In our recent DDR3 vs. DDR2 review we discovered a 16% to 18% improvement in memory bandwidth with the P35 chipset. This translated into a 2% to 5% increase in real world performance in some computer applications. This represents a significant improvement in performance. Our review attributed the increase to the P35 memory controller, as we found no real difference in the performance of DDR3 and DDR2 on the new P35 chipset. Both memory types were much faster on P35.

While it is true that P35 exhibits these levels of performance improvement, the reasons for this performance boost are more complicated than the memory controller alone. If you look back at the test configuration, the P965 tests were run at 2.66GHz with an X6800 driving the tests at 10 x 266, while tests on the P35 used the same 2.66Ghz, but with an E6420 running at 8 x 333 speed.

The reason for these configurations was the problems with ratios on P965 and P35 motherboards. We wanted a test bench that examined the same speeds on all three tested boards. However, there is no way to run the P965 at a 333 base FSB (1333) with memory at either DDR2-800 or DDR2-1066. The available ratios do not allow this setup. Similarly the P35 chipset motherboards will not allow the DDR3-1333 setting unless the CPU is set to the 1333 speed. Setting the P35 to 10x266 limits the top memory base setting to 1066 without overclocking. With limitations on both sides the choice was made to test all boards at 2.66GHz, with P965 at 10x266 and P35 at 8x333.

This is a fair setup for the tested chipsets, since P965/975x were designed for 1066 FSB chips and official DDR2-800 memory and P35 is designed for 1333 FSB processors and official DDR3-1066 or official DDR3-1333 memory. P35 also is backwards compatible with 1066 FSB processors and it can also support DDR2 memory if the board manufacturer chooses. The 16% to 18% bandwidth improvement between these tested and representative configurations remains accurate and translates to a 2% to 5% real world improvement in several real-world tests.

In the review we attributed the performance increase to the P35, rather than the DDR3 memory. Technically this is correct, but there are two components at play in the performance increase. First the P35 memory controller itself is improved to provide better memory performance at the same memory speed and timings, and second the chipset supports the 1333 FSB which also improves memory performance. Since many readers have asked which is the primary factor in the memory performance increases - architecture or bus speed increase - benchmarks were set up to try to separate the impact of the memory controller improvement and the bus speed increase on P35.



Memory Test Configuration

In the original DDR3 vs. DDR2 review a 4MB cache E6420 was overclocked to provide 8x333 2.66GHz testing. An X6800 was used to provide the 10x266 2.66GHz configuration for comparison. For the revised tests, we used the X6800 for all testing since it was simpler to use an unlocked CPU with the flexibility to run all of the CPU configurations needed in our benchmarking. In addition we have seen some variation in the performance of different Core 2 Duo processors and we wanted to remove that variable. While we don't know exactly why, X6800 processors are often a bit faster in gaming benchmarks than some other Core 2 Duo chips. We suspected this might be distorting the real improvement in our gaming test and using the same processor removes that variable.

Memory is now also limited to Corsair DDR3 and DDR2 since there were minor variations in the performance of our two DDR3 memories. This eliminates another potential variable from the benchmarks. All other components are the same as the initial review.

Memory Performance Test Configuration
Processor	Intel Core 2 Duo X6800 (x2, 2.93GHz unlocked, 4MB Unified Cache) 10x266 - 2.66 GHz 8x333 - 2.66GHz
RAM	Corsair CM3X1024-1066C7 (2GB Kit - 2x1GB) Corsair Dominator CM2X1024-8888C4 (2GB Kit - 2x1GB)
Hard Drive	Samsung 250GB SATA2 enabled (8MB Buffer)
System Platform Drivers	Intel - 8.3.0.1013
Video Card	Leadtek WinFast 7950GT 256MB
Video Drivers	NVIDIA 93.71
CPU Cooling	Intel Retail HSF
Power Supply	Corsair HX620W
Motherboards	Asus P5K3 Deluxe (Intel P35 DDR3) Asus P5K Deluxe (Intel P35 DDR2) Asus P5B Deluxe (Intel P965 DDR2)
Operating System	Windows XP Professional SP2
Bench Software	SiSoft Sandra XI SP2 CPU-Z 1.40 Everest 4.0 Super Pi 1.5 Far Cry - River Demo

Benchmarks with the P965 at 1333 FSB were not really of much use in our testing, since the memory ratios would not also allow standard memory speeds of 800 and 1066 when the bus was set to 1333. Also none of the DDR2 memory we had would run at 1333 on the P965. The P965 Is not really designed for a stock 1333 FSB so it was only tested at a 1066 FSB.

