An early look at Phenom dual / tri core performance

http://www.erenumerique.fr/test_pr [...] 04-11.html

The article is in French, but the benchmarks speak for themselves. It also shows the differences between ganged and unganged memory mode, something most articles have failed to mention, so kudos to the author for the added testing. It also serves as a good yardstick for measuring clock for clock performance between K8 and K10.

In a nutshell, a Phenom X3 @ 2.3GHz is ~= to E6750/E6850 in multithreaded benchmarks, with lower single threaded performance obviously. The Phenom X2 @ 2.3GHz is a good match for a E6400.

What is striking to me is that clock for clock, Phenom is not that much faster than K8. Improvements range between 5 - 10% generally. I was under the impression it was a lot higher than that...

These results make it very obvious why AMD is delaying dual/tri core Phenom until Q2 2008. Unless AMD can get the clockspeed up, a Phenom X3 is no better than current C2Ds, and they may as well forget about Phenom X2 until it reaches 3GHz, otherwise it will be outperformed by an X2 6400+.

interesting read

Half-Life... what a mixed bag. Here's what I can draw:

- It doesn't use more than 2 cores.
- There is a disadvantage when moving from 90nm 1mb/core cache to 65nm 512kb/core cache. Whether it's the latency or amount of cache, I can't be sure, but cache seems more likely (affects Core 2 also).
- AMD's limitation to 1mb/core cache creates an illusion of the GPU starting to bottleneck the test, while Intel shows that is not the case.

It's a poor test to judge % improvement. The 2.3GHz Brisbane core you compared with also loses 10fps vs. the 90nm Windsor core when matched at 2.6GHz. And who cares, when everything AMD has loses so heavily to Core 2 anyway?

Why not bring up something more positive, like how Phenom sets the curve for WinRAR or reaches middle ground between A64 and Core 2 in 3dsmax?

Please. I selected all of the benches, not just a few. All of them showed per-core improvement. Did I mention I hate "high and mighty" people?

I already knew that the TLB issue.... Ehhhh, never mind. You'll never change.

You most certainly did not 'select' all of the benches.
Here are all of the benches












Youve made the same comment about me never changing before, and I will give you the same response: You're right, I will never change...I will always despise liars and cheats, and will always work to afford them the credit they deserve.

So which one doesn't show scaling increases? Photoshop? Oh, so you just have to have something to say about the underdog.

Besides Anand got 15% so it shows that benchmarks ARE NOT GOSPEL. They are just a guide. I imagine that if the L3 bug IS affecting general perf (which I think it is) we will see at least 20% and maybe 30%.

It looks liek certain workloads don't hit the problem as AMD says, but if you look at Sandra Memory, NO chip should have higher bandwidth than a Phenom with 1066DDR2, but they ALL do.

I guess AMD is moving totally away from Intel's "methodolgy" so they need K10-optimized benchmarks or they won't see the improvements.

That's my guess though, not meant to be taken as fact, like YOU ALL USUALLY DO.

I can't wait to get a monolithic tri-core. Oh never mind I think I'll just stick to my glued Q6600.

C2D vs K10

Note I'm comparing the 2.33GHz E6550 against the 2.3GHz Phenom X2, so the results will favor the 2.33GHz by ~1% or so, keep that in mind.

3DMark05 GPU: +8.01%
3DMark05 CPU: +16.06%
3DMark06 CPU: +16.35%
Winrar: +8.2%
Photoshop CS3: +29.41%
PovRay: +32.31%
Half Life 2: +28.57%
3DS Max 9: +14.78%
x264 encoding: +9.88%

Average: +18.17% (~17% after compensating for 1.3% clockspeed difference)

Wow. This was more than I expected. A ~17% difference clock for clock between K10 and C2D. By the time Phenom X2 launches, it will be competing against Wolfdale 45nm, which would put the difference out to ~22%.

To put things into perspective, the slowest Wolfdale at launch, the $163 E8200, runs at 2.66GHz. It will take a 3.2GHz Phenom X2 to compete with Intel's slowest C2D in 2008. Unless AMD can get their **** in order real soon, they won't be selling Phenom X2s for over $150.

Anandtech's results were comparing a single socket QC Barcelona vs dual socket DC Opteron, I would assume the dual socket Opteron would be running at a slight disadvantage in terms of per core scaling.

This is a direct Phenom X2 vs A64 X2 comparison, so I would say this is a more accurate reflection on the scaling between K8 and K10.

You're not seriously thinking Phenom performance will jump 20 - 30% due to the L3 bug right?! I am trying to be civil here, but sometimes you make it very hard for people to take you seriously.

Yup, he makes outlandish claims that end up being false. "AMD will change the color of the shipping boxes, that'll give 'em a good 2-3%".

I'll give him credit on this one, he didn't present it as fact, but rather as speculation. However, he needs to be careful, benchmark cherry picking is a Sharikou strategy.

That drop in Photoshop makes me wonder how they "created" the dual core Phenom I am wondering if it has to do with the way the memory controller works in the Phenom compared to how it works in the Athlon 64's making it the reason for the slow down.. We will see when the actual Phenom X2's are released if this is an issue or not.

