To all, based on Anand's K6-2-400 benchmarks, K6-3-400 will be at least 5% better than I previously anticipated.
In my 'old' simulation, I had assumed that the improvements to the CPU core would only be 5%, which would result in only a 1.5% improvement in Winbench 98 scores. (You would see the full 5% on an in-the-cache CPU benchmark like Norton SI.)
In actuality, going by either Anands Winstone 98 or Winstone 99 benchmark, the core CPU improvements ('CXT' on the K6-2, also to be in every K6-3) are an increase of 16% in CPU speed, combined with a 10% reduction in L2 cache and main memory access, mostly due to write combining. With these changes, and an empirical adjustment to the I/O wait time's to account for Anand's other hardware, my math model predicts all the benchmark numbers in Anand's latest report within 0.2 Winbench points. My model predicts 25.16 for the K6-2-400 NON-CXT winbench number and 26.74 for the K6-2-400 CXT-core winbench number, while the actual numbers are 25.1 and 26.8, respectively. So, if anything, applying the 16% CPU and 10% memory access factors underestimates the improvement provided by the CXT core of the new K6-2-400 and (future) K6-3.
So, how good will the K6-3-400 (if there ever is one -- the K6-3 may start at 450 MHz) be on Winbench 98? The following numbers are about 20% higher than Anand's because I am assuming tomshardware.com's (CPU shootout) ultra-fast peripherals, like the IBM UW-SCSI 10K RPM hard disk, 128M RAM etc. Tom's "CPU-shootout" report has more overclocked Pentium II and Celeron numbers for comparison.
CPU...........Actual Tom's..Prediction....Act.Anand..Prediction
PII-400................30.8..........................25.6
PII-450................31.4
PII-504 (112x4.5).33.5
Katmai 500.........................35.8 (guess +7% vs PII-504)
K6-2-350.............27.2........27.3.............24.0........24.0
K6-2-400.............28.3........28.5.............25.1........25.2
K6-2-CXT400....................30.5.............26.8........26.7
K6-3-400............................34.8............................30.1
K6-3-450............................36.3............................31.2
Clearly the K6-3-400 will beat any overclocked or 500 MHz Pentium and come close to a Katmai 500. A K6-3-450 will beat even that.
(PS, yes, I DID account for only 256K L2 on K6-3. I assumed only 10% of L2 cache misses would be in L3. And if you don't believe that a 7% increase in Winbench requires 10-20% improvements in CPU speed and memory access, look at how the results from 350 to 504 MHz are all bunched together at tomshardware.com
Petz |