Since the P965 was not well suited for 1333 testing, the P35 was adjusted down to 1066FSB for comparison of memory performance. With that setting performance could be tested at DDR2-800 and DDR2-1066. However, a 1333 memory ratio cannot be selected with a 1066 FSB on the P35 boards that were tested, so it was not possible to run memory at 1333 and the CPU at 1066. That is the reason there are no 1333 memory speeds reported under the 10x266 CPU timings.



Bandwidth and Memory Scaling

The results chart was greatly expanded for these benchmarks. Instead of grouping together similar timings for easier comparison we have listed every timing used for DDR3 and DDR2 benchmarking. The bold values are the best performing results for that memory speed. We did take a bit of liberty here since the 800 5-6-6 and 6-6-6 timings are considered together, as are the 1066 5-6-6, 6-6-6, and 7-7-7 timings. If that is too broad a consideration for your liking the data is here for you to interpret how you see fit.

The percentages in parentheses after the upper values compare that result to the performance of the equivalent setting on the P965 ASUS P5B Deluxe. This is compared on P965, P35 DDR2, and P35 DDR3 at 1066FSB. As explained on the previous page, 1333 memory speed could not be tested at 1066 CPU FSB which is why that result is missing from the 1066 FSB list.

Results are also present for the P35 boards at 1333 processor bus at the same CPU speed (both at 2.66 GHz as detailed). Again, the limitations of the P965 did not allow us to run comparable 1333 FSB results since the memory ratios are not available for a 1333 FSB CPU.

Standard (Buffered) Sandra XI.SP2 Memory Bandwidth - 2.66GHz
Memory Speed	P965 ASUS P5B Dlx	P35 DDR2 ASUS P5K Dlx	P35 DDR3 ASUS P5K3 Dlx
10 x 266 - 1066FSB - 2.66GHz
DDR2-800 3-3-3-9	5531	5754 (+4.0%)	-
DDR2-800 5-6-6-15	5207	-	-
DDR2-800 6-6-6-15	-	5545 (+6.5%)	-
DDR3-800 6-6-6-15	-	-	5451 (+4.7%)
DDR2-1067 4-4-3-11	5782	6037 (+4.4%)	-
DDR2-1067 5-6-6-15	5712	-	-
DDR2-1067 6-6-6-15	-	5872 (+2.8%)	-
DDR3-1067 7-7-7-20	-	-	5843 (+2.3%)
8x333 - 1333FSB - 2.66GHz
DDR2-800 3-3-3-9	-	6456 (+16.7%)	-
DDR2-800 6-6-6-15	-	6143 (+18.0%)	-
DDR3-800 6-6-6-15	-	-	6156 (+18.2%)
DDR2-1067 4-4-3-11	-	6811 (+17.8%)	-
DDR2-1067 6-6-6-15	-	6621 (+15.9%)	-
DDR3-1067 7-7-7-20	-	-	6613 (+15.8%)
DDR3-1333 9-9-9-25	-	-	6757

The 10x266 results are roughly representative of the improvement in memory bandwidth that can be attributed to the memory controller of the P35. Considering the variables in timing, this ranges from around 3% to 6.5%. Since the DDR-800 results are the most comparable across all the memory types, and that ranges from 4.7% to 6.5% improved bandwidth, we would conclude that the contribution of the P35 memory controller to improved memory bandwidth is roughly +5%.

The 2.66GHz 8x333 results show the total impact of the improved P35 memory controller and the increase in processor bus to 1333. As you can see the total improvement in memory bandwidth is 16% to 18%. This means the increase in processor bus speed to 1333 is contributing about 11% to 13% to the bandwidth increase, since we concluded the impact of the memory controller alone to be about 5%.

Those of you who are now saying, "Just as I suspected, it's just the 1333 processor bus that is boosting performance" may want to reserve judgment until the game tests. You will likely be surprised.

We normally also test memory with buffering schemes like MMX, SSE, SSE2, SSE3, etc, turned off. While these features do provide apparent improved bandwidth, we have found the unbuffered bandwidth to correlate better with real-world application performance. Unbuffered performance does not always follow the patterns of buffered memory performance.

