Athlon 64 Memory: Rewriting the Rules
by Wesley Fink on October 1, 2004 12:45 AM EST- Posted in
- Memory
Performance Test Configuration
The first tests with the new Athlon 64 test bed were in the recent OCZ PC3700 Gold Rev. 3 review. In this round, we extend the Athlon 64 memory tests to some of the fastest memory that we have tested and the newest memory in our labs. Since we have found DDR memory to perform very differently on the memory controller with Athlon 64 chips, we will be including Athlon 64 benchmarks in all future memory reviews.The A64 test bed includes components that have been proven in Socket 939 Athlon 64 benchmarking, such as the Gold Editors Choice MSI K8N Neo2, the completely unlocked Socket 939 FX53, and the OCZ Power Stream 520 Power Supply. Since the Athlon 64 tests represent a new series of DDR testing, we have chosen the current generation nVidia 6800 Ultra video card for benchmarking. We have found the 6800 Ultra to be a particularly good match to nVidia nForce3 Ultra motherboards.
All other basic test conditions attempted to mirror those used in our earlier Intel memory reviews. However, test results are not directly comparable to tests performed on the Intel test bed.
| AMD nForce3 Ultra Performance Test Configuration | |
| Processor(s): | AMD FX53 Athlon 64 (2.4GHz, Socket 939, Dual Channel, 1000HT) |
| RAM: | 2 X 512MB Crucial Ballistix (DS) 2 X 512MB Geil PC3200 Ultra X (DS) 2 X 512MB G. Skill TCCD (DS) 2 X 512MB OCZ PC3200 Platinum Rev 2 (DS) 2 X 512MB OCZ PC3700 Gold Rev 3 (DS) 2 X 512MB PQI 3200 Turbo (DS) |
| Hard Drives | Seagate 120GB PATA (IDE) 7200RPM 8MB Cache |
| PCI/AGP Speed | Fixed at 33/66 |
| Bus Master Drivers: | nVidia nForce Platform Driver 4.24 (5-10-2004) |
| Video Card(s): | nVidia 6800 Ultra 256MB, 256MB aperture, 1024x768x32 |
| Video Drivers: | nVidia Forceware 61.77 |
| Power Supply: | OCZ Power Stream 520W |
| Operating System(s): | Windows XP Professional SP1 |
| Motherboards: | MSI K8N Neo2 |
We have found the fastest performance on AMD Athlon 64 chipsets (nForce3, VIA K8T800 PRO) to be achieved at Cycle Time or tRAS of 10. Athlon 64 platform benchmarks were, therefore, run with the tRAS timing of 10 for all A64 benchmarks.
Test Settings
The FX53 is completely unlocked, something not currently available with Intel processors. This allowed a different approach to memory testing, which truly measures performance differences in memory speed alone. All tests were run with CPU speed as close to the specified 2.4GHz of the FX53 as possible, with CPU speed/Memory Speed increased at lower multipliers to achieve 2.4Ghz. This approach allows the true measurement of the impact of higher memory speed and timings on performance, since CPU speed is fixed, removing CPU speed as a factor in memory performance.The following settings were tested with the six memories on the Athlon 64 test bed:
- 12x200/DDR400 - the highest stock memory speed supported on K8T800/nF3/SiS755 motherboards.
- 11x218/DDR436 - a ratio near the standard DDR433 speed.
- 10x240/DDR480 - a ratio near the standard rating of DDR466.
- 9x267/DDR533 - a standard memory speed used in testing other high-speed memory.
- Highest Memory Speed - the highest memory speed that we could achieve regardless of the multiplier. This setting was generally achieved at a 2T command rate and performance is often poorer than slower memory timings at a 1T Command Rate.
- Highest Performance - the highest memory performance settings that we could achieve. This setting is normally the highest stable speed using a 1T Command Rate.
Command Rate is not normally a factor in Intel 478 tests, but it is a major concern in Athlon 64 performance. A Command Rate of 1T is considerably faster on Athlon 64 than a 2T Command Rate. For this reason, we had added the Command Rate to the timings and voltage reported for each memory speed setting.
