Hacker News Reader

I heard the memory bandwidth is not just slower than on a GPU, as expected, but is significantly slower than Apple’s unified memory.

View on HN

CPU/GPU is decent (800 GB or so), memory is slowish (300GB or so). Some Apple M are slower, some are faster.

Where did you get those numbers from?

DGX Spark has a maximum of 273 GB/s bandwidth in ideal scenarios (hard to reach)

That puts it between an M5 (153) and M5 Pro (307)

The 900 GB/s is from the NVLink-C2C interconnect, if you were wondering about that. They quote "up to 900 GB/s of bidirectional bandwidth between GPU and CPU".

Mind you thats not to/from memory, which indeed only has 273 GB/s.

Ah I see. But the only C2C equivalent on the Apple side is the UltraFusion which is 2.5TB/s if I recall correctly.

Yes its not an "Apple M killer" at all. Also, the available official performance numbers are partially overstated (1 Petaflop is only possible for sparse FP4 models, "in theory").

Perhaps a sobering rule of thumb: if it was actually useful, you couldn't buy them because someone would scoop them all up to shove them in a DC and make money with it.

Reply on news.ycombinator.com