http://nvidianews.nvidia.com/Releases/NVIDIA-Launches-World-s-First-High-Speed-GPU-Interconnect-Helping-Pave-the-Way-to-Exascale-Computin-ad6.aspx
shows only gpu-gpu links (just like SLI).
I think that picture is incorrect. It shows NVlink only between the GPUS, but
the text accompanying it states
"plans to integrate a high-speed interconnect, called NVIDIA® NVLink?, into
its future GPUs, enabling GPUs and CPUs to share data five to 12 times faster
than they can today."
OTOH, it's not as if NVLink is coming soon to any AMD/Intel processor near
you. I guess I wouldn't be totally shocked if IBM did something bespoke to
make use of these ultra-high-end modules. actually, from the looks of them,
I wouldn't be surprised if IBM were physically producing the modules,
since they have decades of high-performance (cost is no factor) MCM experience.
I think the big question is this: what do we need exascale for? projects
like NVLink and MCMs seem to be premised on the idea that spending billions
to develop exascale is justified, since pushing tech, while absurdly
expensive at the time, blazes a trail so that followon products become
affordable. I think this is a pretty dubious proposition, especially because
it is anti-empirical. (ie, if it doesn't work, it's because you didn't spend
enough; if it does work, there's no reason to believe you had to spend that
much. and everyone knows that the performance will come eventually, on its
own. exascale is *not* the same sort of thing as the Manhattan project.)
regards, mark hahn.
_______________________________________________
Beowulf mailing list, [email protected] sponsored by Penguin Computing
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
