I recently received a couple treats from Gigabyte Japan, namely a pair of new Radeon HD 5870 Super Overclock or “SOC” edition video cards. These cards feature a “binned” GPU, a custom PCB with tons of high-end features, and a custom air cooler. They also come with a factory rating of 950MHz for the GPU and 1250MHz for the GDDR5 memory, a significant boost over the 850/1200MHz speed of the reference cards. I took this beast of a card, and as you might expect, I put it through its paces with a buffet of benchmarks to find the absolute limits of performance.
First, I’d like to point out some of the stuff that makes this card special. If you’d like all the marketing speak, please visit the Gigabyte product page. All manufacturers have pages of reasons why their products are superior, but I am more interested in seeing the actual performance. Afterward, if compelled by the performance, I may be interested in how they were able to achieve it. For the rest of you, I will pick out a couple of things that set this card apart from the rest.
GPU Gauntlet Sorting
The first of which, and possibly the most important in my opinion, is the binned GPU. Gigabyte has a new process they call GPU Gauntlet Sorting. This is basically a process by which they select the GPUs with the highest potential for overclocking. Most impressive about this is the fact that it’s a manual binning process done by real people! So I’m very excited to see how this plays out in my overclocking.
The PCB was designed in-house by Gigabyte with top quality components for overclocking. Taking a look around you can see that no expense was spared. It also has very handy voltage read points on the PCB for overclockers to have a convenient location to monitor voltages in real-time.
Dual Fan Heatpipe Cooler
They’ve also slapped a very interesting air cooler on the card. It has two 80mm fans and four 6mm heat-pipes to draw heat away from the GPU. What I find especially intriguing about the cooler’s design is that it does not make contact with anything on the card except the GPU. There are not any heatsinks anywhere else! A hidden benefit for extreme overclockers – the stock cooler attaches with only 4 screws! Very nice touch Gigabyte!
OC Guru – Graphics Card Smart Tuner
I would also like to mention the “OC Guru – Graphics Card Smart Tuner”. Currently, it is the only software overclocking application that is compatible with the SOC card. It allows for software adjustment of the fan speed, and up to an additional 100mV for both the GPU core and the memory. It also monitors voltage, current, power consumption, GPU temperature, and fan speed.
Overclocking on the Stock Cooler
First step was to figure out what this card is capable of accomplishing with the factory configuration. As most overclockers know, reference 5870 cards are known to be very good overclockers. Most will hit 900MHz with ease, and a few are even capable of 1000MHz with a bit of tweaking and an extra dose of the good stuff, aka “voltage”. So I was really hoping to see that type of result here.
I received two of these SOC cards for testing. However one of them right off the bat was not stable at its default settings. A normal user would call it dead on arrival or “DOA” and send it back for a replacement. Seems like a bit of a failure in the GPU Gauntlet Sorting procedure, since this card would artifact in 2D on the Windows desktop. However, I was able to easily overcome the problem by setting the start up voltage to 1.25V with the OC Guru application. So, although only one card needed it, all of my “stock” testing was actually done with a bit of extra voltage.
Test System Configuration
- eVGA X58 E760 Classified motherboard (single card testing)
- Gigabyte X58A-UD7 (CrossFireX™ testing)
- Intel Core i7 920 – CPU & uncore @ 4.0GHz (custom water cooling)
- 3x1GB Crucial Ballistix DDR3-1600 @ 7-7-7-21 1T
- 2x Gigabyte HD 5870 SOC 950MHz/1250MHz and 1015MHz/1375MHz
- Ultra X3 1000W PSU
- Windows 7 x64 Ultimate SP1 (a custom, stripped down and optimized version)
- ATI Catalyst 10.4 driver package (most recent version available at start of testing)
All my testing was run at 1280×1024 resolutions due to the limits of my monitor. With each of the synthetic benchmarks, I set all driver settings to maximum performance and the benchmarks were run at default settings. With the two DX11 game tests, the drivers were set to maximum quality and the benchmarks were configured to the highest detail settings available. I did this to give both camps (benchers and gamers) the info they are looking for. Each test was run three times, each test was completed with a fresh boot and with the CPU priority set to “high,” Aquamark 3 was also manually configured for affinity on all 8 threads. Scores shown for the synthetic tests represent the best out of three; the scores for the two DX11 game benchmarks are the average of the three.
I started by running through my entire benchmark suite with a single card at its stock clocks & voltages, that is 950MHz core and 1250MHz (5000MHz effective) memory speed at 1.2V and 1.6V respectively. After completing all tests at stock frequencies, I found the maximum speed at which either card would complete all tests. I found it very interesting that both of these cards were able to reach the exact same overclock through this benchmark suite, 1015MHz core at 1.3V and 1375MHz memory at 1.6V. Increasing the memory voltage made very little impact on the memory overclocking capability on either card, both of which were stable at 1375MHz (5500MHz effective) with the standard 1.6V. However, after putting both cards into the machine, the additional heat on the upper card became too much and I was forced to drop the memory frequency down to 1300MHz in order to pass all of the tests.
