The third Opteron concerto with orchestra

Just recently we've published the previous article of this series where we've described some aspects of K8 architecture internal organization. It's not that we are going to stop digging deeper, but it's time to estimate the improvements in practice. How good will this architecture behave comparing to Pentium 4 and Athlon XP?

Both us, and readers have been waiting for results for a long time. Some wanted to check if their guesses are right, some hoped K8 would be a "saviour" for their systems, and some - to gloat over. Well, let's hope everyone would be satisfied. If you remember, in the previous article we described the concepts and details of K8's memory controller with the help of low-level tests. There's no sense in repeating these, so we'll just be referring to obtained data when needed. This time we'll examine the results of Athlon 64, Opteron, Athlon XP and Pentium 4 comparison in real apps, as there's already a lot of practical material worth thinking over.

First, we've compared Athlon 64 and Opteron. Second, before the Athlon 64 is officially announced on September 23, we'd like to have at least some performance results to understand this novelty's capabilities. Moreover, after the official announcement (the clock speed is rumoured to be around 2GHz) we will be able to check if the architecture is as scalable as expected. Thus it turns so that we have to measure samples' performance anyway. Third, it would be interesting to check Athlon 64 performance behaviour, as earlier samples are Revision B, but already Revision C is going to be sold. And AMD is also rumoured to improve memory controller efficiency (again?!)

In other words, we decided to publish this testing despite it's not that hot already. Moreover, the concept of thorough AMD64 and K8 architecture examination warns against hot "sensational" testings. Besides, to understand the Athlon 64 position among rival solutions we'll compare its results with high-end Socket A and Socket 478 processors, respectively.

Now that we've convinced (at least ourselves) of the testing necessity, let's list the test equipment required.

Performance examination

K8 platform:

• Opteron 240 (1400MHz) processor, Socket 940
• ASUS SK8N Opteron motherboard on áàçå nForce3 Pro
• Two Athlon 64 processors, 1400MHz and 1600MHz, Socket 754
• Soltek K8AV-R Athlon 64 motherboard on VIA K8M400

Pentium 4 platform:

• Pentium 4 3.2GHz (16x200) processor, Socket 478
• ASUS P4C800 Deluxe motherboard on i875P

Athlon XP platform:

• Athlon XP 3200+ processor (11x200=2200MHz), Socket 462 (Socket A)
• ASUS A7N8X rev2.0 motherboard on NVIDIA nForce2 Ultra 400

Opteron-based system featured 2 x 512Mb Samsung PC2700 ECC Registered memory, but TwinMOS DDR400 (2 x 256Mb, 2-2-2-5) modules were used in other platforms. To avoid other bottlenecks we also used GeForce FX 5800 Ultra graphics card.

In diagrams we'll provide the comparative performance for all the systems and will try to estimate the performance of the yet hypothetical Athlon 64 2GHz with linear extrapolation.

As the result we obtained data for two Athlon 64 clock speeds to which we anytime can append the (0,0) point. Though the dependence of performance on clock speed is non-linear, it's rather close to that in the scale start to neglect the approximation error. However in specific tasks, where the external hardware is the limiting factor, the straight line drawn through two points (corresponding to 1400MHz and 1600MHz), won't cross zero. And the higher the clock speed would be, the smaller the relative boost would be under the same conditions. Thus it would be more correct to use hyperbolic extrapolation below:

P = a*F/(1 + b*F),

P - performance (points / time unit);
F - frequency (clock speed);
a, b - task-specific coefficients for external hardware and processor.

(Hyperbolic dependence results from the fact that a composite process performance "constant" is a geometric average of consequent process performance constants.) In other words, the linear extrapolation is useful to runtime-frequency reciprocal dependencies, not to performance-frequency ones. However, in this article we decided to confine only with the results of hopefully understandable linear extrapolation, as the estimated upper error bound made 2% for our tests.

