[Beowulf] 64-core processor...
hahn at mcmaster.ca
Tue Aug 21 12:09:57 PDT 2007
I googled on "+raw +multicore +mesh", and found, among other things,
a book chapter that went into some detail on the "Raw" prototype:
(access might be limited to campuses that have agreements w/springer.)
I'm guessing that Tilera is a lot like Raw - couple generations more
agressive fab, probably cleaned up design, etc. here's the Tilera guy
talking about tiled multicore streaming architectures:
> From the picture in Ars Technica, there are four memory controllers for 64
> processors. There is no floating point, but there would certainly be room for
> it in a 65 nm version.
I'm pretty sure Raw had FP, and guess Tilera does as well (not DP though.)
> It seems to me that the key difficulty for building larger clusters out of
> these things
> is the imbalance between computing and memory bandwidth.
well, I think the premise of most massively multicore is a dataflow/streaming
design for programs, not a memory-based model. more recv-compute-send
rather than load-compute-store.
one very frustrating thing about Tilera is the total vapidity of their
published docs. for instance, inter-tile latency, which is pretty critical
in evaluating how much cache misses will hurt. heck, I'm not even clear
about whether the caches implement coherency, or whether they're sw-managed.
> and routing, but not so fine for general purpose computing. 16 cores per DDR
> (doesn't say if they are 64 or 128 bits wide, but I would guess 64 based on
> reasonable size
> package pin counts) seems way too skinny a pipe to memory to be reasonable.
four DDR interfaces, which I also would guess would be a dimm wide (64b),
or something like 40 GB/s. that really doesn't seem too terribly shabby ;)
I would be most interested in a detailed comparison of your chip, Tilera's
and just for amusement, Sun's new one ;)
regards, mark hahn.
More information about the Beowulf