Intel Versus AMD: The Truth Unfolds

It has not been easy to find out the real story of how Intel hardware compares with AMD hardware. The reviews from places like Anandtech and Toms Hardware Guide have been so sloppy and unprofessional that no useful information could be obtained. However, after a second review at Tech Report there is finally enough accumulated data to draw some conclusions.

Often when I mention the horrible state of hardware testing today I get some flack from Intel fans who really want to believe that any published test results that show Intel favorably must be true. I suppose there have been a number of people who have suggested that I have been somehow denying reality by not embracing these results. Of course, these people don't understand at all how I think. Poor testing doesn't mean that AMD is better; it just means that you don't know. With proper testing, the results could show that Intel is even better than what was shown with poor testing. I suppose since I'm now talking about testing done at Tech Report some must be wondering if I am endorsing Tech Report as a good fact source. Well, not exactly. The problem is that it has taken two reviews (and sometimes a third) to get the facts at Tech Report. If the testing had been truly prossional then we would have been able to get what we needed from the just one review. But, this is not the case; it is necessary to look at two or more reviews to find out how reliable the tests themselves are. What this means is that some of the data published by Tech Report does not fully show what is happening with the hardware and therefore this data can be misleading (and the author's comments can be incorrect as well). However, by using more than one set of tests there is finally enough information to draw some conclusions. I suppose this is at least better than most other review sites where the testing still shows essentially nothing even with two reviews in hand.

One of the issues that keeps getting tossed out is the idea that Intel processors now draw a lot less power than AMD's. This seems to be due in part to a desire to throw off the very bad reputation that the Prescott cored processors had. Yet, there is no indication of any big advantage for Intel. The highest clocked chips from both companies are rated at 120 watts and the second level from both companies runs about 80 watts. You have to add a bit to the Intel numbers because their TDP isn't a maximum like AMD's and you also have to add the power draw for the Northbridge which is included with the AMD numbers. However, Intel's C2D processors have much better power management than AMD's current Revision F and 65nm Revision G processors so this tends to even out. The sole advantage that Intel currently has for any class is for their quad cored processors and this is only because AMD doesn't have a quad core out yet. You could however argue that if you rate processors by SSE performance per watt then Intel is better. This is true if you use SSE most of the time such as for heavy duty rendering or scientific or engineering calculations. But, for most areas Intel and AMD are about equal. This shouldn't change until late 2007 when Intel releases 45nm chips.

The first benchmark is Cinebench which renders a graphic scene. Surprisingly, Intel and AMD are almost dead even on this benchmark. This is not due to AMD's greater memory bandwidth because this test is not memory intensive. I would have expected Intel to be faster but this is not the case. At the same clock speed the two are tied. This benchmark also does not scale very well; perhaps the code will be better by the time AMD has its own quads out.

The next is Pov-Ray and once again the results are very close; Intel trails by about 4% at the same clock. Either Pov-Ray does not make use of SSE or AMD is gaining more from the 64 bit code. This application is not memory intensive but it does show good scaling even with 8 threads. A dual socket Clovertown Xeon 5355 system should deliver about the same performance as a quad Opteron 8218 system and a single socket Kentsfield QX6700 should be about the same as a dual socket Opteron 2218 system.

With 3DS Max scanline rendering there are two factors. First of all, for a comparable system, Intel is 25% faster at the same clock than AMD. This is probably due to better SSE. Secondly, this code does not scale very well. Intel shows a 22% drop-off when scaling from 2 threads to 4 while AMD shows a 29% drop-off. This looks suspiciously like poor NUMA support. Finally, in contrast to the scanline rendering, 3DS Max's ray tracing code is so poor for multi-threading that it isn't useful at all for comparison beyond 1 core.

In the Quad FX review the Valve Source engine particle simulation numbers are all over the place. I've tried to see if anything can be gleaned from this information but unfortunately there are no matching numbers in the Clovertown review. This is a good example of a benchmark that less knowledgable fans might point to as an Intel win but these numbers are far too erratic to tell anything. The Valve VRAD map build however shows Intel 26% faster at the same clock. There really isn't anyway to tell if this is due to cache, integer IPC, or SSE but Intel is clearly faster. This will be a good benchmark to repeat with K8L. Since K8L should increase cache, integer IPC, and SSE the results should be much closer.

3DMark is not in the Clovertown review but after comparing with an older review with C2D E6300 it is clear that 3DMark 06 CPU Test 1 is good and not simply gaining from larger cache; Intel is about 10% faster.

The MyriMatch test is interesting. It is clear that this code is terrible with multi-threading and quickly loses efficiency as threads increase. However, the numbers drop off not only with more threads but with faster clock speeds as well. Worse still, the Intel numbers show variances of 5% when increasing clock by 15% and the AMD variances are even greater at about 8%. This leads to a very unreliable spread showing Intel is somewhere between 3% and 27% faster. Since I did not see this variance in the other tests and there would be no reason for Tech Reports to have reconfigured the systems during this test I would have to say that the code is at fault and the test is therefore useless.