Overclocking with OC Guru
As I mentioned above, I used the Gigabyte OC Guru software during my testing on air to assist in reaching my maximum overclock. This software works well, and did what I asked of it throughout my single card testing compared to the software program used for most other cards on the market right now, MSI’s Afterburner GPU overclocking utility. MSI has done an amazing thing for the community with their free utility, if you haven’t used it yet, you owe it to yourself to check it out. There was only two things that were really lacking from the OC Guru app though, the monitoring portion needs improvement as there is no way to track the voltage, speeds, temperatures, and fan speeds while benching. There also is no graph or data file that logs this information so it leaves you a bit blind. I used the GPU-Z sensor tab to keep track of these things during my testing. Things were not so pretty however after switching to the CrossFireX™ setup, see more below.
I wanted to add a second card to the mix to find out how much I could improve the results. But there I ran into a problem with my eVGA motherboard. I could not get the dual card setup to run at x16 by x16. No matter what I tried, even multiple BIOS revisions, I could only run in x8 by x8. So, assuming the motherboard was the problem, I switched over to my trusty Gigabyte X58X-UD7 motherboard to complete the testing. While the motherboards themselves may contribute slight differences in performance, testing around the web has shown that differences should be within a percentage point or two.
Using the OC Guru software, I was expecting to be able to complete my testing at the same speeds I previously found the cards stable at, 1015MHz GPU and 1375MHz memory. The first problem I ran into was that the OC Guru application would show no problems setting both cards to my desired speeds and GPU-Z confirmed the changes. However after running any of the benchmarks, my scores would be about the same as prior to the OC, and checking CPU-Z would show that the cards were not OC’ed during the benchmark run. I found that If I backed off the OC a lot, down to about 1000MHz GPU and 1300MHz memory, I could get the settings to stick, but only on the second attempt after booting. For instance, if I booted up, set my desired speeds in OC Guru, then ran a bench, the clocks would always revert to default. Then setting the clocks to the desired speeds a second time, I would run the bench again and the clocks would stick. However, my scores were significantly lower than with stock speeds. So, I’ve omitted the overclocked CrossFireX™ results for now. If/when the problem gets sorted out, I’ll provide an update.
This bench is very heavily CPU bound and as you can see it benefits very little from overclocking or multiple cards. If I were able to run the CPU at about 6GHz, which unfortunately is not possible, then we would see a little more spread in the scores. The result shown is the best out of three runs. Overall highest score screenshot shown.
This bench is very heavily GPU-dependent which is clearly reflected in my results. OCing and multiple GPUs both contribute heavily to the final score; it scales very nicely with multiple GPUs. The result shown is the best out of three runs. Overall highest score screenshot shown.
This bench is mostly CPU-dependent with modern GPUs. We see a little improvement with overclocking and CrossFireX™ but not even close to a linear change. The result shown is the best out of three runs. Overall highest score screenshot shown.
This bench is my favorite of the 3DMark series, and is both very dependent on the CPU and GPU, although with modern GPUs it is also more heavily favoring the CPU overclock. The result shown is the best out of three runs. Overall highest score screenshot shown.
3D Mark Vantage
This is the newest bench in the 3DMark series and is still very GPU-dependent. As you can see, scores scale nicely with OCing and CrossFireX™. The result shown is the best out of three runs. Overall highest score screenshot shown.
S.T.A.L.K.E.R. Call of Prypiat (DX11)
This was one of the first DX11 titles available and is a good test of the full system. All benchmark settings were manually configured for maximum image quality. The result shown represents the average of three runs. Overall highest score screenshot shown.
Alien VS Predator (DX11)
This game bench is one of the newer DX11 benches to hit the streets, and is very GPU-dependent. The result shown represents the average of three runs. Overall highest score screenshot shown.
Overclocking with LN2
So the last phase of my testing involves pushing one of the cards to the extreme limits with liquid nitrogen. I really wanted to score some huge points at HWBot with this card and move my ranking up a few notches. Because my goal was for maximum global points, I stuck with single card benching for this phase. Not only are there more points for single card results, but it’s easier to handle only two LN2 pots as opposed to three and my 25L of LN2 was not really enough to exploit the maximum potential of a triple pot setup.
First order of business was to enable my GPU to receive some extra voltage, which requires a “hard mod”. This involves de-soldering and re-soldering resistors on the PCB and should only be attempted by users who know exactly what they are doing. So I will not go into details about the specific modification here. I was able to change the default voltage, or “VID”, of 1.2V to 1.4V. This particular type of mod allowed me to continue using OC Guru for software adjustment up to 1.5V, so I had some room to play.
