Slashdot Mirror
Sun's UltraSPARC III Processor Shipping
jzuska writes: "Sun Announced it's UltraSPARC III today. The link is here. 29 million transistors and 900MHZ, 9.6 Gigabyte/second processing bandwidth woohoo!!!" And this is an announcement of "volume shipments," not *coughcough* EventualWare. So lessee ... for boxes that you might reasonable put on your desk right now, there are how many architecture choices? (And how many of them run Windows?)
On a semi-related note, Sulka writes: "Motorola just introduced a new G4, the MPC7410. While this is a processor meant for embedded systems, it's essentially the same CPU as the G4 in Macs. What's significant is the price -- the 500 MHz version carries a $195 price tag. This is much cheaper than the Intel and AMD high-end offerings. I wonder how much the G4e is going to cost." Sounds like a cool basis for (awfully) high-end set-top devices at that price, but imagine what that will cost 12 months from now! Yoiks.
SPECfp2000 results are available for the UltraSparc-III 900 MHz. It scored a 482. Pretty damn quick, especially when you consider that its score is more than 50% higher than the Pentium-III 933 Mhz, which got a 305.
Wrong comparison. First off, the proper score to be using is SPECfp_base, not SPECfp_peak. In case you didn't know, while recently some pretty ridiculous SPEC optimizations have been sneaking into the compilers used for SPEC_base, they are at least optimizations included in the standard compilers; SPEC_peak numbers include optimizations which would break any other code, and thus are not considered widely applicable.
Second, you clearly ought to be comparing the US3-900 to the P3-1000. Yes, the GHz P3 was unavailable for 6 months after it was supposedly "launched", but it is available now, or at least as available as the US3-900. Plus, the P3-933 numbers you quoted were hobbled by the horrible i820 chipset (yes, the i840 is only used for high-end workstations, but that's the market we're talking about here, right), and were obtained using an older version of Intel's Fortran compilers. Now, many people have complained that Intel's new compilers are so good as to call into question the usefulness of the SPEC_base benchmarks, since successive versions of the compiler have shown remarkable improvement in SPEC scores on otherwise identical computers. Still, they meet SPEC's rules for base scores, and everyone optimizes their compilers for SPEC, and most importantly the SPEC tests seek to benchmark not CPUs but entire platforms, and the compiler is an extremely important part of any platform.
Thus, the numbers you should have quoted are:
SPECfp_base:
US-III@900: 427
P-III@1000: 327
Looks a lot less impressive, doesn't it. Especially when you consider the fact that any chip with an ISA less than 20 years old ought to beat the pants off x87 in SPECfp, due to x87's crippling 8-register stack-based FPU implementation. Once the x86 chips finally phase out x87 in favor of SSE2 (coming with Intel's P4 and later with AMD's K8 "Hammer" family), they will finally have a decent platform for double-precision fp, and their SPEC_fp scores should rise accordingly.
In any case, considering the US-3 is destined primarily for the server market, the more important SPEC benchmark is not SPECfp but rather SPECint. Let's check those scores, shall we...
SPECint_base:
US-III@900: 438
P-III@1000: 438
Ouch.
Of course, the real strength of Sun's UltraSparc line is its tremendous scalability. Yes, you're overpaying for a 1-way Sun system, or even 2- or 4-way, but what you're paying for is the headroom to later buy a 64-way machine without having to completely switch your architecture. Fine. Considering the poor scalability of x86 and Compaq's awful support of the Alpha platform, it's completely understandable that IT departments continue to overpay for Sun boxes. Intel's flubbing of IA-64 thus far has given Sun a 3 year reprieve, and it'll be another year and a half before a real IA-64 CPU (McKinley) shows up on the scene.
Still, don't try arguing that Sun can compete with anything (except HP, which desperately needs to release a new processor) on straight price/performance. The US-3 closes the gap quite a bit from the extraordinarily outdated US-2, but not all the way. Sun's done a decent job squeezing performance out of an in-order design, but when Intel releases the SPEC scores for the P4 5 weeks from now they're going to make it (and indeed everything else) look extremely bad.