Phenom vs. A64-X2 Scaling

[Arkaign]

If this is true, then Phenom is done for, until they completely refresh it.

http://www.tomshardware.com/20...g_compared/page19.html

Performance per core does not scale as well as with an Athlon 64 X2 core between 2.2 and 2.8 GHz. This means that the performance gains of Phenom at future clock speeds will not be as significant as they have been with Athlon 64 X2 in the past. Let me give you some numbers to give you a better feeling: Athlon 64 X2 wins in 18 of our benchmarks, while Phenom 9000 only scales better in four categories. I would also like to emphasize that we used Asus's BIOS version 0603, which does not include a fix to the Phenom's TLB bug. Hence Phenom runs without any performance limitations.

Sad

C'mon AMD, stop self-destructing and bring something worthy to the table.

 

[Amaroque]

Personally, I take whatever Tom's Hardware writes with a huge grain of salt. His site has posted very faulty and biased info frequently in the past.

I'm aware that the C2D is faster clock for clock, and much more OC friendly. I still don't trust Tom's reviews whatsoever. :roll:

 

[Arkaign]

I know what you're saying. But the scaling numbers are pretty clear. We were told that the Phenom was going to scale really well, and it appears the exact opposite is true. C2D has nothing to do with that article.

 

[jiffylube1024]

One factor which is probably at play is that I think the 2MB L3 cache is fixed in Phenom currently at 2 GHz. So, the 512K caches are scaling up with clockspeed, while the other half of the cache is not.

Isn't Phenom already a total niche product?

 

[Arkaign]

Originally posted by: jiffylube1024
One factor which is probably at play is that I think the 2MB L3 cache is fixed in Phenom currently at 2 GHz. So, the 512K caches are scaling up with clockspeed, while the other half of the cache is not.

Isn't Phenom already a total niche product?

Man, wouldn't that asynchronicity on the L2 vs. L3 caches wreak havoc with latency? Seems like the most likely culprit for the horrible scaling.

But on the flip side, it presents a target for a better revision of Phenom in the future. Either increase the size of the L2 cache and drop the L3, or make them both synchronous.

 

[Amaroque]

Remember when the origional Athlon came out (I believe K7 Argon)? Cache was 1/3 to 1/2 of the core clock speed. It still outproduced the PIII. IIRC L2 cache was at 1/2 core speed except for some of the higher clocked models, and then the L2 was at 1/3 of core speed.

 

[Arkaign]

Originally posted by: Amaroque
Remember when the origional Athlon came out (I believe K7 Argon)? Cache was 1/3 to 1/2 of the core clock speed. It still outproduced the PIII. IIRC L2 cache was at 1/2 core speed except for some of the higher clocked models, and then the L2 was at 1/3 of core speed.

Athlon K7 had maybe a 1-5% advantage over the Katmai .25mu that also had external L2 running at async speeds. P3 Coppermine blew K7 away, until T-Bird leapfrogged it.

Anyway, apples to oranges. Phenom has L2 + L3, and stupidly, the L3 is stuck at a speed which apparently causes the scaling to be far less than ideal. It would have made sense if the Intel and previous AMD processors had a similar weakness, but it looks almost like the Phenom would have been faster with no L3 at all, just a larger sync L2.

 

[Amaroque]

You mentioned in a previous post that the L3 might "wreak havok" I was just pointing out that async L2 didn't seem to do that to the K7. I'm not a CPU engineer however. If someone with a Phenom could run benches with the L3 disabled in the BIOS, I suppose that would answer your question.

 

[Arkaign]

Originally posted by: Amaroque
You mentioned in a previous post that the L3 might "wreak havok" I was just pointing out that async L2 didn't seem to do that to the K7. I'm not a CPU engineer however. If someone with a Phenom could run benches with the L3 disabled in the BIOS, I suppose that would answer your question.

Well, look at K7 800Mhz vs. T-Bird 800Mhz. No actual cpu differences per se, just the die size and integration of the L2 to enable sync operation. The penalty from running the L2 at such a reduced speed was absolutely large. The K7 outperformed the Katmai P3 (also 1/2 speed external L2), but was leapfrogged by the P3 with on-die sync cache, the Coppermine.

So :

K7 core was better than P3 core inherently.
L2 cache penalty was bad enough for the K7 to lose to P3-Coppermine.
Once they both had sync L2, the Athlon once again came out on top.

 

[nyker96]

I remember seening a 2 core vs 2core report that shows about 10-15% improvements from X2 to phenom. So there's some gains to be noted. As for scaling, I think phenom uses some cache system that might be interfering with the results. Besides it's a monolithic 4-core design, maybe not too useful to run it like single coreded. Who knows. Bit apple to orange to me this type of comparisons.

 

[Amaroque]

Yes butt... The L2 was integrated onto the die of the Tbird, so it had something like 100x faster access to the core, plus 2x faster L2 freq. I belive there were other enhancements, but I don't really feel like looking up a K7 or PIII. lol

Just some random thoughts.

 

[harpoon84]

jiffylube1024, the L3 speed is actually 1.8GHz on retail Phenoms, not 2GHz as tested on ES Phenoms.

According to Techreport http://techreport.com/articles.x/13741/3 the 200MHz difference in L3 speed accounts to roughly a 2% performance penalty. I'd estimate that a Phenom 2.3GHz would be about 5% faster if it had synced L3 instead of 1.8GHz. Of course, as clockspeeds increase, the 1.8GHz L3 will become an even bigger bottleneck. Can you imagine a 3GHz Phenom stuck with 1.8GHz L3?

 

[Amaroque]

Oh, and yes, 1/2 speed 512k L2, to full speed 256 L2 did bake a good difference. I'm not sure if this applies to L3 though. Like I said, we need someone to run benches with L3 enabled, and disabled...