Hacker News Reader

Comment by areoform | Hacker News Reader

    > RTX Pro 6000 Blackwell has 96GB of GDDR7 VRAM

A mac studio with 96GB unified memory costs, $5,299.00.

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?

View on HN

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.

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.

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.

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

the last mac pro was rackable, and they discontinued it anyway.

On the other hand, they kept on crippling the mac pro's pcie capabilities. I think the only thing you might put in one of the slots would be nvme cards. I thought there were no graphics cards anymore after arm.

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

Memory bandwidth, ecosystem compatibility, compute power.

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

Is that was true, claude would be written in Rust;

Reply on news.ycombinator.com