My thinking is that if they created it by using the affinity command in Windows is that they may have used two cores that were connected to one memory controllers instead of the ones that are connected to separate controllers. Either way I can't read French to find out what they did so.. If someone knows do tell..

EDIT: Ok nm I see what they did.. Still the issue I talk about is still viable for what they did. I am wondering if those two cores are on the same side of the chip, heck even being adjacent from each other might cause a latency issue..

Now they also know why you can't even quote yourself in YOUR sig. You just take your silly reasoning and run. I'm not going to preface my statements anymore cause it never worked in the past.

My apologies for misinterpreting your post, after rereading it I understand what you mean. To be fair, your point wasn't exactly clear, and it seems others had the same interpretation as I did. And no, not because they are 'biased', but because you worded things rather ambiguously.

Anyhow, that is still a rather bold claim, you are predicting a 5 - 15% increase just from the L3 issue alone.

If you can't dazzle them with brilliance, bamboozle them with stupidity...

I thought you said the L3 bug affected at least 20%?

How come it becomes "grow to 20%" all of the sudden?

Look Baron, no one wants to prove AMD buyers wrong. AMD buyers know what they got when they made that purchasing decision. We only want to prove YOU wrong. Almost all of your claims are based on that random number generator in your head. NOTHING solid.

And I thought TLB errata lead to BSOD, not performance hit?

Really Baron? No one, and I mean NO ONE that I know of make that claim, only YOU. Even those who has extensive knowledge and experience in the field never claimed such number, only YOU

I guess we're all biased rabid Intel fanboy.

I don't agree. I think claiming it was a "native" tri-core was.

I don't really see this as anything different than the 6300/6400 with 2 MB of L2 Cache or the 21x0 line with 1 MB of L2 Cache.

If disabling a single core will for whatever reason let them run higher and can fill a "market" segment it will be a good idea. The question is where will they be priced.

Where did they claim it was a native quad core? I read a lot of news and have never read that they claimed it was a native tri core... But I do agree that it will be a good processor as long as it is priced decently..

After looking at the benchmarks closely I have to say that the Phenom X2 Benched here is being bottlenecked by something likely the point that it is a Quad Core running only 2 cores and likely those 2 cores are on the same Memory Controller (Phenom has 2 64-bit Memory Controllers one on the left and one on the right..) which would explain the Drastic change in performance from the Dual to the Tri-core on the PS CS3 benchmark..

Ding Ding Ding....we have a winner. Congrats to Yomamafo1 for calling it, here:
http://www.tomshardware.com/forum/ [...] 8_160.html
Posted on 11-20-2007 at 04:30:03 AM

Good call Yomama

I would really, really doubt this, as it's hard to squeeze 3 equivalent cores in circular fashion into a square. You could still efficiently fit this on a wafer with a tetris-like design, but that opens a whole set of problems with the patterning and circuit testing machines.

Far more likely, they're just fusing off bad or weak cores off a quad.

Yomamafor1, will you marry me?

I do think there is a problem with Phenom that will be worked out .No ,I know nothing of cpu more than tweaks and little basics,but this steak is not done .I believe the tri-core will come out with faster clocks (2.6 and up,cant recall where I read this)the end or 1Q 08 begining of 2Q 08.Which will be closer to the time of the Sb700 chipset and R700 gpu.For now the Phenom is a sucky over priced econo cpu.I'm patient

For those that are interested

After taking all of the scores and finding out how much faster AMD (unganged 2.3) is from moving from dual to tri, from tri to quad, and from dual to quad, and for Intel from dual to quad (2.4Ghz).

I have come up with this information

AMD Dual to Tri = AVG Increase = 26%
AMD Tri to Quad = AVG Increase = 20%
AMD Dual to Quad = AVG Increase = 54%
Intel Dual to Quad = AVG Increase = 45%

Of interest is also the point that AMD NEVER decreased in speed in any of the benchmarks although Intel Lost performance in both the 3DMark 05 GPU Bench (93% of the speed) and HalfLife 2: Lost Cost (96% of the speed) while AMD improved in speed by 3% and 1% respectively.

So I guess at least this shows that their Native QuadCore design does have it's benifits at least with scaling.

Oh and for those that wonder Intel did not once scale better in any given benchmark.. The minimum Scale difference was in Photoshop with 0.01% difference and a maximum of 22% in Winrar.

I will be interested in seeing Intells native quad core in action......I wonder if they will be able to get it right. Also I was reading that Intell will release an native 8 core cpu next year>>>>?

Quote from an article I was reading on Economist.com

"AMD faces three handicaps. First, it has only two chip fabrication plants, or “fabs”, compared with Intel's 15, which still supply three-quarters of all processors for PCs. This makes AMD more vulnerable to manufacturing problems: its newest chips are not as fast as promised and have been hit by production delays. It also means that Intel can usually introduce new manufacturing processes faster. With Penryn, Intel has started using hafnium, an obscure metal, to insulate the transistors in its chips. This has made it possible to shrink the tiny switches even further without leaking too much current. AMD will not reach the same level of miniaturisation until mid-2008."

Here is the link for the full article if you are interested.

http://www.economist.com/business/ [...] d=10180738