Unbuffered Sandra XI.SP2 Memory Bandwidth - 2.66GHz
Memory Speed	P965 ASUS P5B Dlx	P35 DDR2 ASUS P5K Dlx	P35 DDR3 ASUS P5K3 Dlx
10 x 266 - 1066 FSB - 2.66GHz
DDR2-800 3-3-3-9	4226	4370 (+3.4%)	-
DDR2-800 5-6-6-15	3668		-
DDR2-800 6-6-6-15	-	3853 (+5.0%)	-
DDR3-800 6-6-6-15	-	-	3988 (+8.7%)
DDR2-1067 4-4-3-11	4608	4706 (+2.1%)	-
DDR2-1067 5-6-6-15	4389	-	-
DDR2-1067 6-6-6-15	-	4484 (+2.2%)
DDR3-1067 7-7-7-20	-	-	4311 -1.8%)
8x333 - 1333 FSB - 2.66GHz
DDR2-800 3-3-3-9	-	4536 (+7.3%)	-
DDR2-800 6-6-6-15	-	3975 (+8.4%)	-
DDR3-800 6-6-6-15	-	-	4098 (+11.7%)
DDR2-1067 4-4-3-11	-	4926 (+6.9%)	-
DDR2-1067 6-6-6-15	-	4557 (+3.8%)	-
DDR3-1067 7-7-7-20	-	-	4547 (+3.6%)
DDR3-1333 9-9-9-25	-	-	4702

The expanded unbuffered results are somewhat different than the unbuffered results in our earlier review. The memory controller contributes from 2% to 8.7% of the memory bandwidth improvement. With our target of the DDR-800 results as most comparable the contribution to improved unbuffered bandwidth is 5%, which is about the same as the buffered contribution. This is pretty unusual, since as you can see below the 1333 line, unbuffered bandwidth improvement is 4% to about 12%, which is lower than the buffered total improvement. For buffered memory bandwidth the memory controller is contributing about 5% of the improvement and the increased processor bus is adding 0% to 6%. Since unbuffered results normally correlate well with gaming performance this hints at the somewhat unexpected gaming results.



Number Crunching and Gaming

The only real improvement in our tests is a 16% to 18% improvement in memory bandwidth on P35, whether it is running DDR2 or DDR3. Does the memory bandwidth improvement translate into any real improvement in system performance? Where there are performance improvements are they likely attributable to the improved memory controller or the faster CPU bus?

SuperPi 1.5 at a 2M setting was run in all memory and speed configurations. This utility benchmarks system performance in pure number crunching as it calculates pi to the number of decimal points selected (in this case 2 million). SuperPi is a very simple program but due to the size of the data set memory performance is often a critical factor.

Super Pi 1.5 - 2.66GHz Time in Seconds - Lower is Better
Memory Speed	P965 ASUS P5B Dlx	P35 DDR2 ASUS P5K Dlx	P35 DDR3 ASUS P5K3 Dlx
10 x 266 - 1066 FSB - 2.66GHz
DDR2-800 3-3-3-9	46.05	45.59 (-1.0%)	-
DDR2-800 5-6-6-15	47.28		-
DDR2-800 6-6-6-15	-	46.77 (-1.1%)	-
DDR3-800 6-6-6-15	-	-	46.95 (-0.7%)
DDR2-1067 4-4-3-11	45.39	45.03 (-0.8%)	-
DDR2-1067 5-6-6-15	45.72	-	-
DDR2-1067 6-6-6-15	-	45.16 (-1.2%)
DDR3-1067 7-7-7-20	-	-	46.03 (+0.7%)
8x333 - 1333 FSB - 2.66GHz
DDR2-800 3-3-3-9	-	45.78 (-0.6%)	-
DDR2-800 6-6-6-15	-	46.08 (-2.5%)	-
DDR3-800 6-6-6-15	-	-	46.89 (-0.8%)
DDR2-1067 4-4-3-11	-	45.20 (-0.4%)	-
DDR2-1067 6-6-6-15	-	45.81 (-0.2%)	-
DDR3-1067 7-7-7-20	-	-	45.93 (-0.5%)
DDR3-1333 9-9-9-25	-	-	45.77

With all tests run with the same X6800 CPU at 10x266 and 8x333 processor speeds, there was no clear performance improvement in SuperPi. It is interesting that most of the best (lowest) SuperPi results were with the P35 at 1066 processor bus. In most cases SuperPi performed better at 10x266 2.66 than at 8x333 2.66 which was something of a surprise. SuperPi performance is a very slightly better as a result of the P35 controller. The 1333 performance is still a bit faster than the P965, but the faster processor bus seems to actually degrade SuperPi performance a tiny amount.

Gaming

A standardized game benchmark was chosen from our memory test suite to determine if the better P35 memory bandwidth improved gaming performance. The Far Cry - River demo was run for 3 loops and results in fps were averaged over the 3 runs. This benchmark was chosen because we are familiar with how it behaves in memory performance tests. A full suite of gaming tests will be run in the P35 performance launch on May 21st.

