Petz,
The reason AMD is not pursuing a 166x2 FSB with PC2700 for Athlon XP's is because it only results in 5% performance improvement at most. Various overclocking sites have proved this but you and Joe insist on continuing to overestimate the advantage of doubling memory bandwidth.
A potential comparison would be to compare KT-133A vs. KT-266A. Sintle to dual channel should be slightly faster, since getting, for example, 16 bytes would take 1/2 of the memory clock cycle less if the transition was from single SDR to dual SDR, instead of single SDR to single DDR.
Dual channels would also require some incrase in latency.
Why is that?
The clawhammer has double the cache of the XP.
Clawhammer will not be compteting with Athlon. All the benchmarks will be Clawhammer with 512k L2, single DDR vs. Northwood 512K L2, dual DDR. Later, it will be vs. Prescot with 1 MB L2, dual DDR. So AMD will have to bring Sledgehammer to desktop, not to be blow away by Prescott. Why not do it a couple of quarters earlier?
Because AMD likes to snatch defeat from jaws of victory. Hammer could have shown up with a roar, instead, it will arrive with a whimper.
Alternatively, AMD could make a Clawhammer socket version of Sledgehammer -- fewer pins but with 1M L2.
Didn't you notice that all the benchmarks are becoming a form of stream benchmark? First of all, you have the real stream. Then you have SpecFP, the stream-like media conversion, encoding / decoding benchmarks. Then, you have what used to be application tests, but these have been turned into stream bench, by running a stream like application in the background. Heavy multitasking apps stress L2 size (and wayness) + memory bandwidth, since contents get evicted a lot under heavy multitasking.
Now, with all the benchmarks rigged this way, and AMD fully aware of the fact how the benchmarks have been rigged, AMD knows where the trap is, but they are walking right into this trap.
Joe |
