Why can't AMD or Intel make an alternative to CUDA?

>CUDA is used for deep learning because nvidia developed tools for it.
That is outside the reach of the consumer market. Do YOU have fucking $10,000 to fork over? Also even in the HPC market CUDA is still struggling to compete with AMD epyc servers because despite quadro cards having more FP64 on paper than a DP epyc server blade it's very hard extracting all that performance out of 5,000+ GPU cores than it is to out of 64 CPU cores.

bhphotovideo.com/c/product/1402941-REG/pny_technologies_vcqgv100_pb_quadro_gv100_graphic_card.html

serversdirect.com/Servers/id-SD-1123US-TR4/Supermicro_A_UltraServer_1123US-TR4

youtube.com/watch?v=d94N2Lu4x9s
The idea is to create a vendor neutral stack which could be used with any GPU be it from Intel or AMD or NVidia. That would be a good thing. Right now it's nothing more than a proposal.

*blocks your path*

Attached: it just works.png (500x347, 75K)

You mean like OpenCL, SYCL, C++AMP, HC, or HIP?

>HIP
This

AMD has OpenCL, but Nvidia is spending a lot of money on devs to keep the on CUDA. Nvidia being Nvidia at keeping their monopoly.

On top of it, Nvidia artificially make their OpenCL cumbersome to use to push devs towards CUDA. But if you want something hardware-agnostic OpenCL is the way to go.

It's pretty sad given how exponentially much more dual precision compute performance AMD cards have than Nvidia ones for a lower price. vega 64 has more than 3X the FP64 than the rtx 2070.

techpowerup.com/gpu-specs/radeon-rx-vega-64.c2871
techpowerup.com/gpu-specs/geforce-rtx-2070.c3252

I, too, had my entire life ruined by nvidia

>fucken wood screw motherfuckers
what did he mean by that?

>Being this retarded

hello newfag

>wood screw mother fuckers.
Massive kek
I wish OpenCL got to market a little earlier. Could've came out before CUDA got the monopoly of the market established.
if only "MUH GAYMES" actually used the true raw power of amd... I see the teraflops and wonder why they aren't like miles fucking better...then in some gaymes they actually are...too bad gaymurs just say "amd cards suck" rather than put 2+2 together and say "This software is really poorly written and isn't using the hardware"

>I wish OpenCL got to market a little earlier. Could've came out before CUDA got the monopoly of the market established.

OpenCL was a shitty CUDA clone proposed by Apple who didn't like the fact that CUDA effectively locked them to a single vendor. They literally didn't even get a design committee together for an entire year after CUDA was out in the wild (sampled earlier). And then Khronos just shat something out in a few months and it's been considered a mistake ever since despite non-Nvidia companies constantly pushing it

Proprietary
I've suffered a lot because of this.
No choice but Nvidia to train my neural networks.

>CUDA based video encoding for example is about the same as SW video encoding using the ultra fast preset.

NVENC is not CUDA based, it uses a "fixed function" encoding accelerator. Apparently Nvidia provides a library that has some features that use CUDA for enhancements but it still relies on function specific hardware for encoding. Also, as of Turing's "improved" NVENC, review sites are saying it looks about as good as medium speed CPU encoding, which blows all of its competitors out of the water.

>OpenCL
its crap and about 5 years behind cuda

>train my neural networks
deepfaking an ugly celeb's face on a pornstar I presume

It's fucking proprietary. They're not allowed to by law.

I unironically think a lot of low level stuff developed by RedHat and their employees are of good quality.

In theory OpenCL was the choosen one, AMD,Intel and Apple just build shitty implementation only C language, Apple goes full Metal, AMD made RocM, OpenCL makes a lot better for 2.2 plus high level language with SyCL but nobody implement it, New Intel strategy is wait Vulkan get full compute capacities OpenCL, Intel developer new libs vulkan drive for OpenCV.

Intel will tried Vulkan,AMD full autism RocM and Apple avoid Vulkan and bans CUDA/OpenCL for new Metal 2.

AMD did. OpenCL and Vulkan computer exists.

Attached: 1548016194010.jpg (400x346, 19K)

You know that most of machine learning research is done on consumer cards, right?
Fucking Nvidia sold a ton of 1070s and 1080s only because stupid researchers still haven't realized monopolies and proprietary technology are not good.

Is PlaidML any good? I hear the support for opencl cards is now there.

I've never seen an HPC system with AMD cpus. Only Intel.

>I wish OpenCL got to market a little earlier. Could've came out before CUDA got the monopoly of the market established.
Ironically, China is the only real hope for OpenCL. They would rather "cheap out" and not pay premium on proprietary software.
No one cares what you have seen. You are a literal nobody with a made up anecdote.

>stupid researchers aren't buying my amd garbage and waste time while trying to make it work, but something that works for them
wooooow

Idiot

Maybe because competitive amd cpus have been ou for less than 2 years?
BTW the most powerful cpu-only supercomputer will be amd, at the HLRS

It's called opencl

Does AMD provide the same level of software support as nvidia for cuda? No. AMD went literallly "just fix our shit lol"
Now kill yourself.

>You are a literal nobody with a made up anecdote.
Fuck you, here's your spoonfeeding: top500.org/list/2018/11/

>AMD will become relevant, guys, I swear!

>2 top spots are IBM
>number 3 is chink
>1 Faildozer CPU in top 10
>I have seen only Intel
way to confirm you are a NEET.

Do people really use this?

Yes. I've been using it for around 3 years.

OpenCl for AMD, CUDA for Nvidia

It's the same shit, just compile your code for whatever target system you're going to use.

Or do what everyone else does, add support for both.

I haven't run into code that only works on one platform. Just code that runs more efficiently/faster on one or the other.

Software vendors are not stupid enough to chop their user-base in half.

>Yes. I've been using it for around 3 years
Link. Want to try.