4 ms·
Intel attempted a GPU that was effectively a cluster of x86 cores with a massive process shrink. https://en.wikipedia.org/wiki/Larrabee_(microarchitecture)#Dif
by techrat 6y ago
Intel attempted a GPU that was effectively a cluster of x86 cores with a massive process shrink.
https://en.wikipedia.org/wiki/Larrabee_(microarchitecture)#Differences_with_CPUs https://en.wikipedia.org/wiki/Larrabee_(microarchitecture)#D...
Not quite as old as 486s, but damn close: The cores were all based on the P54C Pentium, in order execution only type processing cores.
I remember hearing about an Intel Dedicated GPU that was supposed to be all 486 cores, but my ability to Google it seems to be failing me so I suspect it was probably my conflating an IOE type cluster of cores of the Larrabee with the 486.
It appears that the Larrabee came in variants of 10 (at 65nm) 40 (at 45nm), 48 (at 32nm) cores on the dGPU.
https://www.anandtech.com/show/2580/6 https://www.anandtech.com/show/2580/6
In the end, Intel expected people to be able to run x86 code on these dGPUs...
- theandrewbailey 6y agoLinus (Tech Tips) got his hands on a prototype Larrabee card, but wasn't able to get it running. https://www.youtube.com/watch?v=um-1fAVU1OQ https://www.youtube.com/watch?v=um-1fAVU1OQ
- temac 6y agoIt even kind of shipped in the form of Xeon Phi, without graphics though, including in the first version a precursor of AVX-512 (and in next versions subsets of AVX-512 that are not exactly the same as the ones most in use today... it's a mess...)
- fulafel 6y agoThe chip was not really like a cluster, the memory semantics were fully coherent SSI like we're used to on x86 multicore systems. (Possibly you meant cluster in a chip design sense, not in the distributed computing sense? In Larrabee/KNL context, clusters usualy meant racks of machines HPC style...).