Author: Pieter-Jan Plaisier
What is HWBOT Research Lab?
As you all know HWBOT has changed since October ‘09 as we now have two people working full time on the site. This allowed us to do the necessary improvements to the HWBOT website’s looks and features of which you can see the result by browsing around the front-page. The development phase of the bot required more time than we originally anticipated as in the past six months we’ve spent most of our day and night-time to the project. In the next months and years, we will continue to expand the HWBOT front-page and features, but next to that we also want to take on different other challenges. For instance, using the data available in the HWBOT database (thanks to all of our beloved members) to produce more valuable performance and overclocking charts. In addition, we’d also like to spend more time on the science part of overclocking and translate our findings to what users find interesting. Overclocking performance scaling, guides for a better understanding of overclocking, power consumption measurements and so on.
These projects will fall under the (currently-still-beta) HRL, or: HWBOT Research Lab, tag. It’s simply a name to separate the normal news and articles from the more overclocking research minded articles. Do note that the name can still change and that the exact fields of the HRL have yet to be determined. It’s a beta-project.
Introduction to this article
About a week ago, on the day Thuban was released publically, we presented you our findings on the overclocking capabilities of this new hexacore architecture. As you might have read, the current generation of X6 processors are not yet fully at the level of the Deneb C3 silicon revision, simply meaning they don’t clock as high just yet. The clock frequencies reached on liquid nitrogen were disappointing in some way, although the results in the temperature range of +40°C to -40°C were quite impressive, leaving the Deneb C3 behind by a couple of 100MHz. Currently, we see that the 1090T’s are reaching up to 6.7GHz suicide, 6.2GHz fully benchmark stable and almost 7GHz using Liquid Helium!
Another aspect of the new Thuban is the (advertised) improved memory overclockability. As many AMD overclockers know: neither the Deneb C2 nor C3 were very good at overclocking the memory beyond DDR3-1900, especially if stability was required. Mainboard manufacturers have learned from the Deneb overclocking problems and with the release of a new chipset, the 890FX, most of them have designed all sorts of fixes and tweaks for the new PCBs in order to improve the memory overclocking. That combined with the improved Thuban silicon could mean new heights in memory overclocking.
When I say ‘could’ I’m saying that the improved overclockability is a fact. The days of struggling below DDR3-1800 are over, at least if you have the right mainboard. As I’m writing this, we’ve seen results close to DDR3-2100 CL7-7-7-21 32M stable, which is far from non-impressive. We expect clock frequencies to rise even higher with new mainboards and bioses being released. All this MHz-madness is fun, agreed, but it does bring up a few questions. Is it really worth to have such high memory clocks? How does the memory controller frequency relate to all this? When does overclocking become … pointless?
The answer to that last question is of course “never”, but the two other questions are a bit trickier. When browsing through overclocking forums, I notice that people have different opinions on how interesting overclocking really is. As far as I’m entitled to do so, I would split up two different branches: on the one hand there’s the theoretical scaling limitation and on the other hand the practical side of the performance scaling. The latter is what most people are looking for, without any doubt, as that’s where other variables such as voltage and cost form a bottleneck.
In this article, we’ll be having a look at the memory subsystem performance scaling. Important to note is that this (short) article only deals with the theoretical and practical performance gains coming from overclocking and will not deal with voltage, cooling and other practical issues.
Page 1 – Introduction
Page 2 – Test setup and methodology
Page 3 – Theoretical Gain – DRAM bandwidth figures
Page 4 – Theoretical Gain – L3 cache bandwidth figures
Page 5 – Practical Gain – WinRAR
Page 6 – Practical Gain – TechARP X264 HD Benchmark
Page 7 – Conclusive thoughts