Re:Yes, but... (also long) (Score 2)

I'd love to see Intel and AMD quit spending all their money (or the vast majority of it) on speed improvements, and dump a huge amount of it into redesigning the PC architecture. Honesly, I think there is a huge amount of $$ to be made in the chipset integration market. Intel has done this to a certain extent, and now owns probably 60% of the primary MB chipset market. However, nobody is doing anything interesting.

AMD has made a step in the right direction with the adoption of the EV6 bus, but Intel is still running the GTL+ bus, which sucks. I'm talking about as a bus architecture, not comparing MHZ of the bus.

I'd love it if I could buy a 600Mhz PC with an advanced memory and I/O subsystem (maybe like that on the SGI O2 or VisualWorkstion series). The key here is that the push has to come from Intel or AMD. They're the only ones who can design the CPUs to take advantage of such a design. SGI's problems (besides internal ones to the company (like no clue about how to sell into the NT market)) were that the CPU was still a "typical" CPU, and the custom chipset had to work around its limitations.

I hate to plug it, but Microsoft's X-Box seems to have at least some of the specs that you're looking for, although they seem to have limited it so that they can sell it as a $400 console. If I read the specs correctly, they're planning on 200Mhz DDR on a unified memory bus. Unfortunately, I think that this technology is probably not going to go too much further than nVidia and our favorite software giant. But hey, it's a PIII architecture that's started to catch up with the SGI 02.

Make sure that they can sustain 10 GByte/s throughput from main memory to the caches to the CPU.

As for 10GB/s on the memory bus, you're looking at:

• 128-bit memory bus @ 640Mhz SDR or 320Mhz DDR. As far as I know, nobody is getting good yields on 200Mhz DDR parts at this point.
• 256-bit memory bus @ 320Mhz SDR or 160Mhz DDR. IIRC, most designers aren't using 256-bit data paths because of signal noise.
• 2-channels of RAMBUS would have to run at 1250Mhz -- that's PC2500 RDRAM.

So, I'm afraid that 10GB/s *anywhere* is a bit too optimistic.

Look at methods for allowing non-core devices to tap directly into the memory system. AGP is sorta-kinda there, but not even close to some of the state-of-the-art stuff I've seen out of SGI, HP, and Sun (particularly the first two).

And one of the problems with updating the PC design is cost: while the chip manufacturers might be willing to fork over the dough for a higher-speed chipset, only people who are building new systems with no legacy peripherals will buy an entirely new chipset. And even small increases in cost seem to cause a lot of technologies to languish - how many 66Mhz or 64-bit PCI cards or slots do you own?