After physically modifying the card for additional voltage, I sprayed the card down with a silicone conformal coating, and insulated the motherboard and the card. Then I strapped on my kingpincooling TEK9 5.0 LN2 container and got down to business.
I was fairly limited in the quantity of my testing due to a limitation of only 25 liters of LN2. So I started by testing the graphics card sub-zero with the CPU on water cooling. I’m not providing scores with the CPU on water cooling because that wasn’t the purpose of this testing. Finding the clock limits of the SOC was the goal, and that was accomplished with 3DMark Vantage. I was able to successfully complete the bench with the card at 1300MHz for the GPU and 1350MHz for the memory (effective speed of 5400MHz). The GPU needed 1.5V while the memory remained at 1.6V (stock). What I found very interesting is that the GPU only needed 1.4V to remain stable through the benchmark at 1250MHz. That’s a lot of extra voltage for that last little bit of speed.
Here are some pictures of the system while I was searching for maximum GPU overclock.
After finding the basic limitations of the card, I moved the CPU to LN2 cooling as well in order to maximize my benchmark scores. Here are some pictures of the system all setup:
Aquamark 3 is a pretty old bench now, and does not push modern graphics cards too hard. This is heavily CPU-dependent, so pushing my CPU to 6138.8MHz I was able to reach a score of 413,090 which I was pretty happy about.
3DMark 03 is very GPU-limited, and here you can see that the result is very high for a single GPU. However, compared with the top scores at HWBot, it is still a ways off from the best. While other 5870s have attained GPU speeds of up to about 1500MHz, my 1250MHz limitation meant a score of only 107,856 in this test.
3DMark 05 is just the opposite of 03, and like Aquamark 3 it is very CPU-limited. Maximizing CPU speed is very important, and with my CPU at just over 6.1GHz I was able to reach a respectable score of 47,144.
3DMark 06 was the only real problematic bench for me during this session. It was a bit of a disappointment to me, but I really did not have enough LN2 to waste on a bench that wasn’t playing nice. So, after a few unsuccessful attempts I moved on to 3DMark Vantage.
With 3DMark Vantage, the GPU tests are very GPU-limited. The CPU test does take advantage of Hyper Threading and puts an extremely large load on the CPU, but the CPU test score has less weight in the overall score. With those things in mind, while benching Vantage with a single GPU, the GPU clocks are vital! CPU clocks are not as important. I had my CPU clocks very conservative here, so there is definitely a little room for improvement. But alas, the GPU clocks are just not high enough to compete at the top, so there wasn’t much point in pushing the CPU higher.
The surprise with Vantage was the fact that I was unable to get my GPU to complete the test at the previous speed of 1300MHz, so I had to settle for 1275MHz GPU and 1350MHz memory speed. This resulted in a very respectable score of 27,419, unfortunately, that’s still a long way off from the top 5870 scores at HWBot which exceed 30K points.
After getting a solid 3DMark Vantage score, I moved back briefly to 3DMark 06, but was still unsuccessful in completing the test with the remaining LN2, so I had to complete my testing without a 3D Mark 06 score.
Well honestly, this card is a mixed bag for me. I like the stock cooling system; effective, quiet, and most importantly it’s easy to remove. I like the component selection and the attention to detail with overclockers in mind. I like the idea of the binning process, but this is also where things turn sour for me. These GPUs were binned on air, which is great for the majority of the market, the gamers, but does nothing special for hardcore overclockers.
Reference cards have been available for months prior to these custom cards, and we have seen many users reaching 1000MHz core speeds. Although 1000MHz is not guaranteed on a reference card, I’d be willing to bet that a vast majority are capable of 900MHz or more. In addition, memory overclocks of 1300 are the norm, and 1350 are not uncommon.
So this really turns into a question of value. I’m going to make some generalizations here that some of you may disagree with, but this is how I see it and if you want to know how I would want to spend my hard earned money, here it is.
With reference cards selling for less than $400 these days, the Gigabyte SOC card commands a $100+ premium. If a reference card can be expected to reach 900MHz core and 1300MHz memory, and a Gigabyte SOC card can be expected to hit 1000MHz core and 1350MHz memory, then the question becomes this: “Is the Gigabyte SOC worth a 25% increase in price for about a 10% increase in performance?” My answer, no!
The overclocker in me is even more convinced. These cards were not binned with extreme cooling, so they are not necessarily going to be good overclocking cards on LN2. I only tested one of the two with LN2, but even if the 2nd one was amazing on LN2, the fact still remains that there is no guarantee of top shelf overclockability with extreme cooling. So again, if it was my money, I’d go for the best priced 5870 available and push it as far as you can!
“Thank you!” to Gigabyte Japan for supplying the pair of SOC cards, the UD7 motherboard, and the LN2 used in this review. Stay tuned for a more in depth look at the UD7 as well as the new UD9 coming very soon. Please feel free to ask questions or voice your comments below, thanks for reading!