Saturday, December 22, 2007

Intel GPU/CPU combo in 2009?


Really integrated graphics may be coming in 2009

Penryn is almost here; Nehalem is due next year.

But Intel may be planning to follow them up with new chips integrating a GPU with the CPU in 2009 - at least according to the Reg.

Bloomfield will apparently feature a 1366 pin socket (ouch) and feature "QuickPath" - Intel's answer to AMD's "HyperTransport" along with up to three channels of DDR3 memory. The new core apparently will have 8MB of L2 and use QuickPath to connect to a "Tylersburg" northbridge that will provide PCIe 2.0 lanes.

Strangely enough, less than six months later there will apparently be a "Lynnfield" going to a 1160 pin socket with four cores and hyperthreading, 8MB of L2, dual DDR3 channels, using PCIe as a chip to chip bus. Frankly, this sounds weird to me. Why on earth would Intel go to a 1366 pin chip, to go to a 1160 pin chip less than six months later?

The only way that this would make sense is if one of the chips was destined for the low end integrated market.

Let's peek into our crystal ball... to reduce costs, you have to simplify designs. What if Lynnfield integerated the GPU? With PCIe integrated onto the chip, it is not farfetched to consider that Intel may also consider integrating a couple of SATA channels and some USB2.0 channels... then you could have a nice little quad core "SOC" (System on a Chip) allowing very low part count integrated low end designs.

Such an SOC really would only be suitable for low end machines as the GPU would have to share memory bandwidth with the four core / eight thread CPU portion, so even a hypothetical dual channel 2000MHz DDR3 system with close to 32GB/sec bandwidth would be hard put to exceed the performance of an 8600GT - and it would starve the processor for memory bandwidth at that.

To make matters even more confusing, the referenced roadmap also shows a mainstream "Clarksfield" part using a 989 pin rPGA socket. UGH.

PLEASE PEOPLE: We don't need so many sockets!

No comments: