Company of Heroes Complete Edition (2009) benchmark with Turion 64 at 2400Mhz and GTX 275:

4x AA was used, the rest is maximum quality. Some options allow Ultra quality that cannot be selected (Shader Quality, Terrain Detail). It can probably be enabled only with dx10. Environmental reverb cannot be enabled.

Low of 3.9fps is probably a benchmark glitch. The value is not reliable.

The binned and OCed registered ECC memory does seem to have a small edge in games, average fps is better in all games over 2x unbuffered sticks. It was much less convincing in 3d mark, pc mark and cinebench.

2x unbuffered 1GB DDR 400 sticks at CR1 had very little OC headroom, it was a big problem to stabilize CR1 in the first place. Running them with CR1 at DDR480 is sci-fi. They may OC well with just 1 stick, but then you are limited to 1GB RAM and this is only suitable for Windows 98. In this case you would probably do better with single rank stick of 512MB. For Windows XP you really need at least 2 GB ram.

Based on my results with binning of 2GB sticks, I would recommend to buy 4x more sticks than you intend to use. Only buy them if you find them for a fraction of ebay price and only from a seller who has 100s of them.

I only used default voltage, I do not overvolt anything. Lancaster supports up to DDR 500 in BIOS.

I already have the MSI K8N NEO Platinum, but I am keen to try the Nforce4 - MSI K8N Neo3 (MS-7135) - PCI-E + AGP. Any review from Vogons for this board?

I checked manual of MSI K8N Neo3 and it looks like a decent choice. It has detailed memory timing configuration, multiplier and voltage settings in Cell menu. Let us know if it supports registered ECC memory. The best CPU for it is Turion 64 ML-34 / 37 OCed to 2.2Ghz. As GPU use GTX 260/275/285.

I'll try to grab it on auction if it won't cost too much. Just for s*it and giggles...

I do not think you got "real"AGP on that board! It is just plain PCI! I would still get it because nforce4 is the best platform for socket 754 and you can use cheap PCI-express cards.

Yes, my point was to have PCIe on 754 for a few bucks and to try Nforce 4 on 754. For AGP I already have an nForce 3 250Gb. I saw AMA3400BEXSAR for cheap (10€), so maybe that can be a funny combination with 3850 or something like that?

You can buy it. It is a mobile Clawhammer 3400+ at 2.2Ghz. https://www.cpu-world.com/CPUs/K8/AMD-Mobile% … 3400BEX5AR.html

It is AMA3400BEX5AR, not AMA3400BEXSAR

Attaching aida64 memory benchmark to show there is no noticeable performance drop with ECC on (without scrubbers) or off. The differences are within margin of error. Having ECC for memory OC is a significant advantage as it will either correct the error or stop the machine to prevent data corruption. We saw how elusive Command Rate 1T was with 2 unbuffered sticks and it required many hours of testing and multiple testing programs. With ECC memory you should memtest with ECC on and off.

I got to test upper limit temporarily, to make sure things wouldn't get unstable if base frequency was increased by 1Mhz. It would not be good to be in small island of stability.

When 273 base frequency was selected in BIOS, memory was still stable (ecc was disabled for memtest), memory benchmark is attached. We are getting into very low latency territory, just 58.8ns. Memtest86 reported errors with 278 base frequency. I didn't try increasing voltage. Normally I use base frequency 267 in BIOS. It may be possible to get to 58ns or slightly less by further binning or more voltage.

I also retested my other Qimonda sticks at DDR480 and current timings and found out 1 other stick also passes memtest. This would make OC success rate about 50%, but I would still recommend buying 4x more with small number of sticks, or 3x with very large number of sticks.

Temperature wise at DDR480 memory chips feel warm, not lukewarm and not painfully hot. I would estimate the temperature to be about 42'C. I do have plenty of airflow in the case as it's a mesh case. If you increase voltage or have less airflow use a heatspreader. The cheapest way to get heatspreaders is probably by buying super cheap FB-DIMM memory. Nobody will ever use that memory again as it is server specific. You can get it for like 1€ per stick.

New batch of cheap Qimonda 2GB registered ECC sticks arrived for binning so I got to test them out.

