Slashdot Mirror
Sun's Zippy New Chips
Mark the Revelator writes: "Reuters has a story about Sun unveiling it's latest and greatest UltraSparcIII chips. The new chips are being made by TI and are the first UltraSparcs to use copper instead of aluminum for transistor connections. Although they're supposed to compete with Intel's Itanium chips, they only run at 900MHz ... for now."
Won't somebody PLEASE think of the child processes?!?!
--
"Outlook not so good." That magic 8-ball knows everything! I'll ask about Exchange Server next.
Hmm, Not really.
I mean, yeah, they're totally different. And they're faster clock-per-clock (with added benefit to FP stuff).
But a 1.4GHz Athlon blows away a 7-800MHz UltraII for most kinds of computation. A 1 GHz Athlon seems to be about (42, 29) on the (retired) SPECint95/SPECfp95. A 450Mhz Ultra-II (not Ultra-IIi, I'm looking at results for an SPARCstation Ultra-60) gets about (20, 27). That's a bit faster int clock-per-clock, and a lot faster FP. Note that for practical stuff (databases, web, whatever) int is more important. Of course benchmarks are hard to interpret, but this gives you an idea. All the SPEC benchmarks are available at www.specbench.org. Of course there are no Ultra-III results, but I'm guessing it's not going to be 2x as fast as the best x86s (at least I'll wait to see the results before I believe it).
You use a Sun because you want an architecture that will scale smoothly up to 64-way (I *guarantee* that will be faster than any single x86 machine).
Actually if you want to both go fast at the low end and scale well, you can buy an RS/6000 -- IBMs Power3 and Power4 chips are absurdly fast and scale very well (and actually focus on memory bandwidth for database performance). But a bottom-of-the-line Sun is a lot cheaper than the cheapest RS/6000.
Full disclosure: I work for IBM (in software) and I've seen a good bit of internal stuff about IBM chips, esp. the upcoming Power4. Most of that information has now been published in MicroProcessor Review and is now publicly available, I think you'll find it if you poke around...
(even more amusing full disclosure: I'm a huge fan of old Sun stuff, their machines are beautifully engineered. i use a couple old 32bit sparcs for all kinds of things)
Just because the MHz on the Sun equipment (900MHz) is lower than the current Pentium (1.5MHz), don't be fooled into thinking the Intel hardware is better. What matters after all, is throughput and pumping that data. Check your specs!
Check this 4 CPU Intel vs the 1 CPU Sun considering plain speed...
CINT2000: Intel Corporation Intel D850GB motherboard(1.5 GHz, Pentium 4 processor) - 536 524
CFP2000: Intel Corporation Intel D850GB motherboard(1.5 GHz, Pentium 4 processor) - 558 549
CINT2000: Sun Microsystems Sun Blade 1000 Model 1900 - 467 438
CFP2000: Sun Microsystems Sun Blade 1000 Model 1900 - 482 427
CINT2000: Advanced Micro Devices Tyan Thunder K7 Motherboard, 1.2GHz Athlon MP Processor - 522 495
CFP2000: Advanced Micro Devices Tyan Thunder K7 Motherboard, 1.2GHz Athlon MP Processor - 481 433
Throughput on the Sun with 2 CPU, but strangely enough, none for any Intel hardware. Throw a 2 CPU AMD in there, though...
CINT2000 rate: Sun Microsystems Sun Blade 1000 Model 2900 - 10.7 9.97
CFP2000 rate: Sun Microsystems Sun Blade 1000 Model 2900 - 10.2 9.09
CINT2000 rate: Advanced Micro Devic Tyan Thunder K7 Motherboard, 1.2GHz 2CPU - 10.8 11.1
CFP2000 rate: Advanced Micro Devic Tyan Thunder K7 Motherboard, 1.2GHz 2CPU - 8.30 9.14
"Beware of he who would deny you access to information, for in his heart, he dreams himself your master."
They make all of Sun's UltraSparc chips, and also manufacture other, more esoteric things - like dual core chips (DSP and ARM, known as OMAP).
All in all, TI is much, much more than calculators.
Part of Sun's success is how well they address the bus/throughput issue, as opposed to 'other' computer architectures. And that's why JUST comparing MHz is like comparing apples and oranges.
