[foobar33333]

Nvidia is basically dead at this point, no support for standards that have been around for years. And now they are including malware in their cards to slow down unapproved calculations.

[daniel-thompson]

> Nvidia is basically dead at this point

As of this morning's Nvidia's market cap ($320B) is bigger than AMD's ($100B) and Intel's ($255B).

[curt15]

CUDA is an industry standard, and AMD's lack of support is why NVIDIA enjoys a monopoly in the deep learning community.

[opan]

It's proprietary. People being vendor-locked into a proprietary thing doesn't mean everyone can or should support it.

[MaxBarraclough]

Is this due to CUDA being a technically superior product, or is it just that it's entrenched?

OpenCL/AMD is a usable and mature platform, right?

[my123]

OpenCL as implemented by AMD is definitely not very usable.

AMD themselves aren’t really committed to OpenCL given the state of their OCL 2.x drivers and went to their own AMD-only path (ROCm/HIP), which is broken... (and only supported on a single gen of customer GPUs today, Vega) It doesn’t even work on RDNA or RDNA2 officially today.

[solarkraft]

I've been told that OpenCL is in a whole different league than CUDA due to a worse API and much smaller ecosystem.

I have, however, seen attempts to use Vulkan for compute. Could that be a viable way forward?

[MaxBarraclough]

I'd say OpenCL is a pretty good API, it's just that because it's in C and it allows for quite fine-grain control, it takes a bit of orchestration to get anything off the ground. If you want more convenience, there are a few wrappers for it, in various languages.

It's true that its ecosystem is sadly lacking.

[my123]

Using Vulkan for compute is still a much worse developer experience, it isn't really a reasonable alternative.