The critical question in 2002 for AMD is “how much bandwidth can the Athlon (as opposed to any particular platform for it) handle?”
AMD apparently plans to stick with a 133/266MHz FSB throughout 2002. Are they doing that because faster isn’t readily available/reliable yet, or because their processors can’t handle faster?
If the XP or Throughbred or Barton can’t effectively (as opposed to officially) handle much more than a 133/266 FSB, Intel is likely to gain a significant advantage in some performance areas the second half of the year with either PC1066 RDRAM or dual-channel DDR.
The only way you can see whether or not the CPU can handle 2.7GB/sec of bandwidth is to run the whole operation at 166Mhz. Not FSB at 133Mhz, memory at 166Mhz. Unfortunately, that’s just what the reviews I’ve seen did.
By running memory asychronously, you introduce all sorts of potential latency problems. The only way you can prove whether or not that’s actually happening is to run the whole shebang synchronously, and see the differences.
Even more basically (presuming your benchmark truly tests CPU bandwidth), by running memory faster than FSB, you are imposing a bottleneck on the system. It’s like attaching a fuel line that can only pump 20 gallons an hour when the fuel tank can provide 25. The engine will get 20 gallons an hour, but that doesn’t “prove” the engine can only handle that much. You have to change the fuel line and see what it does with 25.
You cannot prove a limitation by testing it within the limitation. You prove (or disprove) a limitation by testing beyond that “limitation” and seeing if it’s really there or not.
I swear, some are treating the 133/266Mhz FSB like it’s as sancrosanct as the speed of light. Maybe it is to some, but not to overclockers.
Yes, there’s other factors involved.
Yes, even at 166Mhz all the way around, you’re only going to see a few more percentage points improvement at best for memory bandwidth-intensive operations, but all that proves is that there’s too much of a memory fetish out there. Faster memory simply doesn’t scale much compared to other improvements.
Yes, I’m sure there is some point where the Athlon says “No mas,” and it could be somewhere between 2.1-2.7Gb per second. We don’t really know, and these tests don’t prove it.
The point to all this is that there’s a much more important issue here than how well the KT333 does, or can be made to do. If the bottleneck is the fuel pump, that’s relatively easy to upgrade. If the Athlon engine is the bottleneck, AMD has a much more serious potential problem.
Via plans to introduce dual-bank DDR for both PIVs and Athlons. The PIV version is the the P4X600; the Athlon version is the KT400. If the first does the PIV some real good, but the second doesn’t. AMD is going to have a rough time in certain markets until Clawhammer.
P.S. What may be the real reason why we didn’t see 166Mhz all around is that the available KT333 motherboards may well have not had that option built in. I sort of doubt Via reference boards have it, and according to its manual, the Gigabyte GA-7VRXP only goes to 161Mhz.
Under those circumstances, you need to force a higher FSB through programs like CPUFSB.
That’s what needs to be looked at, and what I’m planning to look at in the near future.