Hacker News Reader

Comment by eachro | Hacker News Reader

How is AMD doing in the GPU for deep learning business? Have they completely ceded that market to Nvidia?

They've made a CUDA competitor (https://rocm.github.io/) and it works well enough to port Tensorflow to (https://github.com/ROCmSoftwarePlatform/tensorflow).

Not super impressive, but we'll see what happens I guess.

Is running TensorFlow on GPUs still relevant now that google has their TPUs? Other dedicated hardware is also coming.

Yes. It will always remaining relevant. Google wants a continued supply of GPGPUs from anyone who will sell it to them. They will take TPU to the max, but they are not in it to compete. They are in it to protect themselves first and to have leverage.

yes. GPUs end up being far more flexible and the latest generation data center GPUs (Volta V100) have dedicated tensor cores. Also, not everyone can (or wants to) use Google Cloud.

I don't really follow the ML space, but when you have a good GPU in your PC (which I assume quite a few people here have) you can learn/use Tensorflow by running things locally instead of paying for Google Cloud etc. Why pay extra money if you already have a good GPU in your computer (unless you really need to decrease the time it takes to run the computation etc.)

They emphasized on their earnings call that their biggest increase in spending is on the software side. They know they have a lot of work to do there and are investing in it, but that it will be a multi-year process.

I believe CUDA severely hampers AMD's efforts in deep learning.

AMD's primary focus is on Epyc building presence in data centers. They are aiming for 5-6% market share by end of 2018 and 15% by the end of 2019.

Frankly, this is starting to become less of a hurdle in that pure CUDA code is now rarely written in the DL community (instead opting to call optimized libraries like CuDNN). The biggest issue is replicating all the good "core computing libraries" that NVIDIA offers: CUBLAS, NPP, CuDNN, Tensor-RT, etc.

Frankly, if AMD can start putting this stuff together over the next 2-3 years it's usually not too bad to have deep learning frameworks leverage this stuff.

Their GPUs for deep learning are nearly non-existant, as others have mentioned there are projects and support, but... not worth the effort essentially.

On the other hand, with the rise of their cheaper CPUs and APUs, I'd actually expect that to change. From a power perspective and eventually the ubiquity, it'll make sense to provide more support.

I will say you can use their gpus for deep learning. Not super performant and stuck on tensorflow 1.4, but it's not terrible, and certainly better than CPU.

It is extraordinary how much AMD is faltering for deep learning, with most projects being built for only nvidia. Their ROCM/HCC thing is neat, but at this stage I would expect AMD to have large scale involvement with all major deep learning projects, ensuring that they fully leverage AMD GPU hardware. Instead it's just a couple of half-baked, half-ready sort of ports.

How is it extraordinary at all?

AMD was literally going bankrupt like 2 years ago. They fired 30% of their staff and then reverse-mortgaged their headquarters. They bet their entire company on "Zen" and... luckily, it paid off.

https://www.theverge.com/2012/10/12/3496006/amd-layoffs-10-p...

https://arstechnica.com/information-technology/2013/03/amd-s...

AMD finally has money to actually invest into long-term projects. But AMD just didn't have the money to do so between 2012 and 2016.

Unfortunately, it will take until 2020 before AMD puts something together. It just takes a long time to build software and hardware. But ROCm shows how AMD is beginning to accelerate, finally, into the GPGPU compute sector.

I think the question is how many truly great engineers would it take to build some of the necessary APIs. Why spend hundreds of millions in hardware R&D and then underinvest in software that makes the hardware valuable. While the CEO has been doing a stellar job Turing around the company it still feels like they underinvested in software.

I'm not sure if you really understand the peril or how deeply AMD cut back on their R&D Efforts.

Small Cat (Puma cores) were cancelled and that entire department was laid off. This is the core that won AMD the Xbox One and PS4. Small Cat was also their low-power architecture designed for tablets (notice that AMD literally gave up on the growing Tablet market?)

AMD cut DEEP to stay out of bankruptcy. They were cutting known, successful projects and were still losing hundreds of millions of dollars per quarter.

Lets be frank here: if AMD were today, in a position where they had better ROCm or OpenCL or Tensorflow support but say... no "Zen" or "EPYC" ?? They'd be worse off. The CEO made the right bet to focus on Zen as their comeback strategy.

Zen is successful because it demonstratively does something better than Intel. It also doesn't need much software support or tooling, and was designed to be relatively cheap to make. I don't think AMD would have done the same with the NVidia GPGPU compute world, especially since CUDA is a proprietary technology. Even if AMD were to make something better than NVidia, they wouldn't reap any rewards because CUDA is a big moat.

Businesses "reverse mortgage" their real estate all the time, including extraordinarily profitable businesses. I also don't see what some links from 2012 have to do with "like 2 years ago", though we all know that AMD went through some down times -- the CPU side did, though the GPU side never did and kept paying the bills.

A segment of their business was promising, and they could see a hugely emergent demand growing. Instead of getting in on the projects that matter, they just ignored it turning it into an entirely nvidia market. Yes, that is irrational and self-defeating. If they were so self-sabotaging in the GPU market they should have sold that business (and it makes one wonder why they ever bought it in the first place).

Of course CUDA is a moat. AMD sat doing nothing allowing it to get dug further and deeper, becoming completely synonymous with deep learning. A small scale, inexpensive software project to keep official, supported branches of major projects wholly functional on AMD hardware (e.g. Tensorflow) would have done great things. Instead they made some completely irrational solutions that tried to move the market somewhere completely different.

This has nothing to do with how they bet on "Zen". It isn't either or.

AMD started off with massive layoffs in 2012, and continues through 2015.

https://venturebeat.com/2014/10/16/amd-misses-earnings-targe...

https://venturebeat.com/2015/10/01/amd-to-lay-off-500-people...

2016 AMD began to stabilize, as they focused on Zen. They still were losing tons of money per quarter, but they refocused their efforts. And 2017 they finally delivered. AMD didn't actually turn a profit until late 2017.

https://www.fool.com/investing/2018/02/05/amd-turns-its-firs...

Between 2011 and mid-2017, AMD was losing money every quarter. There's a reason why the stock was below $2/share for a long time.

Reply on news.ycombinator.com