reach a similar spot.

That said, my hunch is that the legacy and compatibility that the general purpose hardware makers have to carry forward will become a problem. But until that happens (I.e. specialized HW delivers better bang/$) in a 2-10 years, they will likely figure out how to alleviate that and/or develop more specialized hardware.

Seems that if Tesla was able to build its own custom hardware and have it working in 2019, other carmakers which are much larger and profitable would easily be able to do it, if they cared to.

Domain expertise is a finite resource and couple that with the fact the automotive manufacturers are not adept at higher level software development (as opposed to embedded)

I'm not surprised companies like Ford have continued to push out dates

Tesla's chip had 7x the performance of NVIDIA. Tesla hired a chip architect from Apple to get the upperhand.

There's an obvious first-mover advantage though... there's not that many Jim Kellers around, and they got one.

Actually Intel is at the forefront of deep learning software as well:

- MKL-DNN (rebranded DNNL) is unique and has no equivalent in the ARM space. It even supports OpenCL

- OpenCV

- MKL

- PlaidML and NGraph follow the current trend of deep learning compilers

- They had Nervana Neon as a deep learning framework which I'm pretty sure contributed to Nervana acquisition as at the time it was even faster for convolution than Nvidia CuDNN.

I'm really looking forward to their Xe GPU. Also I expect them to leverage MLIR or contribute heavily to the linear algebra dialect -> LLVM IR optimization passes.

Who is using Intel for ML? Saying you are cutting edge by some acquiring is something else than actually pushing the envelop.

Reply on news.ycombinator.com