Athlon 64 Preview

[Wraith]

--" a higher count of L2 cache increases per-clock performance on average because it has less stalls waiting for system memory."

Which doesn't help much if you're limited by the speed of the core rather than the memory, as you're implying. A good prefetch will do more for you in most cases than the larger cache, which really helps only when there's a dataset to be worked on that fits in it (but wouldn't in the smaller one).

--"The design, the process, they all mesh together to give you a pretty lackluster performing processor. "

Again, you're completely missing the point of the PR ratings. A PR rated 2800 processor is going to perform pretty similar to aonther PR rated 2800 processor. By definition.

And no, you don't understand the difference between process and design issues.

--"So why did they elongate the pipeline if that was the goal of the design?"

Higher IPC and better scaling. We've already seen it definitely has the former. We won't be able to judge the latter for a couple years.

Remember how poorly the first P4s overclocked?

--"If the market is shrinking, it's AMD that's getting the shaft, and not Intel"

True enough, but you don't understand why. It's always been the case that market problems will effect AMD more negatively than Intel.

--"how is this "pushing it forward"?"

Intel's been pushing their ramping on the current core forward. Haven't really paid attention to Prescott dates.

--"BTW, AcesHardware posted news about the POVRay benchmark running on a 1GHz Itanium 2:"

They should have a good Operton benchmark up this week, judging by the hints they've dropped on their message board.

Wraith
"Sanity is a madness put to good uses."
-- George Santayana

[Urban Ranger]

Originally posted by Asher
I don't think you're thinking this through well enough -- a higher count of L2 cache increases per-clock performance on average because it has less stalls waiting for system memory. It improves IPC somewhat, which is what you need when your clockspeeds are less than stunning.

You are assuming that the execution units are fully utiliised at every clock cycle, than increasing data feed can increase IPC. This is however not the case, so you have to look elsewhere.

Originally posted by Asher
So why did they elongate the pipeline if that was the goal of the design?

To solve the problem of execution units under-utilisation.

[Asher]

Originally posted by Wraith
Again, you're completely missing the point of the PR ratings. A PR rated 2800 processor is going to perform pretty similar to aonther PR rated 2800 processor. By definition.

But it doesn't.

They should have a good Operton benchmark up this week, judging by the hints they've dropped on their message board.

April 23rd...

UR:

You are assuming that the execution units are fully utiliised at every clock cycle, than increasing data feed can increase IPC. This is however not the case, so you have to look elsewhere.

If you look at benchmarks in the real world, increased L2 cache does help modern x86 processors. It's a simple fact.

It helps the P4 more than the Athlon, but it helps nonetheless by increasing IPC.

[Wraith]

--"But it doesn't."

Yeah, pretty much it does. Especially considering that this is an old stepping and none of the software has even been recompiled to take advantage of the extra registers yet, much less the 64 bitness (which, you should know, Epic is planning to use in their next Unreal engine).

--"April 23rd..."

And here we are.

Opteron, with a non-NUMA aware OS still destroys the gig faster Xeon in most applications. It's currently weak in media encoding and archiving, but the DivX guys have been rather enthusiastic about the difference 64 bit support will make, so that shouldn't last too long.

It also scales a whole heck of a lot better, and again that's even with the non-NUMA aware OS. Once the 64 bit apps like DB2 and so on are around this should be quite good.

AMD looks posed to take a large chunk out of the 1-8 way server market this year. They've got two Tier 1 vendors (IBM and Fujitsu-Siemens) on board, and are completely blowing away the Xeons in most server applications. And after things are going for a while they'll be making big inroads into the up-to-32 processor field, as Newisys has already announced a 32 processor server for later this year.

AMD's finally doing something right. This is the high-margin market. AMD should finally start making money.

Wraith
"Everybody's entitled to a little confusion in their lives. I practically thrive on it."
-- Commander Cain

[Asher]

Originally posted by Wraith
Yeah, pretty much it does. Especially considering that this is an old stepping and none of the software has even been recompiled to take advantage of the extra registers yet, much less the 64 bitness (which, you should know, Epic is planning to use in their next Unreal engine).

As I understood it, Epic plans to compile for x86-64 (not like it takes any effort), but only the content creation stuff would probably require x86-64.

And here we are.

Opteron looks like a killer Xeon competitor.

But as some of the benchmarks have shown, it's nowhere near the Itanium II ballpark (as someone in this thread implied it'd be competing with).

The workstation performance of Opteron isn't that great, but for servers it looks fantastic.

BTW, have you heard about the new emulation software for Itaniums that would make a 1.5GHz Itanium II behave like a 1.5GHz Xeon? I know that's still pretty sad for 32-bit performance, but it's faaaar better than what they've got now.

[Asher]

Itanium II (1GHz) vs Opteron (1.8GHz).

The 1.5GHz Itanium II with 6MB cache is due out soon, and a 32-way system with it is now the fastest computer in the world to run TPC-C...

[Wraith]

--" As I understood it, Epic plans to compile for x86-64 (not like it takes any effort), but only the content creation stuff would probably require x86-64."

Right. They're still making a gaming engine, so they need to have as many people as possible be able to use it.
It will be very interesting to see what kind of performance boost the 64 bit version will get, though.