The Euler 3D computational fluid dynamics test is a good example of how misleading the data can be and how equally far off the author's interpretation can be. He claims that the code is both floating point and memory intensive. Yet, when comparing the two sets, there is no evidence of any drop-off due to memory. With double the memory bandwidth, the dual socket woodcrest system is actually slower because of the higher latency FBDIMM's. There is no gain at all from the higher memory bandwidth, therefore the code is not memory intensive in the sense of saturating the memory bandwidth. Intel seems to be about 18% faster. However, some of the scores are far worse than 18% and this again brings up suspicions of NUMA problems. Looking more closely, we see that in spite of claims that the Quad FX tests were done using a NUMA compatible OS, there is clear proof of the opposite. Quad FX and the dual socket Opteron 2218 system are very similar as both are dual core on dual socket. And, their scores are almost identical even though the FX is clocked 15% faster and uses faster memory. Apparently, the interleaving that was used with the Opteron is making up for the faster speed of FX, but this could only happen if the FX were running without NUMA support. This means that for any code in the FX review that uses large working sets the FX processor will show an anomalous dropoff. This is why one review was not enough; you can't see this without comparing with the second review. So, Intel is 18% faster rather than the 73% as would be suggested by the FX numbers. Even with the interleaving hack this very poor number would have been cut in half and then half again with proper NUMA support.

The Folding @ Home series is interesting. But, there are errors in the totals for both reviews. The actual totals are pretty close and would give Intel about 6% more speed. However, what we really want is the best rather than totals or averages and this would give Intel about 25% more speed. However, it is stated that the same code for AMD uses 3DNow! rather than SSE so the actual difference could be less. However, I think 25% due to better SSE is reasonable. This is a benchmark to watch because it should have an increase with K8L unless the code is not actually making use of the AMD hardware.

The Sandra benchmarks are both good and bad. The Multimedia Integer tests are way off. I'm not sure what they are testing but whatever it is does not show anything real about the processors. If these scores were taken at face value Intel would be 253% faster in integer. Or to put this a different way 1 C2D core would be 89% as fast as 4 AMD cores. This is yet another published score that unknowledgable fans might tend to mention as a sign of the phenomenal speed of C2D. However, this is obviously impossible theoretically, architecturally, and from the fact that no other test shows speed like this. Therefore, we have to ignore this test. However, the Multimedia Floating Point test looks good. Intel is 79% faster. This is reasonable since Intel can easily do 100% faster with SSE and could theoretically, with perfectly matched data, actually hit 4X AMD's speed. However, this is not a likely thing to occur. This is another benchmark that should become about equal with K8L. Perhaps by the time AMD releases K8L we will know what is wrong with the Integer test.

The Windows Media Encoder shows a significant drop with 4 threads, but it drops about the same for Intel and AMD. With the original code Intel had a 23% lead, however with the newer 64 bit code this lead drops to only 15%. This is true for both 2 and 4 threads.

PicColor contains some anomalous scores which lessons confidence in the set's overall accuracy. The extremely low cpu usage graphs add to the uncertainty. The best I can say is that for most tasks Intel seems to be about 12% faster but is about 60% faster on floating point.

Panorama Factory is not memory intensive. It shows poor scaling with 4 threads and truly horrible scaling with 8 threads. However, the scores appear consistent. Intel is 7% faster. The only odd score is the FX-74 score which is slower (12% instead of 7%). This seems to be due to NUMA problems for the large datasets. Apparently, interleaving is enough for this task but FX-74 without even that shows a drop-off.

Tech Report has used three different methods of power consumption testing with the three reviews I referenced. This somewhat diminishes the value of the power consumption tests. For example, the Quad FX review does show some advantage for Intel, however the load application is Cinebench which has poor scaling. In the Clovertown review the load application is Pov-Ray which scales well and the power consumption is very similar. Realistically, Intel only seems to gain with lower power draw with its Kentsfield QX6700 processor. There is no real difference with dual core/single socket between Intel and AMD nor with dual socket.

I generally give Intel a 15% integer lead so I'm surprised that Intel doesn't seem to show this kind of lead for a lot of the tests. For example, I usually match X2 4200+ with E6300, 4600+ with E6400, and 5000+ with E6600. Yet, it seems for many tasks, Intel only appears to be 10% faster. This would mean that 3800+ would match with E6300, 4200+ with E6400, and 4600+ with E6600. This would make 5000+ equal to E6700 but AMD would still have no procesor to match X6800. This seems to be about right if we exclude SSE. However, if SSE is needed then Intel is the current bargain.