64-bit battle lines form

AMD IS NOW out with Opteron, IBM has just updated the rest of their RISC line with Power4+ and Intel is closing on the Madison Itanium2 launch. How do the battlelines look for the remainder of this year, and early next year?

During 2003, barring any yield issues, Opteron may reach 2.4 GHz - further tuning of the initial 0.13-micron part should easily yield that speed, since it has longer pipelines than the 0.13 um Barton AthlonXP which is already well abover 2.2 GHz actual clock speed in the 3200+ PR rating. Further major speed-up will require the move to 0.09 um (µ, microns) in 2004.

Power4+ is now at 1.7 GHz, holding the current 64-bit speed record, at least until Madison comes out, possibly dethroning it in quite a few benchmarks. The next major speed boost for IBM should come around yearend with the new Power5 - speeds around 1.8 GHz, but most importantly quite a bit more performance per clock with 2-way multithreading, more registers for better FP and improved cache & memory paths - expect Power5 off-chip 36 MB L3 cache to have higher bandwidth than some competing CPU's on-chip caches - no, I did not mention UltraSPARC III by name here. By the end of 2004, IBM should be ready to move to 0.09 um process with Power5+ with up to 2.5 GHz speeds.

With Madison firmly on track for announcement in late summer, Intel will have another top performance contender: the 1.5 GHz beast with monstruous 6 MB on-chip cache, biggest-ever on any general-purpose CPU. Expect this thing to be in the first or second place in many benchmarks. The low-voltage, 1 GHz 1.5 MB cache Deerfield will branch out to the blade arena, while the next high-power flagbearer, early next year, is expected to be Madison 9M, with clock closing onto 1.8 GHz and with, as the name says, even more humungous 9 MB on-chip cache. You might see a major FSB boost on this Madison rev, either 667 or 800 MHz net throughput, probably coupled first with a non-Intel core logic chipset (HP and IBM).

2004 will also see the very last Alpha, unless someone outside HP takes over the poor thing. The 0.13-micron EV79, the last incarnation of the seven year old EV6 core, will still hover around the performance top, sadly hinting what could have been achieved with Alpha if it was supported and developed wholeheartedly. The 1.6 GHz mark should be easily achieved (and well exceeded in fact) but it remains to be seen whether it will be deliberately throttled down in the shipping systems as not to threaten the new "crown jewel" of EPIC proportions...

SPARC? Well, US-IV got to be there or many heads would roll... or maybe not, excuses can always be cooked up. Anyway, being basically a dual-core update of US-III, US-IV is not expected to offer much more in single-CPU performance - what means it will be right down at the bottom of the garbage dump, speed-wise. Even if it reaches 1.5 GHz by yearend, US-IV per-core SPEC, Stream, Fluent etc performance will have trouble even matching the current McKinley 1 GHz Itanium2, not to mention its up-to-date competitors at that time.

Here's a little "teaser" table for the CPU-hungry - again, these are just unofficial estimates by independent industry sources, haha... µ

64 bit CPU shootout - first block late 2003, second early 2004
CPU	Max Clock	Cache on chip	Rpeak GFLOPs	Rmax GFLOPs	SPECpeak int2000	SPECpeak fp2000
Itanium 2	1.5GHz	6MB	6	5.5	1200	1900
Power4+	1.7GHz	1.5MB	6.8	4	1150	1700
Alpha EV7	1.15GHz	1.5MB	2.3	1.8	900	1500
Opteron	2GHz	1MB	4	3	1250	1350
US-IV	1.4GHz	1MB?	2.8	2	900	1300
Itanium 2	1.67GHz	9MB	6.7	6	1350	2200
Power 5	1.8GHz	2MB	6.8	5.7	1500	2500
Alpha EV7	1.6GHz	2MB	3.2	2.5	1300	2100
Opteron	2.6GHz	1MB	5.2	3.8	1600	1700
US-IV	1.5GHz	1MB?	3	2.2	1000	1400