--"But as some of the benchmarks have shown, it's nowhere near the Itanium II ballpark "

Depends what you're talking about. Itanium II still beats it in most FP stuff, but at a price. The Opteron really cleans up in Scalar SSE-2, so it depends on your app. There will be a some things that the Opteron will definitely beat the Itanium in as far as price/performance goes.

--"The workstation performance of Opteron isn't that great, but for servers it looks fantastic."

What are you basing that conclusion on? I haven't seen any good workstation focused benchmarks yet, and Ace's review indicates good performance in a lot of workstation apps, like 3DSMax.

--"BTW, have you heard about the new emulation software for Itaniums"

I've heard of it, yeah. Guess they had to give the ex-Alpha team something to do
Suspect it'll be about as popular as FX!32 was, too.

--"and a 32-way system with it is now the fastest computer in the world to run TPC-C..."

You should check that submission closer. That was a 32 system cluster...

And as far as POVray goes, I'd wait for independent confirmation of that before I'd take it without suspicion.

Wraith
1066 - Too much Saxon violence

[Asher]

Originally posted by Wraith
What are you basing that conclusion on? I haven't seen any good workstation focused benchmarks yet, and Ace's review indicates good performance in a lot of workstation apps, like 3DSMax.

Tom's Hardware -- your favorite site.

http://www.tomshardware.com/cpu/20030422/opteron-23.html#3drendering

The Xeon is actually cheaper and wins in most Workstation benchmarks. Lightwave, Cinema4D, 3D Studio MAX, mp3 maker, MPEG2 encoding...in fact the only workstation programs the Opteron won were WinRAR and SETI@Home.

The multitasking performance in particular, the Xeons dominated...

You should check that submission closer. That was a 32 system cluster...

Oh? Well, either way...

[Wraith]

--"Tom's Hardware -- your favorite site."

I figured as much, just wanted to be sure.

Tom's makes too many factual errors in his review for me to take it seriously.

And who spends several grand on a workstation machine to put a Radeon 9700 Pro in it? Sorry, but this guy is just impossible to take seriously any more. It's not like he doesn't have actual workstation graphics cards to test with, he does reviews of them too.

Edit:

LMAO! BTW, remember that he's testing a 1U Newisys Opteron server... it doesn't even have an AGP slot. It's on integrated video. So much for his workstation graphics tests.

Wraith
There is no substitute for good manners, except, perhaps, fast reflexes.

[Urban Ranger]

Originally posted by Asher
If you look at benchmarks in the real world, increased L2 cache does help modern x86 processors. It's a simple fact.

It helps the P4 more than the Athlon, but it helps nonetheless by increasing IPC.

That's because a failed branch prediction will kill the P4 far more than the Athlon due to the deep pipeline.

But my point hold nevertheless, the 256K increase in L2 cache helps the Barton some, but doesn't seem enough to justify the increase in cost.

[OneFootInTheGrave]

auch...

Mighty MSFT favours AMD Opteron

Who would have thought so ...

http://www.microsoft.com/exchange/techinfo/planning/2000/perfscal.asp

[Asher]

Where's the IA-64 benchmarks for that.

[OneFootInTheGrave]

Originally posted by Asher

Where's the IA-64 benchmarks for that.

well when you mentioned it...

double ouch

Quad Opteron Benchmark Results (AMD)
By Brian Neal
Thursday, April 24, 2003 7:54 AM EDT

Recently, some readers have asked how the Opteron stacks up to the Xeon MP. As you know, the Xeon MP currently runs at up to 2.0 GHz and features as much as 2 MB of L3 cache in addition to a 512 KB L2 cache. It also uses a 400 MHz FSB. The Xeon MP is used for Xeon-based systems configured with four or more processors. While we haven't benchmarked a Xeon MP system ourselves, there are several results published on the Internet, some of which are highlighted below.

Thanks to Chris at AMDZone for letting me know about these Microsoft Exchange 2000 benchmark results. The results not only include a number of Xeon MP-based systems, but a 1.8 GHz quad-Opteron system from RackSaver as well. The Opteron has a messaging throughput of 15,520 (MMB2) with a response time of 88 ms. The nearest competitor is an IBM eServer xSeries 360, a 4-way 2.0 GHz Xeon MP system with a messaging throughput of 13,200 and a response time of 183 ms. The 4-way Opteron results can be found here, with the Xeon MP results here.

The same quad-Opteron system has also been benchmarked in TPC-C and results have been submitted. Equipped with 4 CPUs and 32 GB of RAM, the Quatrex-64 4P Opteron (Executive Summary, Full Disclosure Report) manages a tpm/c of 82,226 at a price/performance ratio of $2.76/tpmC.

This compares favorably to a Dell PowerEdge 6600 (Executive Summary, Full Disclosure Report), a quad 2.0 GHz Xeon MP system also configured with 32 GB of RAM. The 6600 rates at 78,117 tpmC at a price/performance ratio of $4.85/tpmC. Both the quad Xeon and Opteron systems are using MS SQL Server 2000 The Opteron also rates slightly faster than HP's 4-way 1 GHz Itanium II system, the rx5670 (Executive Summary, Full Disclosure Report). That system has a throughput of 80,571 at $13.26/tpmC. There's also a less expensive configuration running Linux, that delivers 80,495 tpmC for $5.30/tpmC.

from http://www.aceshardware.com/

not too shabby for AMD...