Hacker News Reader

Comment by asdff | Hacker News Reader

512gb being more expensive than just two 256g is a little strange. You mean you can get the same exact total cluster ram but double the compute/storage nodes and save $700?

View on HN

Which is an especially interesting option given built-in RDMA support with reasonable performance, including fully-connected graphs up to 6 machines…

Didn't take me long to find a nice proof of concept, 4x mac studio cluster in this case:

https://www.jeffgeerling.com/blog/2025/15-tb-vram-on-mac-stu...

you're still limited to TB5's 120Gbps speeds vs 1.2Tbps memory bandwidth on M5 Ultra

What sort of jobs are memory bandwidth bound that people are running? Will they actually see 10x speed improvement or hit some other real world bottleneck? Are these jobs "short" on the order of seconds or minutes or is the node being hammered for two weeks straight? I'm just curious about it.

> What sort of jobs are memory bandwidth bound that people are running?

CFD, for me. Depends on how you segment the domain, but the speed hit can be significant.

The 512GB is likely targeting people who want 1TB RAM (2x cluster).

It was the same with the M3 Ultra. Two 256GB is a better deal if you want to bother connecting them with thunderbolt.

It might be possible to use that, because you're of course effectively doubling your memory bandwidth at the cost of a slow interlink. (well, still >100 gbit/s, ie. still better than pcie bandwidth in 5090s or RTX6000 card)

yea but who needs that clutter on their desk

Reply on news.ycombinator.com