We ran our standard suite of memory performance benchmarks - Quake 3, Super Pi 2M, and Sandra 2004 UnBuffered. Since the results for Athlon 64 tests are new, we are now including Sandra Buffered (Standard) test results as well as Sandra UnBuffered test results. Return to Castle Wolfenstein-Enemy Territory has also been added as a standard memory benchmark.
Facebook
Twitter
RSS
47 Comments
View All Comments
saechaka - Friday, October 1, 2004 - link
i can't seem to find a legit place to buy that ocz 3200 rev. 2. any suggestionsAvalon - Friday, October 1, 2004 - link
Excellent article. It's good to know what different memories can do on the Athlon 64 platform.ramclocker - Friday, October 1, 2004 - link
14The psu is probably around 20A on the 12V...I know from my testing 20A doesn't cut it anymore on a high end gaming/benching rig....you also have to remember that at high speed the ram will be drawing high levels of current also and the board will draw higher current due to heat etc.
I found the article an excellent read due to the fact it finally proved to me with reasonably tight timings running high fsb over 2-2-2 at 200 is the way to go...running 2.4gig for all tests Wes was the wise move here...great work.
Blappo - Friday, October 1, 2004 - link
The computer would probably wouldn't use more than 250W. I understand that you don't want to mention the make and model. The nVidia 6800 Ultra draws most of its power through the 12V connection to the PSU, where the ATI 9800 Pro draws its power from the AGP slot. What is the max current rating on the 12V rail for the 465W PSU that you were using? I agree that a high quality PSU is needed (although not necessarily high max rating).Wesley Fink - Friday, October 1, 2004 - link
#12 - The Asus A8V is reviewed at http://www.anandtech.com/mb/showdoc.aspx?i=2128 and compared to other 939 boards. In memory testing we use a standard test bed to minimize variables.#11 - The 90nm Athlon 64 tests should appear next week. We have just received 90nm 3500+ and 3000+ processors. AMD did not do a media launch on these processors, so we had to find them on the open market
#9 and #11 - A Value RAM roundup is in the works, but it has been moved out a while because of a large number of new launches this month.
Deuce - Friday, October 1, 2004 - link
It sure would be nice with tests also conducted on the Asus A8V. I'm still deciding between the two.PrinceGaz - Friday, October 1, 2004 - link
Just to follow up that comment, I suppose the DDR533/2.4GHz results are actually the most useful out of them all when it comes to comparing those particular modules. All of them were fastest (at 2.4GHz) at that speed, except for the OCZ PC3200 Plat Rev.2 which was marginally faster at 8x300 for DDR600.Anyway, I'm looking forward to the reviews of the desktop 90nm A64 processors, and especially finding out how well each of them overclocks.
And also the Value Memory review you promised a few weeks ago :)
PrinceGaz - Friday, October 1, 2004 - link
Although all the (admittedly premium quality)memory could reach very high speeds, that didn't have much impact on performance.Taking the highest clocking brand as an example, the OCZ PC3200 Platinum Rev.2, from the DDR400 2-2-2 speed to the DDR534 2.5-4-3 speed which was the best result still at a CPU speed of 2.4GHz, the results were
Quake 3: 516.3 -> 525.8
Super PI: 80 -> 79 (lower is better)
Wolf: 110.8 -> 112.7
So running the memory at DDR534 instead of DDR400 provided less than 2% increase in performance. This is to be expected when you compare the real-world performance of S754 and S939. The only thing that is important is that the memory can do 1T command-rate to the maximum overclock of your A64 at default multiplier.
I think the results on the highest memory performance page are probably misleading to some readers. It shows the Crucial Ballistix coming in at 536.5fps on Q3. Looking at the results I see that was at 9x278 for a CPU speed of 2.5GHz. Your CPU was able to reach over 2.6GHz so the performance in real world tests would have been somewhat higher with a 10x multiplier. Sandra results are irrelevant to most people.
It would be better if you included an additional test in addition to Highest Memory Speed, and Highest Memory Performance. This would be Highest CPU Speed where the CPU is maxxed out, and the memory run at whatever multiplier gives best performance on real-world tests (ignoring Sandra). I suspect the results would be a *lot* closer.
AtaStrumf - Friday, October 1, 2004 - link
Now about some value RAM tests? These modules are just too expensive for most of us.Jalf - Friday, October 1, 2004 - link
Or maybe the "average" user would rather blow $200 on 1GB memory ;)