AMD Radeon HD 7950 Launch Review

Product: AMD Radeon HD 7950 Video Card
Company: AMD
Author: James Prior
Editor: Charles Oliver
Date: January 31st, 2012

Power, Temps, Overclocking

AMD promised lots of overclocking headroom at the AMD Tahiti tech day. Something they also said that at the Northern Islands and Evergreen presentations before. This time, however, it's a little bit different.

There are three controls to work with for AMD Radeon overclocking - first engine clock, then memory clock, and finally PowerTune level. The engine clock governs the whole GPU clocks, as AMD doesn't run clock islands for their shader cores as NVIDIA does. Memory clock is the base clock rate, which is quadrupled to get the GDDR5's effective data rate. PowerTune level adjusts the point at which the intelligent onboard monitoring intervenes and adjusts core clocks to stay within the chosen TDP. This can be adjust incrementally between -20% and +20%; 180W and 220W respectively, for the Tahiti Pro.

For our testing we adjusted Powertune to +20% - and then closed Overdrive. To adjust clocks we installed the new version of Sapphire TRIXX, which allows a larger range of clock speeds and the ability to specify a custom fan profile. Our engineering sample reference designed HD 7950 has the standard AMD cooler, which is not the best option if you want quiet overclocked performance.

Without adjusting core voltage, we were able to run the Radeon HD 7950 at 1025MHz, an increase of ~28%! The memory, while specified to run at 1250MHz/5Gbps, appears to use the same chips as the HD 7970 (at least, on our reference engineering sample board) and easily hit the HD 7970's 1375MHz specification. Our final memory overclock was 1590MHz/6.36Gbps, a substantial ~27% increase from stock speed. With the default voltage limits reached, we used Sapphire TRIXX's voltage adjustment to give us more headroom. Setting a core voltage of 1.237v, an increase of 147mV over stock, we were able to run an engine clock of 1200MHz, a whopping 50% increase over the stock clocks. For this level of performance we opted to create our own fan profile to manage the thermals of the card.

TRIXX Software

The Tahiti Pro ASIC is a 'binned' version of the full design, with various factors contributing to the final product specifications - not the least of which is fabrication factory yield. The missing CUs could be disabled due to imperfections in the manufacturing process, which allows an otherwise unusable part to be sold and generate revenue. On the Cayman series, AMD surprised everyone by 'permitting' (or, at least looking the other way when it happened) unlocking of the missing elements in the Cayman Pro design. No such luck this time with the Tahiti Pro, which may be an indicator that nonfunctional parts are more prevalent than before (which would stand to reason given the new process and associated problems) or possibly just a change in marketing strategy; unlocking didn't work out the way AMD hoped.

Shroud removed

AMD's new GPU core design is voltage limited in terms of clock speed, and that voltage limit is defined by AMD's aimed for TDP. The dual PCIe 6pin connectors plus mainboard slot offer a total of 225W, under PCIe specification, and AMD has very sensibly set the HD 7950's maximum board TDP at 200W. When the PowerTune limit is raised by 20%, to a 240W cap, there is still electrical headroom in the power provided to the card.

Using a Kill-a-watt meter we measured full system power draw in three conditions; idle, media playback, and load. The idle condition is with display blank enabled, at the Windows 7 desktop. Media playback is recorded during 1080p full screen Blu-ray playback using CyberLink PowerDVD 10 Mark II, with GPU acceleration enabled (ATI AVIVO/NVIDIA CUDA). Load condition is the highest reading seen during 3DMark11 Extreme preset run of GPU test 2.

Additionally we recorded GPU temperatures under these conditions; taking readings after 30mins of the workload condition has elapsed, with the stock fan profile in use, except for overclocked which used a custom fan profile.

We reran 3DMark11 Performance and Xtreme tests to see what improvements the new clock speeds had given:

  HD 7950 HD 7950 OC % Difference
PMarks 6801 9152 +35.5%
XMarks 2249 3128 +39.1%

The overclocked 7950 scores ~15% faster than the stock 7970, and 38% faster than the GTX 580 SC (45% in extreme preset). Will that translate to more gaming performance?