Comment by agibsonccc | Hacker News Reader

Comment by agibsonccc | Hacker News Reader

Parent Full thread

Just going to say, just because tensorflow supports it won't mean much. The ecosystem as a whole is still largely cuda. This includes the database tech(kinetica,mapd), the resource managers (mesos/yarn), not to mention HPC.

There's also just not much incentive for people to move to AMD here. There's a reason NVIDIA gives out these graphics cards like candy to academia. It's because they want to make their margins in data center while keeping the broader developer community locked in to cuda.

People on HN make broad and sweeping comments about "if you open it they will come and everything will be magically better".

It's a lot more complicated than that. The market incentives just aren't quite there yet. Could it happen one day? Yes. Will it happen today? No. It's going to take a lot more than this for other vendors to start supporting AMD.

What I will say: There will be competition and the space is heating up. AMD is one player.

Now let me put my commercial hat on here: What would it take for me as a deep learning vendor to support/care about AMD?

1. Show me the money. I need a clear revenue stream. AMD has some hope here. Customers don't care about which gpu they use as long as it fulfills a use case they care about.

2. Code: Show me something that's actually a robust. A 1 off fork of any deep learning framework (see the random 1 off caffe forks that aren't actually caffe by nvidia, intel,amd)

3. Share the burden. Put the code out there and support the broader community. Actually maintain and follow up with the latest innovations (hint: this is hard. throwing code over the wall once doesn't mean anything)

4. Amazon will be key here. Google cloud (look google is great but they aren't the leading cloud player by a long shot and likely won't be anytime soon) - get them on board with some AMD cards.

5. (Disclaimer I know the founders) - make opencl not suck. https://vertex.ai/ is an interesting player in this space.

So look, I won't say it's impossible. Let's just not ignore the actual commercial market forces that are also at play here.

There used to be this gigantic company that ruled networking world. Its name was Cisco. No one could touch it. Costs for a basic series XX with 4 slots was N, they charged X, for 6 slots, it was N+20, they charged 4X, for 16 slots it was N+200, they charged 30X.

What is Cisco is the current question.

> What is Cisco is the current question.

A $190 billion company earning 4x what nVidia does, a mere $9-$10 billion in net income, with 12x the cash of nVidia. A gigantic money printing machine, is what it sounds like they are.

Have you heard about the company called Juniper? Brocade? Arista? The only reason they got anywhere is because of cisco's "We are cisco.Eff you!" attitude.

Nowhere did I say it's never going to happen. These things take time to evolve. I even talked about the market forces and what's currently plausible.

There's always big companies that get disrupted because they kill good will. That's the start to something. It takes more than 1 move. Will NVIDIA make this mistake over time? Likely.

Will it be AMD and Tensorflow support that does it? Not by a long shot. It will be by a series of players providing the right incentives. My response (nor any response to this) should not be treated as binary. Ultimately in order for a company to be "disrupted" you need to actually analyze and exploit the market forces involved here. just writing this off as "open ecosystem is all we need" is dangerous.

I have to say I disagree with you on the AMD/TensorFlow point here, Adam. Your previous points are completely valid. Nothing of note will happen in 2018, I expect. But by early 2019, maybe - if AMD (or Intel) gets their act together on the software side of things. I don't think the "community" will do it for them. Buy maybe a big player like Amazon will have had enough of Nvidia and support an open alternative. Although we support Cuda in our version of YARN (Hops Hadoop), I expect we'd add ROCm or OpenCL or whatever - if it was a serious alternative. That would happen quickly. The problem, of course, is that we would want great support in TensorFlow first before we do that. Data scientists need a seamless transition - including from a performance perspective. For us, that also means support for distributed deep learning (Ring AllReduce over infiniband). I don't expect that will happen in 2018, and could take until 2020, if i'm being realistic. That means when AMD finally get some good DL libraries, Nvidia will still have one-up on them with distributed DL (reduce training time linearly with more GPUs).

The other wildcard i haven't seen people mention here is the Neural Network Processor (NNP) from Intel Nervana. The hardware has potential. As long as the software doesn't force us to use BigDL, it has potential.

I'd expect competition to heat up in 2018.

