Hacker News Reader

Comment by jacquesm | Hacker News Reader

How does that compare dollar-for-dollar to running Tensorflow on a dedicated Linux cluster with GPUs?

Mainframes also have a lot of PCIe channels you can attach GPUs to. IO capacity and hardware offloading has always been the most notable difference between a mainframe and an x86 box.

I wouldn't be surprised if, with some clever programming, one could pipe data from disk directly to GPUs and back to disk without involving the CPUs that run user programs except for the thinnest slice of time.

Do the specific Nvidia cards you'd ideally want for ML have s390x drivers?

That's something you'll need to ask Nvidia.

But, considering what you paid for the mainframe, you probably have the budget to commission them to write drivers just for you (or to license their source and compile it yourself).

(or to serve them a National Security Letter that says "give us your source" and compile it yourself)

Can’t stop laughing sadly.

Btw given they do linux can they twist so that channel is partition into for that linux container. In the new world just driver through.

IBM has built supercomputers with POWER machines stuffed full of GPUs, in close partnership with Nvidia, for years. I think they might still be the only ones offering systems designed for sustained NVLink bus saturation. So yeah, you could do it - and it would be awesome... but it would be a silly misuse of a mainframe where a cluster would make so much more sense.

Probably much more expensive.

Probably depends on whether you've already got the mainframe or not, and whether you've got the dedicated Linux cluster or or not.

Reply on news.ycombinator.com