As a general rule, I don't see CPU applications as being DRAM bandwidth limited. Good CPU's are more like a sports car than a train. They need to be agile, and switch directions quickly. Tasks which require lots of DRAM bandwidth are better run in special purpose processors, like graphics processors or network processors.
Okay, so based on these criteria (which given the architecture of the P4, I think Intel would probably disagree with to an extent), how do you evaluate the P4 overall? What changes do you think would be required to keep the P4 competitive once the SSE2 optimized K8 / Hammer processors are on the scene???
As it's only fair, I'll go first and allow you to perform microsurgery as required...
I don't see the K8 series succeeding as a high-end server processor. IA-64 will be too well entrenched at that point. I do, however, see th K8 series as a real heavy hitter on the PC desktop. To keep the P4 competitive, I believe that Intel needs to consider the following upgrades to the P4 core: 1) Reduce the penalty of a branch mis-predict. The Northwood is rumored to have twice the on-die L2 cache of the Willamette (512 vs. 256). This probably won't be enough, though it's an economical first step at 0.13 micron. It also creates the potential for a "Celeron" version of the P4 with half the on-die cache. 2) Beef up the x87 and integer units. The double-pumped ALU's are novel, but they don't seem to be providing a knock-out punch at present. Maybe it's software at this point... |
