Pentium 4: the speculation continues

Here is a picture of the retail package of the Intel Pentium 4 and i850 motherboard taken from PC Watch. Although the official release is Monday, there is no stopping people from speculating and giving their opinions. This is how the news takes on a momentum of its own and becomes a feeding frenzy. The early skinny on Willamette was that its FPU was not as powerful as the Athlon leading to this type of prognostication (thanks to NeWs dAwG for the link). Just a few days ago, Australian IT showed benchmarks since removed but reposted at Insane Hardware of the Pentium 4 turning in high scores on the SPEC benchmark. tnaw_xtennis followed up with their take on what was behind the higher than expected scores (hint: FSB). Today, TR reader holmanator points to AMDZone's front page where you can find more P4 vs. Athlon benchmarks. The full post can be found at JC's PC News'n'Links and Silicon Investor. The Silicon Investor post includes benchmarks of the P4 1.6 GHz and a description of the test systems. The scores found on both websites originally come from c't magazine and seem to confirm the benchmarks from Australian IT. To what can we attribute these results? Is it the better than expected FPU or the front side bus that is the X factor? The answer, as Ryu Connor has speculated, may have something to do with memory bandwidth.

When Rambus and the Intel i820 Camino chipset came out last year, people were quick to point out that Rambus DRAM was too expensive (which it was) and the benchmarks showed no great performance leap over PC-100 SDRAM on a BX board. But Rambus' memory bandwidth advantage is very real. However, it is crippled by the higher memory latency. No less an entity than Anand took an unusual amount of criticism for his controversial two-part performance article on Rambus (part one and part two) earlier this year. Part two of the analysis contains graphs and benchmarks. You don't hear many hardware websites say this but using Occam's razor, it may just be that the i820 and even i840 motherboards are hamstrung by the 133 MHz front side bus which is not taking advantage of the amount of memory bandwidth available to it. has got a Pentium 4 reviewer's box (how? don't ask, don't tell) and they have run some benchmarks including UT (which may be broken ~ see "The Damage Report" on why). They make some fairly bold statements about the Pentium 4 and the Intel i850 chipset. You can read their conclusions here. utilizes the work of Tim Wilkens.

In a reply to a post below, Tim Wilkens contends that SPEC is not using optimized math libraries. To see what he means, take a look at a side by side comparison of unoptimized and optimized benchmarks. The unoptimized left and optimized right graphs show the result of DGEMM which is a basic linear algebra routine of multiplying two matrices A and B and adding the result to a third matrix C: A X B + C. The left graph casts the Pentium 4 in a more favorable light but in a real world situation, who isn't going to use optimized math libraries? The right graph was compiled using the free, prebuilt ATLAS (math) libraries. ATLAS optimizes itself automatically for each individual platform independently. You can download the ATLAS libraries here. Although the left graph purports to show the superiority of the Pentium 4 (basically SPEC as it is), the right graph (essentially SPEC with independent optimizations for each platform) is more representative of a real world situation and clearly illustrates the strength of the Athlon FPU. In addition to showing the limitations of the Pentium 4 processor's FPU, it also inadvertently reveals the synthetic nature of benchmarks like SPEC. Tim Wilkens is the author of a number of scientific benchmarks and has been submitting quality work over at JC's. Click to find more information about QMC, Primordia, and Linpack.

Speaking of benchmarks, I saw on JC's PC News'n'Links that ZDNet has updated their Business Winstone and Content Creation Winstone benchmarks to 2001.

Tip: You can use the A/Z keys to walk threads.
View options

This discussion is now closed.