Intel has recently released its new line of Pentium III Xeon CPUs, based on the .18 micron process. We take a look at its performance, utilizing a dual CPU configuration on an Intel i840 plaform with 256 MB of Rambus memory as a testbed. Here are all the details of our findings.
I would like to see a dual PIII 667 setup, with similar hardware run against a dual xeon setup. They are virtually the same cpu now... not like the PII 400 against a PII 400 Xeon.
both can run on a i840 motherboard, both only have 256k of full speed cache, both can use Rambus, both are running 133 mhz bus, both can run 4x/pro AGP, and both use the .18 die Coppermine core... so why is one so much more exepnsive?
damn does that hurt
Give me fuel, give me fire, give me that which I desire
Umm... You know, I'm pretty much not convinced by this benchmark. All in all, it looks more like a paid commercial, than like any professional (or even amateur) benchmark. Among the many problems:
A) It uses just _one_ benchmarking application. Sandra. Sorry, this may be a popular benchmark, but it's also known to have its share of problems. Mostly due to the fact that it's a purely synthetic micro-benchmark, which bears NO resemblance with what happens in real life applications.
Briefly: I'd like to also see something like Sysmark, specviewperf, maybe even a Volanomark for those of us who use Java. You get the idea... something that at least tries to reflect real life performance.
B) It looks to me like a contest with only one contestant. OK, it's good, fast, etc, but compared to WHAT? How fast would a similar setup with dual non-Xeon Coppermines run? How fast would a dual _Celeron_ run in a similar setup?
C) While singing praise to RDRAM (again!) may be fun, I'm not convinced for the same reason I mentioned above. Fast compared to WHAT? And how do we know the high figure isn't just Sandra artificially scaling across CPU's and their caches? How about a comparison with the same dual-Xeon system running off SDRAM? Mind you, we don't even have a PC133 capable chipset yet, and Intel's MTH is known to be slow like heck. But it would be to at least have _something_, no matter how flawed, to compare the results to.
D) HOW and WHERE did that conclusion about RDRAM's low latency get in? Does Sandra show memory latency anywhere? I don't think so. Sounds to me like something squeezed there just for the hell of it, not like a conclusion of the benchmark, as it was presented.
E) Since the Xeon is presented as a solutions for high end servers and workstations... How's about trying the same thing with 1 GB memory in the machine? Or would it hurt someone's interests to show how badly RDRAM performs when you add more of it?
All things considered, dunno, the whole article looks to me like "oh, look, I ran the race _alone_ and came in the first place. I must be the fastest man on Earth." Combined with yet another singing praise to Brave Sir Intel and Brave Sir Rambus, defenders of all that's sacred. With all due respect, did Sander buy shares in Rambus, or what?
Moraelin -- the proud member of the Idiots' Guild
This doesn't make much sense to me really. Intel's distinguishing feature for the Pentium Pro/Xeon has always been the large, full speed cache. When the Coppermine's dropped the PIII's cache to 256k full speed I figured intel had finally got everything worked out. Their budget CPU would have 128k, middle end would have 256k and high end would have 512k or more. Is Intel really going to only make 256k Xeons? Because in many server situations it would actually create a bottleneck to go from say a 550mhz CPU with 2mb cache to a 866mhz with 256k cache. Servers are much more dependant on I/O and Memory performance than raw mhz these days. Our main Novell server is only running at 10%-20% cpu load max but the disk, memory and PCI bus are all maxing out. I know producing large cache chips is quite expensive (the reason PPro's were so expensive, if the cache was bad, the whole chip had to be tossed). So what's Intel planning here?
Amen, Lyam, when he mentioned Solaris, I thought exactly of seeing it benchmarked against a Sun UltraSparc machine. I mean, if he can set up a benchmark under Solaris x86, it should be trivial to re-compile it for the sparc. (And I wouldn't imagine he was recommending it for that configuration, without actually testing it first )
And you're right about the price. The price of getting a decent amount of RDRAM for a server puts you about in the price range of a dual CPU Sun server. And those ARE known to have mondo memory bandwidth, plus a good 2MB of L2 cache per CPU. (Hint: no RDRAM used for that bandwidth, just a very wide bus. Exactly the opposite approach.)
Moraelin -- the proud member of the Idiots' Guild
I am also somewhat disgruntled with reviews such as this Dual Xeon Review. The Iwill board is one that has potential, however synthetic benchmarks like Sandra are OK for what they are, but are not very representative of actual performance. A more telling reveiw would have been whether or not there was any performance difference between PIII Cu and PIII Cu Xeon boards and what justifies the price differential of $50-$100 per processor.
Another good review would have been to use some real world benchmarks against a real dual Xeon 500 with 1MB cache and see how the two compare. Real Xeon 500 w/ 1MB cache currently go under $400 each, which is pretty competative and when combined with memory that is a lot less expensive than RDRAM makes for a pretty compelling server or high end workstation platform.
How about using Photoshop? How about using some sort of server test? How about using Seti@home as a test? Even using CLIBench would give more information than just basing performance on Sandra...
What is the difference between the Xeon reviewed in this article and a regular P3 anyway? And unless I'm going blind (a possibility), the article never mentions the Xeons' speed. Further, the benchmark used doesn't even stress a CPU's L2 cache well (reference http://www.firingsquad.com/hardware/p3vc2/page5.asp)!!
I'm not one to post these kinds of messages, but what good is this article anyway? And RDRAM has a higher latency than regular SDRAM the last 10 articles I read...
I think the review is in error. I believe L2 on die cache of the pIII cu Xeon is like other Xeons available in 512KB, 1MB, and 2MB, or will be soon. The info on the Intel site is only current up to 500 and 550 MHZ Xeons. Look here: http://www.intel.com/ebusiness/serve...oductbrief.pdf
I have a dual PIII-450(512k cache) with 384Mb PC100, UDMA 33 disk Maxtor 10,8Gb + 8,4Gb WDC+4,3Gb WDC,cd teac 32x,dvd creative 2x + dxr2, cdr teac 6x/24x, matrox G400MAX 32Mb and creative sb128 pci. With sisoft sandra is my cpu performance is 2831 it/s and the new XEON III has only 700 it/s more (3500it/s).this mean that the xeon III 666Mhz is only 20% faster than a
Why should i buy a new Xeon and pay 3 thimes more that i have for my comp(i paid 2000$ in october 1999 - this is the price in slovenia today 1750$)
[This message has been edited by nebivedu (edited 04-19-2000).]
Apparently, the review is correct in that there is virtually no difference between a "Cascades" XEON and a PIII Coppermine. However, a "Tanner" XEON does/will include 1 and 2 MB L2 Cache. See a nice roadmap that explains this at Tom's Hardware Guide: http://www6.tomshardware.com/cpu/00q1/000228/ir-10.html
I too would have liked to see this review compare an 800MHz PIII Cu vs. a 800 MHz Cascades Xeon, just to verify that there's no appreciable difference. I'd also be curious to see how a 550MHz Tanner Xeon stacked up against a 800MHz Cu.
Not to knock in this benchmark too badly but two things:
1) What power user _EVER_ uses RAMBUS? It's for gamers because SDRAM is faster and has lower latency in large quantities (Abolve I think it was 128 or 256, but I don't recall.)
2) My 40 MHz SPARC32 CPU benches roughly 10% higher than this thing. I know an equivelant SPARC based system should run faster because its RISC, but 40 MHz 48 Mb FPM RAM and 1.2 Gb Fast SCSI-2? That's not evey approaching equivelant. I think you need to rethink how you do your benchmarking.