That's a pretty gen-eric name. Runway has the same problem.

This driver doesn't support CUDA.

This comment should be pinned at the top.

Isnt mlx a cuda translation later?

No, MLX is nothing like a Cuda translation layer at all. It’d be more accurate to describe MLX as a NumPy translation layer; it lets you write high level code dealing with NumPy style arrays and under the hood will use a Metal GPU or CUDA GPU for execution. It doesn’t translate existing CUDA code to run on non-CUDA devices.

My understanding is that MLX is Apple’s CUDA, so a CUDA translation layer would target MLX

No, it’s not. MLX is Apple’s NumPy more or less.

Does tinygrad support MLX?

Pretty misleading. This driver is only for compute not graphics.

As a sizable share of the market is going to want to use this for local LLMs, I do not think this is that misleading.

Most people I know are not using TinyGrad for inference, but CUDA or Vulkan (neither of which are provided here).

GPUs can do graphics too?

Graphics was not what came to mind when I saw the headline.

Graphics is typically what comes to my mind when people talk about graphics processing units

The latest MacBook Pros don’t even need external GPUs to run AAA games.

What do you mean when you say "run"? Low graphics 45 FPS at 720p? Or ultra graphics 120 FPS at 4k? My assumption is that a fairly large part of that space is inaccessible with the integrated GPU.

The term eGPU gives it away, but is inaccurate.

Something like eNPU or eTPU seems more appropriate here.

You probably need $100K of hardware to run Sora.

Full quality Sora yeah probably needs serious hardware. distilled version on a 4090 though? maybe. danjl earlier in this thread made a solid case for just distributing the weights and letting people run locally. the SD/ComfyUI crowd already does this daily. OpenAI won't because deepfakes. they already had a mess WITH server-side moderation. open weights with zero moderation, good luck with that PR.

> OpenAI won't because deepfakes.

Do you think someone would spend 5 or 6 figures on a license and hardware to create deepfakes?

People pay for OnlyFans accounts; why not this?

Most people just want emulation which was solved a while ago. Recompilation is mostly for modders which is a very small community.

The whole decomp and recomp scene has really started popping off this past year in the emulation community. The N64 recomp tool was the first one to really get interest going.

PowerPC doesn't have the organic ecosystem that ARM has.

Not now. But my question is, what was stopping IBM from doing what Arm did? We are where we are now and it's too late. But as far as I can see, there was nothing too special about Arm as compared to PowerPC back then, on a technical level.

IBM wasted plenty of effort on Itanic but at least they were smart enough not to cancel any of their architectures.

Probably Intel and AMD aren't willing to do this deal but Arm is.

IBM actually owns x86 rights still. They last used it to do something similar called Lx86 which ran x86 VMs on POWER CPUs.

Developing a good x86 CPU is far beyond IBM's abilities. The rights aren't enough.

Price competitive to AMD and intel? Sure. Abilities? There is no magic, the Tellium and Power11 are each as complicated as something like Epyc and the former has both a longer and taller compatibility totem pole than x86.

Anyway this post was never about building ARM or x86 CPUs, the point is they could have done a zArch fast path for x86 for "free", so there is some other strategy at play to consider doing it with ARM.

Maybe a better link: https://www.arcee.ai/blog/trinity-large-thinking

They could have had file permissions. They could have had a package manager instead of third-party installers.

Also note that Microsoft Office has a long history of not following Windows rules. Microsoft didn't even set a good example.

In the era of Windows 95, having a network connection was still a rarity. Expecting modern systems to have package management and sandboxing mechanisms would have been 20 years ahead of their time.

They had. Sort of. MS-DOS was a single user system. See attrib for details.