Hacker News Reader

Comment by Datenstrom | Hacker News Reader

> new Intel chip features for machine learning tasks and new Intel memory technology being with customers such as German software firm SAP SE (SAPG.DE) could give Intel an advantage in those areas.

I hope AMD turns their attention to machine learning tasks soon not just against Intel but NVIDIA also. The new Titan RTX GPUs with their extra memory and Nvlink allow for some really awesome tricks to speed up training dramatically but they nerfed it by only selling without a blower style fan making it useless for multi-GPU setups. So the only option is to get Titan RTX rebranded as a Quadro RTX 6000 with a blower style fan for $2,000 markup. $2000 for a fan.

The only way to stop things like this will be competition in the space.

View on HN

Their GPUs already have really good performance for machine learning tasks, but it doesn't seem like there's a whole lot more they can do about people overwhelmingly using CUDA. Their ROCm software stack has gotten reasonably good but getting developers to buy into is hard with how much inertia NVIDIA/CUDA has.

Why cant AMD make their gpus to be CUDA compatible?

If I remember correctly, CUDA is pretty tightly tied to Nvidia’s GPU architecture, so it would be very difficult to get the same things to run on AMD chips without them making huge changes to conform to Nvidia’s way of doing things.

I think there have been a few projects that try to translate CUDA stuff into OpenCL or other AMD-compatible compute platforms.

I'd say it's likely because NVIDA's CUDA compiler produces code for NVIDIA GPUs, and AMD would have a lot of (questionably legal) reverse engineering ahead of them in order to support the same code on their own GPUs.

If you're asking why AMD doesn't make a compiler for CUDA source code that targets their own GPUs--that's basically what ROCm currently does. They're pushing their CUDA alternative, called "HIP", which is essentially just CUDA code with a find-and-replace of "cuda" with "hip". (And other similarly minor changes.) They have an open source "hipify" program that does this automatically (https://github.com/ROCm-Developer-Tools/HIP/tree/master/hipi...).

So, basically, AMD GPUs are already sort of CUDA compatible: just run your CUDA code through hipify, then compile it using the HIP compiler, and run it on a ROCm-supported system (which, for now is the most spotty of all of these steps IMO).

ROCm is supported natively in TF 2.0, so they will benefit from that.

Is it? TF2.0 Alpha and Beta release notes don't mention anything about it, and the official docs at https://www.tensorflow.org/install/pip are quite explicit that the tensorflow-gpu supports only CUDA cards.

The last time I looked at ROCm (which was something like a year ago) it was "supported" as in "there's reports on the internet that someone got it to work" but when I tried, I couldn't get it to work and it really wasn't worth the effort. If I'd dig out a machine with an AMD GPU right now, can I get it working (like, train MNIST or some other helloworld'y system) within an hour, and is there documentation available on how exactly that should be done?

https://pypi.org/project/tensorflow-rocm/

>So the only option is to get Titan RTX rebranded as a Quadro RTX 6000 with a blower style fan for $2,000 markup. $2000 for a fan

for $2000, surely you can install an aftermarket cooler or watercooling loop?

If I was doing a personal build in a tower I would probably look at a watercooling loop. The build is 8x GPU rackmount for a small business though and I have not found an aftermarket blower style cooler anywhere. I have found a discussion where someone brought a rackmount case to a machine shop and had custom airflow guides put in between the cards and a hole cut in the top of the case to direct airflow out of since the backplate prevents normal airflow [1].

[1]: https://forum.level1techs.com/t/cooling-8x-titan-rtx-in-a-se...

I am sure it is aimed at Intel fanboys if you know what I mean :). The people who live by Intel as a brand no matter what and will defend its superiority regardless of facts. And yes for 2000$ one could probably build a liquid cooled super rig. I for one am happy there is finally some competition.

NVIDIA won't ever come out with a blower style Titan again. Look at the changes they made to their EULA for the driver stack as it explicitly states that it is not to be a data center product. Which is a shame really.

It kind of amazes me that there can be a EULA for software that is absolutely required in order to use physical hardware that you’ve purchased. Guess that’s the world we live in now...

I mean would it make you feel any better if they just put the requirement on the hardware itself?

I mean people seem to get non-commercial software but why is non-commercial hardware weird?

I mean it makes sense if they just choose not to sell the cards they don’t want in datacenters to people who will use them that way.

What weirds me out is that you could have purchased the card, eventually decide to use it in a way that goes against their EULA, and they could just decide to take away your access to the software required to use the hardware.

I guess it’s more or less the same feeling as certain people have about Windows licensing or buying games on certain platforms. You’re buying a license to use whatever it is rather than actually buying the thing, and this kind of feels like an extension of that to hardware.

It doesn't stop the 'raid the local Microcenter' problem though -- which I was guilty of at my last job. As it turns out people realized that at scale it was actually cheaper to just buy out big-box stores and play the silicon lottery with GeForce cards than buy Nvidia's high-end cards.

Now look, Nvidia wants to make more money, let's not pretend that there's really any other primary motivation here. And segmenting the people who make money with their cards and the people who use them for entertainment is a pretty solid way to do that. However, the secondary reason for this is that on the consumer side people complained that stores were constantly sold out of new cards from large businesses buying them all up. Some stores implemented rationing schemes but the 'final solution' it seems is to just stick a line in the license that says you can't stick these cards in your DC.

OTOH, Intel found ways to segment without attempting insane things like "oh but I don't allow you to put a Core i5 in a datacenter".

A kind of clause which might very well be void in tons of jurisdictions, by the way.

Because software you license, hardware you buy.

Mainframe customers face more onerous licensing terms from their vendors.

I get the sense that is more "you need to pay for data center usage". The GPU manufacturers used to make a ton of money having special drivers for industry. That used to be for CAD, but over the years they lost that cash cow, and now they want it back for ML volume customers.

SAP is a steaming pile of garbage and needs to die. Any company that has SAP and isn't in the midst of getting off of it needs to fire their CTO.

SAP is a prime example of how attempting to be all things at once can make a program ridiculously unusable.

Reply on news.ycombinator.com