Therefore we should obtain slightly steep performance estimation for a 2GHz processor. The newer core revision might add a bit as well, but we won't consider this for now, of course. Besides, raw Athlon 64 drivers might snap some performance too.

In general, this calculation is called to demonstrate the approximate capabilities of Athlon 64 2GHz. And when it's available we'll compare the actual results with those we've achieved here.

Test results

First, let's look at game tests (note the asterisk on hypothetical Athlon 64 2GHz clock). The most indicative here are the low-resolution results where graphics card influence is minimal. But besides, we'll provide higher-resolution results, as those who'll be buying new processors won't play at 640x480, of course.

Well, Pentium 4 looks good, but we remember that Athlon 64 starts from 2GHz. In this case the former has small chances to win though it's the highest-end for now. By the way, it's interesting that the previous-generation results actually correspond to 1600MHz sample performance. This means AMD has made a considerable step forward comparing to K7. Now let's look at separate low-res tests in detail.

You can see that Athlon 64 2GHz should either be better, or equal to the rival. Pentium 4 wins one test, but the general situation is better for AMD. We can easily consider that the novelty will make a good debut in 3DMark2001 SE. And what about 3DMark03?

From the angle of this test, Athlon 64 2GHz slightly lags behind Pentium 4 3.2GHz. Nevertheless, you can see the progress comparing to the previous generation, as the difference might be twice as small. It's also interesting that Opteron-based system has lost in every test so far. This is the result of both processor and memory lower clock speed, and higher registered memory latency. All these factors indicate the server platform won't be good for Direct X games. But still it's only a server :). Let's now proceed to Return to Castle Wolfenstein.

Here we see that Opteron lags in OpenGL games as well, so it seems senseless to build a gaming system on it. Our favourites - Pentium 4 3.2GHz and Athlon 64 2GHz - perform side by side, but the former is slightly better. However, considering that RtCW engine is aligned to Intel's solutions, the result is still very good for our newbie. It's also indicative that in third app in a row Athlon XP 3200+ is very close to Athlon 64 1600MHz. Well... I wonder how AMD will rate its 2GHz processor? :) But, let's proceed further to the AMD-aligned Serious Sam 2.

Just wow! Athlon 64 1600MHz looks way better than its double-clocked rival, while the hypothetical Athlon 64 is simply unattainable for Pentium 4. A clear win. Even Athlon XP 3200+ looks rather good on the level of 1.6GHz next-generation model.

Let's finish with games for now, as you can see the novelty should be good for them. Now let's move to professional 3D graphics in SPECviewperf 7.1.

In this test Athlon 64 has only slightly lagged behind Pentium 4 3.2GHz sometimes outrunning it, but ceded to Athlon XP with higher clock speed. However, we noticed that Opteron has been losing to its "junior" brother (sometimes considerably). It seems that professional apps are sensible to both bandwidth, and latency. We can't think of any other explanations for this. I wonder how they are going to sell Opteron for workstations, if it cedes to a cheaper desktop processor? Of course, this is a junior model, but AMD marketing specialists should think it over to position the processor correctly until faster DDR400 registered memory is available.

Now let's compare Pentium 4 and Athlon 64 computing capabilities.

AMD camp hasn't had a chance to celebrate for a long time, as due to 3ds max optimization for SSE2 and Hyper Threading the final rendering speed was considerably higher on Pentium 4. AMD engineers were able to reduce the gap, but Intel's model is still the leader in general. Nevertheless, we must mark the considerable improvement comparing to Athlon XP 3200+, as already Athlon 64 2GHz should outrun this previous-generation AMD model despite the lower clock speed.

To check, if the win is accidental or not, let's use CPU RightMark capable of measuring performance of different computing methods.

AMD improved the core computing performance having added the SSE2 support and maintained the same x87 instruction performance as before. The SSE2 optimization boost is slightly smaller than Pentium 4's, but still Athlon 64 2GHz is right behind Intel's top model.

