Apple's MLX adding CUDA support

I wonder how much this is a result of Strix Halo. I had a fairly standard stipend for a work computer that I didn't end up using for a while so I recently cashed it in on the EVO-X2 and fuck me sideways: that thing is easily competitive with the mid-range znver5 EPYC machines I run substitors on. It mops the floor with any mere-mortal EC2 or GCE instance, like maybe some r1337.xxxxlarge.metal.metal or something has an edge, but the z1d.metal and the c6.2xlarge or whatever type stuff (fast cores, good NIC, table stakes), blows them away. And those things are 3-10K a month with heavy provisioned IOPS. This thing has real NVME and it cost 1800.

I haven't done much local inference on it, but various YouTubers are starting to call the DGX Spark overkill / overpriced next to Strix Halo. The catch of course is ROCm isn't there yet (they're seeming serious now though, matter of time).

Flawless CUDA on Apple gear would make it really tempting in a way that isn't true with Strix so cheap and good.

For the uninitiated, Strix Halo is the same as the AMD Ryzen AI Max+ 395 which will be in the Framework Desktop and is starting to show up in some mini PCs as well.

The memory bandwidth on that thing is 200GB/s. That's great compared to most other consumer-level x86 platforms, but quite far off of an Nvidia GPU (a 5090 has 1792GB/s, dunno about the pro level cards) or even Apple's best (M3 Ultra has 800GB/s).

It certainly seems like a great value. But for memory bandwidth intensive applications like LLMs, it is just barely entering the realm of "good enough".

You're comparing theoretical maximum memory bandwidth. It's not enough to only look at memory bandwidth because you're a lot more likely to be compute limited when you have a lot of memory bandwidth available. For example, M1 had so much bandwidth available that it couldn't make use of even when fully loaded.

GPUs have both the bandwidth and the compute. During token generation, no compute is needed. But both Apple silicon and Strix Halo fall on their face during prompt ingestion, due to lack of compute.

Compute (and lots of it) is absolutely needed for generation - 10s of billions of FLOPs per token on the smaller models (7B) alone - with computations of the larger models scaling proportionally.

Each token requires a forward pass through all transformer layers, involving large matrix multiplications at every step, followed by a final projection to the vocabulary.

Obviously I don't mean literally zero compute. The amount of compute needed scales with the number of parameters, but I have yet to use a model that has so many parameters that token generation becomes compute bound. (Up to 104B for dense models.) During token generation most of the time is spent idle waiting for weights to transfer from memory. The processor is bored out of its mind waiting for more data. Memory bandwidth is the bottleneck.