Investigations into Socket 939 Athlon 64 Overclocking

Closing Thoughts

This has been a huge undertaking, and we hope that you enjoy the results and overclocking information. Many guides of various forms already exist on the internet, but we really wanted to take a look at some of the options offered by the Venice core processors as well as trying to get by with value RAM instead of higher cost alternatives. After all, why save $60 on the CPU only to spend the same amount of money upgrading the RAM? So, other than a look at how to overclock, what have we learned?

The biggest deal for many people will be the results using standard PC-3200 RAM and higher memory dividers. While you won't be able to match the performance of a system that uses better quality RAM, the largest margin of victory was still under 10% (not counting instances where 2T command rates were used). The average was closer to 5%, and realistically, you won't notice a 5% performance loss. Outside of games, the performance losses attributed to value RAM are even less, with video encoding only losing a few percentages in speed. What it really comes down to is cost. We used a $190 CPU with $85 and $150 RAM. Going with a 3000+ and the value RAM saves almost $125 and should get you about 90 to 95% of the performance of the more expensive setup. That $125 could then be put towards a faster GPU, which will have a far greater impact on games than a 200 or even 400 MHz CPU upgrade.

Other than the value vs. quality RAM debate, what about the overall experience of overclocking this configuration? We were pretty impressed with the "budget" DFI motherboard. While I haven't personally used the Ultra-D, there were few problems on the Infinity that I couldn't deal with. Even with a good motherboard, though, overclocking can be exasperating at times. With value RAM rated at DDR400 2.5-3-3-8, it was at times difficult to get those timings even at speeds slightly below the rated DDR400, particularly on the high end of the overclocking scale. While I'd be willing to run 2.6 or even 2.7 GHz with the OCZ RAM, I have a feeling that long-term stability with the value RAM might require dropping to 2.5 to 2.6 GHz instead, or else increasing the timings to 3-3-3-8 or even 3-4-4-8. And speaking of timings, the 2T command rate should almost always be avoided. Results for a couple of settings were included, and overall, you would be better off running 100 or even 200 MHz slower with 1T command rate.

Something else that all of the graphs and results don't indicate is some of the oddities that can come up with overclocking. The SATA drive would sometimes make strange noises during the Windows XP boot sequence, almost like the HDD heads were seeking back and forth across the drive. Everything seems to indicate that the overclock is somehow to blame, and while a BIOS update might be able to address this particular issue, the end result is that XP would sometimes take up to 5 minutes to load at higher overclocks. Once loaded, everything worked fine, and the HDD was still running at full SATA spec. Another possible cause for delays in loading XP could be the networking subsystem. The testbed was connected to a gigabit switch, and we've experienced issues with network stability on overclocked PCs in the past. An MSI K8N Neo Platinum, for example, drops network connectivity after a day or two in many - but not all - overclocked configurations. We didn't experience this particular problem during testing of the DFI board, but it's something to look for on your own systems. Dropping the speed of the HyperTransport bus also helped avoid some - but not all - of the hard drive access delays. As we said, be prepared for some strange behavior now and then during overclocking.

The final comment that we want to make is about the long term viability of overclocking. We started this article with a warning, and we'll end it the same way. While we haven't encountered problems with the CPU yet, that doesn't mean that the chip won't simply die in a few weeks, months, or hopefully, not for years. Higher voltages in particular can affect CPU life, as they can accelerate electron drift. As we couldn't get to 2.7 GHz without running at 1.750V, we're a little hesitant to recommend that speed as a long-term solution. Given that 2.6 GHz is, at worst, only 4% slower, we'd recommend that as a better solution and go with the 1.650V setting. That's similar to how Intel supposedly binned CPUs back in the socket 7 days: they would reportedly increase CPU clock speeds until the chips failed, and then sell them two bins below the maximum stable clock speed. Whether that's rumor or in truth how they operate (operated?), running at speeds slightly slower than your "stable" maximum will be preferred by many. Crashing even every couple of days or once a week is too reminiscent of the Windows 95 era.

We've tried to get across the point that there are no guarantees with overclocking. Even with that disclaimer, we're pretty confident that the vast majority of Athlon 64 Venice chips will run at 2.4 GHz, and probably even 2.6 GHz. It may require higher voltages, better cooling, or relaxed memory timings, but with the right combination of parts, it's a relatively safe bet. Worst case scenario, try running at 3-4-4-9-2T memory timings, then try running at those timings and PC2700 or even PC2100 on the RAM. If it's still unstable, it might be your motherboard or some other factor holding you back. Even a 15% overclock is still pretty good, though, and you can probably get that without any special equipment other than an enthusiast motherboard.

That closes up this overclocking article. We have several similar articles planned, though we're interested in feedback from the readers. Was this too superficial? Do you want more details on tweaking memory timings beyond what we've mentioned? Or is the mix of benchmarks, settings, and results about right? Let us know. This article was long, with a large portion dedicated to introducing the uninitiated to the art and practice of overclocking. Future articles in this series will focus more on the end results and refer back to the concepts presented here. As always, any recommendations and comments are welcome.