But things are not that good in scene rendering, as Athlon 64 lags behind Pentium 4 almost corresponding to the clock speed difference and loses to even Athlon XP. Moreover, AMD just has nothing yet to answer to HyperThreading with.

Now to archiver tests. As we remember, these applications are most sensible to memory latency as well as the bandwidth. Thus we can expect Athlon 64 to outrun the elder server brother again. Besides, it's interesting to compare it with the previous-generation Athlon XP 3200+.

We were right - Opteron was about 10% slower. You can also see that Pentium 4 3.2GHz results are easily achieved by 1.6GHz model - especially on the background of 3200+. Athlon 64 2GHz should naturally become the clear and evident leader.

Now let's take a look at Pentium 4 favourite - video and audio encoding.

Athlon 64 just failed. In audio encoding Athlon 64 2GHz should only catch up with its previous-generation sibling (and outrun it with the same advantage at the same clock speeds), while the rival CPU is far ahead. It seems AMD processor performance depends only on clock speed here, which means SSE2 optimization is not enabled. But things get stranger in video encoding. This is the only real app where Pentium 4 is considerably better than Athlon 64. Moreover, in the similar CPU RightMark the Athlon XP 3200+ system outperforms the 2GHz successor by 30% (while their clock-specific difference should have made only 10%). We can think of two possible reasons (or even their combinations):

1. DivX just doesn't know anything about Athlon 64, and it doesn't enable optimization under Athlon XP (this seems the most possible, as similar cases have been before);
2. This is the result of nForce's DASP. In this article you can see that Athlon XP shows worse results on KÒ600 similar to those of Athlon 64. We might assume that exactly in this task NVIDIA's DASP works best.

Anyway, Pentium 4 wins in this round, and we'll try to return to video encoding in the future.

Finally, the dessert - SPEC CPU2000 results.

Athlon 64 2GHz should perform on the level of modern processors. Pentium 4 3.2GHz remains the formal leader, but the SPECint difference is too slight to be a serious factor. Actually Intel clearly wins in SPECfp only with about 20% advantage.

On this conciliating note let's draw necessary conclusions.

Conclusions

1. Athlon 64 2GHz expected on September 23 should be a competitive product. At least, it looks better than we thought it would :). We hope our extrapolation1 wasn't that rough. Anyway there's not much time to wait. According to our calculations, already at 2GHz Athlon 64 should lead or catch up in almost all tests. If AMD achieves 2.2GHz in the nearest future, this processor will have all chances to win.
2. We didn't expect the difference between Opteron and Athlon 64 with the same clock. Though 2 memory controllers are more than 1, it turns out that a faster single is better than a slower pair. However, this should change with the announcement of Registered DDR400 ECC. And now it's necessary to re-conceive the older Athlon 64 testings from the Internet based on 1-way Opteron systems performance and the assumption that Athlon 64 should be a more expensive and thus faster variant. This turned out to be wrong!
3. The Web brings rumors that on September 23 AMD should announce two Athlon 64 models at once, with the second being designed for Socket 940! In other words, designed for the same boards, as Opteron. But we wonder what memory such processor will use? Registered ECC is expensive and, as we've showed, almost senseless and bad for a desktop system! The increased latency will negate higher bandwidth. We believe such Athlon 64 would work with usual DDR400! This makes more sense, as if Athlon 64 maintains its low latency, the wider memory bandwidth would only be better! Moreover, it provides some logic, as the Socket 754 model would catch up with the rival, and Socket 940 Athlon 64 would become a kind of a desktop Hi-End. Still we must remember these are only rumours, so let's wait for the facts.
4. We can assure you the new generation will show the significant advantage comparing to the previous solutions. Moreover, the first Socket 754 model actually performs on the level of top Pentium 4 and Athlon XP. This makes upgrades sensible, just don't forget to consider the price.

We believe we've completed our task and compared the three architectures and demonstrated the supposed Athlon 64 performance. Stay tuned to next articles!

Write a comment below. No registration needed!