Or perhaps a better anology is comparing a Formula 1 Racing car stuck in down-town NYC Traffic, versus a 6 cylinder Honda Accord on flat, wide-open highway in Montanta, during the daytime when the weather is perfect.
healyourchurchwebsite.com - WWJB?
Where the real advantages come in is with things like memory architectures (eg, memory interleaving) and bus speeds (where the system bandwidth is more than an x86 solution) which is relevant in databases. Added to that, you can scale these up much more (the E6800 can have 24 900MHz CPU's, for instance; Fujitsu have recently released a 128 CPU system based on their USII clone at 500+MHz).
If you want a measure of raw CPU performance, check www.spec.org; currently, the fastest single CPU systems are Intel P4's (although some alphas come damn close). The Sun 280R doesn't come close to that, although it is faster than its clock speed would suggest...
How in a civilised society can we sit back and let this apocolypse happen? I say its time to end this now. Boycott processors. Save the instructions
Those terms don't apply well with modern processors. Pentiums class processors are primarly CISC with some RISC features/ideas (not many though). The Sparc family has been RISC with a lot of complexity thus making them be more CISC than say the Alpha. That has historically been why their clockspeed is lower than alpha, but still performs about the same for general purpose computing.
The Itanium is a branch off of a different tree, Very Long Instruction Word, which is a branch off of RISC. VLIW let's a compiler pack multiple commands to multiple execution units into a single long word. The idea is to use very RISCy commands to keep a superscalar set of execution units more fully utilized. Great idea, if your compiler can do it.
Just look at the requirements to run the various Windows OSs. When Windows 95 came out, the bare minimum to run it was a 386DX at 33mhz, 4MB RAM, and a 100MB hard drive. Windows ME requires, at a minimum, a 150mhz Pentium, 32MB of RAM, and 480MB of hard drive space. The RAM requirements have quadrupled, the hard drive space has gone up by a factor of five, and CPU power has gone up by somewhere around a factor of 10. (I know that there is some disagreement about what the actual minimums are, but I believe these to be in the ballpark and they illustrate my point.)
So, if you want to find out what the CPU is capable of, dump the OS, write an application that taxes the CPU, and run it on each. (No, you do not need an OS to run a program.) Until you do that, you're just tossing around meaningless numbers.
Really? By what measure? CISC is generally much more efficient with respect to code size, an important consideration in embedded systems.
I'll assume you were talking about the performance domain. Be careful with your categorizations. There are no "pure" RISC or CISC designs anymore. O-O-O superscalar architectures have pretty much killed any simplicity in so-called RISC designs. Now it's true that uniform instructions make O-O-O much easier. But vector processing and multimedia operations don't really qualify as RISC in the classic sense.
Sun has made some obvious mistakes in the past: fixed-size register windows and delay slots come to mind. Like Intel/HP they have in the past made the mistake of thinking that the compiler can do more than it really can (at least at this point). Parallelism is hard enough to extract at run time. It's much more difficult at compile time. Some of this has to do with maintaining the separate compilation model and speed/memory complexity issues (many compiler algorithms are NP-complete).
And of course, all CPU vendors except Intel/HP have made the mistake of having an inadequate number of general-purpose registers. Ironic, eh? :)
That's not to say the compiler can't do more. It can do a lot. Unfortunately, CPU vendors have not provided the necessary hardware to make this possible. In the future you will see a style much more similar to IA-64: the hardware and compiler conspiring together to extract parallelism, save power, etc.
Here's something to think about: the original intent of RISC was to allow simpler pipeline stages and higher clock speeds. So why does a CPU implementing a CISC-ish ISA have a 50% higher clock rate than a RISC-ish ISA implementation? Deeper pipeline, sure. But don't let labels fool you. There is much more going on in the architecture world.
I do agree with you on the scalability issues of SPARC systems. That's their bread and butter.
As you implied, SMP performance is extremely important to people who buy Sun.
In this case, you wouldn't care much how an individual processor performed; you are most concerned with the performace of, say, a 32-way system and it's ability to quickly shuttle data between processors, memory, and disk.
Our beloved Athlon only scales to 2-way, and it's SMP architecture is now being entirely redesigned with the NUMA hypertransport.
Sun probably suffers in raw MHz and SPEC scores because they put so much effort into the SMP aspects.
And, of course, Sun outsells some (arguably) better technology (Power, Alpha) because they are much more open and their service organization is superior.