> RTX Pro 6000 Blackwell has 96GB of GDDR7 VRAM

https://www.apple.com/shop/buy-mac/mac-studio/m3-ultra-chip-...

LLMs can re-write and cross-translate software really well. Why does CUDA still have a $11k price premium?

You can get 5-10x the perf out of the blackwell under the right workloads. It has faster VRAM (> 2x) and can do a lot more matmuls (>> 10x).

They're both good value (or crazy expensive) depending on how you look at it.

It depends how you price the ability to run a particular model at all, vs run the model quickly and serve several parallel streams.

The Blackwell is going to be a lot faster for the same memory capacity than a Mac Studio, surely?

Because people still pay for them, and mac studios aren't usable on server racks, as classical UNIX variants.

Maybe Apple should bring OS X Server back, with a rack like drawer.

CUDA ecosystem is more than running PyTorch.

Because a lot of people are buying them at that price.

Memory bandwidth, ecosystem compatibility, compute power.

As others have stated, the Blackwell runs circles around the Mac performance-wise. I haven’t run the numbers for this specific comparison, but in some cases the nvidia solution is also more efficient (tokens/W). Some people are willing to pay that premium for the above, just like they’d pay more for a 4-door sedan from BMW as compared to a Toyota Corolla.

  LLMs can re-write and cross-translate software really well.