IBM's BlueGene/Q super chip grows 18th core

The mystery surrounding the number of cores in the 64-bit Power processor that will be at the heart of the 20 petaflops "Sequoia" BlueGene/Q supercomputer has been finally cleared up. Back at the SC10 supercomputing conference in November 2010, a software engineer working on the BlueGene/Q system told El Reg that the processor …

[Anonymous Coward]

what? (again)

"keep that clock speed rock solid and therefore be able to synchronize events across those 1.57 million cores"

You're saying the cpu cores are all clock-synced, over a warehouse full of racks, at GHz frequencies? Uh huh.

Dealing with clock-skew across a *single* *core* is non trivial. Who fancies a whip-round for a copy of Hennessy & Patterson for the author? Call it an investment.

And the 18th core is there to 'increase yields' and is typically disabled, make naff all sense. If you build it, you enable it. If you can't get all working, the chip goes cheaper. That makes sense. So perhaps there's something else going on like it's disabled to keep within it's power/thermal limits. Or there's some sly marketing connivance happening.

[Gary Bickford]

Pretty standard way of increasing yield

I haven't RTFA, but I suspect that there is a slower cycle time related to syncing around the torus.

Also, I think that if they promise 17 cores and build 18, they can still sell chips with one dead core as a full product. Similarly, hard drives are generally sold as having a capacity based on N sectors (or blocks) when there are actually N+ sectors, and the dead blocks are just marked out. So in that sense, it does increase yields. That's my thought, anyway.

[Anonymous Coward]

@Gary Bickford

(orig AC here) re. the cores, still doesn't make sense. They can promise 17 or 18 cores, and sell the 18ers for a bit more, but still sell the 17's.

The comparison with the HDs doesn't fit; if a HD had 17 or 18 blocks (instead of many millions) then... you see my point. So would the HD manufacturer's marketeers. Especially if a HD cost as much as one of IBM's monster chips.

Just seems odd.

[networkboy]

This isn't consumer electronics

You bin sort for yield on consumer electronics, sure. If it works sell it for more $$ and make more $$. In the HPC space reliability is king, so you take an improved yield and any that are still good can also be used as spares.

[David Halko]

Nice to see...

Nice to see a crossbar in the middle!

IBM is making POWER looks like a traditional SPARC processor!