2 run at 267Mhz (DDR 533). One appears to have faulty ECC mode (even at DDR400), so functions only as a registered stick. No point in returning it as it clocks very high with tight timings and that is more important than the ECC function. The other stick functions in ECC mode as well. When enabling of ECC mode fails, I get 1 long and 2 short beeps. Both were memtested in single pass, 2nd stick also in ECC mode in addition to non-ECC mode. More testing will be done.

To ensure we are not in a tiny island of stability, I memtested it also at 273Mhz (DDR 546) and it passed. All using default 2.5V voltage of course, nothing is overvolted. I checked temperature of the sticks and it's about 42'C as previously. No need for a heatspreader in my case. At 300Mhz (DDR 600) both sticks no longer boot even with conservative timings.

I ran Aida64 memory benchmark at 267Mhz (DDR 533) in both non-ECC and ECC mode, no difference. We get about 4180MB/s read and write speed and 54.2ns memory latency which is excellent. Latency is lower than what I had with 2 unbuffered sticks (about 57ns). We get lower L2 cache latency as well.

At 273Mhz (DDR 546), we get insanely low latency 52.9ns

When it comes to timings, I used the same timings as for 240Mhz (DDR 480) previously.

DDR 533 should negate the bandwidth advantage s939 enjoys. Because we have just one stick, it clocks very high. I wouldn't expect two sticks in s939 to run at DDR 533/546, 4 sticks will probably top at about DDR 450. I have noticed it in AM3 as well with unbuffered ECC sticks. One 8GB stick works at DDR3 1600 and adding another one causes trouble and requires bumping voltage by 0.1V.

I also tested the 2x 1GB Micron registered DDR333 that I have but neither of them works at 267Mhz (DDR 533) at the same timings as Qimonda. My binning revealed there can be vast differences between individual sticks of the same model, some don't like higher frequencies, some don't like tight timings, some don't like Command Rate 1T and even ECC mode can be faulty. Therefore, if you buy them, you need to get them really cheap, ideally from a local store that collects server memory, who don't speak English and normally don't ship abroad. Do not consider eBay, they are too expensive there to make it practical. You should buy 4x more sticks than you need for binning purposes. If you want exceptional sticks then possibly even 8x more. For s939 that means 16-32 sticks, a big amount. Buying 1-2 sticks is not going to work and you will be disappointed.

AlexZ wrote on Yesterday, 07:32:

I also tested the 2x 1GB Micron registered DDR333 that I have but neither of them works at 267Mhz (DDR 533) at the same timings as Qimonda. My binning revealed there can be vast differences between individual sticks of the same model, some don't like higher frequencies, some don't like tight timings, some don't like Command Rate 1T and even ECC mode can be faulty. Therefore, if you buy them, you need to get them really cheap, ideally from a local store that collects server memory, who don't speak English and normally don't ship abroad. Do not consider eBay, they are too expensive there to make it practical. You should buy 4x more sticks than you need for binning purposes. If you want exceptional sticks then possibly even 8x more. For s939 that means 16-32 sticks, a big amount. Buying 1-2 sticks is not going to work and you will be disappointed.

Have you tried flashing the ram to think it's a smaller capacity? I just mention that because I must have cleared out all my smaller better clocking sticks out a long time ago. Maybe If a 1gb thinks its a 512 it might clock better with tighter timings?

I do not want smaller capacity, I want 2GB. I bought those 2 Micron sticks cheap only to test whether the board and CPU boots using registered DDR. SPD settings are not relevant for timings, I use manual settings. I will check using SPD Tool what is in those sticks. The only use could be perhaps to disable ECC function in that stick that fails if it is enabled so that the ECC option isn't shown in BIOS at all. It isn't easy to recover from that problem.

AlexZ wrote on Yesterday, 21:30:

I do not want smaller capacity, I want 2GB. I bought those 2 Micron sticks cheap only to test whether the board and CPU boots using registered DDR. SPD settings are not relevant for timings, I use manual settings. I will check using SPD Tool what is in those sticks. The only use could be perhaps to disable ECC function in that stick that fails if it is enabled so that the ECC option isn't shown in BIOS at all. It isn't easy to recover from that problem.

Just wondering if the ram could be binned a step further if one didn't need the capacity?

Edit: Flashed a 1GB stick of AMP memory into 256MB, and it did allow me to drop down from cas 3 to 2 (Memtested) . Obviously no use to your project of maxing out the 754 but for me I don't need that much ram on a Clawhammer.