Re: AMD/Intel. We've been waiting for them to get their act together for years. Nervana could be great but I'm going to wait on that one. So far their "launches" have been nothing more than marketing fluff.

As for your projections about tensorflow. It won't be tensorflow. Tensorflow will be 1 of many frameworks. Look I like HOPS but you guys push tensorflow explicitly. A startup running its own hadoop distro that happens to push tensorflow isn't going to move the needle. You guys are great middleware I'm sure but I haven't seen the customers where it might be viable. I hope you guys continue to grow though! While you're doing that, most hadoop vendors are focused on moving up the stack. It will take people with actual resources to move the needle in terms of enterprise adoption.

Amazon is doing this with mxnet and EMR, MapR is pushing tensorflow in their serving. CNTK is being pushed in SQL server and HDInsights. There's some competition there.

What I'm getting at here is: it will take multiple vendors and competition. I'm going to place my bet on the bigger players already involved with the foundations first though. Open standards (addressed below) where it commoditizes the chip will be key. The storage infra will follow from that. It should be something that doesn't displace current infra but allows interop.

Things like nvvm from the mxnet folks, onnx (where the framework doesn't matter anymore!) being pushed by the various hardware vendors etc will move the needle. You need buy in from the actual big players who can upfront the development time in to making these things viable alternatives.

For your "seamless transition" I'm not sure that would be that hard done right. Supporting "great tensorflow" can come in multiple flavors. As a separate issue, tensorflow's production story is horrible. That's another topic I could rant about all day though. It ultimately comes from abstracting it away though. That by itself is a hard problem (Disclosure: I have my own solution for this that I won't talk about here just know I'm biased :D)

Lastly, I question whether opencl can even be a viable alternative. It's a fragmented inconsistent standard with a worse API than cuda. One reason it "won" is because it's in general cleaner and a clear leader in the space.

Yeah, ROCm is the most viable candidate as of today. In general, Nvidia are not good for middleware vendors. They want to be one, but don't offer a platform that integrates with anything. Licensing costs for the DGX-1 are insanely high. My problem is mostly from a data scientist perspective - teams don't need a few high performance GPUs, like a couple of DGX-1 boxes. They need a hundred 1080Tis, maybe complemented by a DGX-1 (which would cost the same a 2 DGX-1s). That way they can do lots of parallel experiments (hyperparam optimization), and distributed training. Making GPUs a scarce resource just re-inforces the lead of the hyperscale AI companies, who have thousands of GPUs available for their data scientists.

Oh I agree that GPUs should be more commodity. We might see alternative ASICs rather than GPUs come out though. I'm personally more interested in seeing that succeed than just confining the solution space to gpus and discrete gpu competition. I'm just not keen on trying to predict what will win (I really don't know) I just have criteria I would be looking for before trying to implement support for it in either my deep learning framework or trying to support anything for customers.

A dick move like this by Nvidia will only spur the development of alternatives to CUDA. CUDA isn't the only game in town but until now nobody had much reason to invest in improving the alternatives - CUDA was free, worked and the hw was cheap and ubiquitous.

Right so like my other replies, I don't claim that alternatives won't come up eventually. I also don't claim that this is exactly how nvidia can unseat themselves.

What's wrong here with many of the guesses people are making though: They think simple open standards are enough.

It's a lot more than that. Don't alienate the market forces that also need to move to make this work. Too many coders write off the real business incentives side of this.

Viable competition will come from multiple vendors and possibly backwards compatible/interchangeable standards that don't require many changes in code. Give me a path forward as a vendor. Show me the money. Put code out there and maintain it/keep it up. Show me it's going to stick around.

It's like any SAAS you'd pay for or an IDE you invest in, vendors want to see an ecosystem that fulfills a set of requirements and allows them to get work done for their customers.

Right now cuda is still that best tool. I covered what will likely need to be considered in my parent post so I won't reiterate that here.

I only ask that folks don't assume that "open source linux drivers and open standards and 1 open source framework with opencl support" are enough to unseat nvidia. The market won't move for that.

btw since you know the founders of vertex.ai you should tell them their cert is messed up they probably want to do something like https://hackernoon.com/set-up-ssl-on-github-pages-with-custo...

Thanks will pass it on!

Reply on news.ycombinator.com