Far Cry - HOC River Time in Frames Per Second - Higher is Better
Memory Speed	P965 ASUS P5B Dlx	P35 DDR2 ASUS P5K Dlx	P35 DDR3 ASUS P5K3 Dlx
10 x 266 - 1066 FSB - 2.66GHz
DDR2-800 3-3-3-9	101.26	106.01 (+4.7%)	-
DDR2-800 5-6-6-15	97.76		-
DDR2-800 6-6-6-15	-	102.80 (+5.2%)	-
DDR3-800 6-6-6-15	-	-	102.29 (+4.6%)
DDR2-1067 4-4-3-11	103.04	107.65 (+4.5%)	-
DDR2-1067 5-6-6-15	102	-	-
DDR2-1067 6-6-6-15	-	106.06 (+4.0%)
DDR3-1067 7-7-7-20	-	-	104.62 (2.6%)
8x333 - 1333 FSB - 2.66GHz
DDR2-800 3-3-3-9	-	106.30 (+5.0%)	-
DDR2-800 6-6-6-15	-	103.01 (+5.4%)	-
DDR3-800 6-6-6-15	-	-	103.39 (+5.7%)
DDR2-1067 4-4-3-11	-	108.00 (+4.8%)	-
DDR2-1067 6-6-6-15	-	106.61 (+4.5%)	-
DDR3-1067 7-7-7-20	-	-	105.87 (+3.8%)
DDR3-1333 9-9-9-25	-	-	106.7

Using the same CPU for all tests actually shows greater improvements in gaming performance than was seen in the DDR3 vs. DDR2 review. As you can see in the comparisons above the P35 memory controller contributed 3 to 5% improvement in gaming performance. Our target DDR-800 results showed a 4.6% to 5.2% improvement, so we conclude the P35 memory controller improves gaming performance roughly 5%

Results below the 1333 line represent total improvement, which varied from about 4% to 6%. With the target DDR-800 performance improvement ranging from 5.4% to 5.7% total, and the memory controller alone contributing 5% improvement, we are left with about 0.5% performance improvement in gaming due to the higher processor bus.

These results were a bit of a surprise. Almost all of the improvement in real-world gaming came as a result of the improved P35 memory controller, with about half a percent attributable to the 1333 processor bus. The Far Cry benchmarks show most of the real world gaming improvements are the result of the P35 memory controller.



Final Words

Many readers raised the valid concerns of whether the improvement in Bearlake (P35) memory performance was the result of the improved memory controller or whether it was merely a side effect of the faster 1333 processor bus. We tried to answer this in comments to the original review, but the question deserved additional testing to clearly show where the performance improvements are originating.

In the area of memory bandwidth improvement, it appears that the memory controller is roughly adding about 5% to P965 performance while the faster 1333 processor bus contributes an additional 11% to 13%. The total improvement in memory bandwidth for P35 compared to the P965 is 16% to 18%. Memory is just one small part of overall system performance, so this does not translate into a 16% improvement in system performance. System performance is improved, but real-world performance improvements are smaller.

In pure number crunching it appears that the P35 only slightly improves performance, in the 0.5% to 2.5% range. It is interesting that almost all of the SuperPi improved performance is a result of the improved P35 memory controller, since the higher 1333 processor bus actually decreases performance a very small amount compared to the 1066 bus.

Gaming performance with Far Cry benchmarks was found to improve 3% to 6% on the P35. About 5% of that increase was due to the improved P35 memory controller, with the higher 1333 bus only adding about 0.5% to the performance.

The memory performance improvements for P35, as stated in the initial review, are pretty impressive for just a chipset upgrade. The memory bandwidth improves 16% to 18%, with about 5% of that increase due to an improved memory controller. The rest is the result of the higher 1333 processor bus. SuperPi performance improves slightly - 0.5% to 2.5% - almost all due to the improved P35 memory controller and gaming performance improves about 5.5%, again almost entirely the result of the improved memory controller.

With this closer look at the components of the P35 improvements in memory performance, we can put to rest the either/or vein of questions about the increase in processor bus. The improvement in P35 memory performance is a result of a combination of an improved memory controller and a faster processor bus. It is not the result of either factor alone. In some cases, like memory bandwidth, the processor bus is the larger contributor. In others, like gaming, the improved memory controller is the bigger factor.

Regardless of which aspect improves memory performance the most, however, it is clear the P35 is a better performing chipset than the earlier P965. If you are in the market for a new system, the P35 is a good chipset for building a performance system. This is particularly true if you intend to use a 1066 FSB processor with the "free" 25% overclock that is possible by just setting 1333 instead of 1066 in the motherboard BIOS.