Overclocked: Our Custom Radeon HD 5870 Roundup

Overclocking

As 2 of our 3 cards were significantly geared towards overclocking, we tried our hand at overclocking each of them. As our ultimate goal in reviewing any product is to evaluate it for long-term use, this philosophy holds over for how we handle overclocking. Our overclocking results are based on what we believe would be an overclock suitable for long-term use as opposed to something that may top benchmarks but potentially be unstable.

For Cypress cards, our ground rules are that we want to keep the GPU below 95C load under FurMark, as we are not confident about the GPU’s longevity past that mark. Furthermore all cards must be able to complete an extended period of FurMark without crashing, artifacting, or downclocking due to VRM protection. Finally every card must be able to complete our benchmark suite under the same circumstances. This is something our reference 5870 can achieve, so we shoot for the same thing out of our overclocked cards.

With that out of the way, we’ll start with the Sapphire Toxic 2GB, the non-overclocking card of the bunch. As it doesn’t feature voltage adjustment, it can only go as far as its stock core voltage of 1.2v can carry it. Furthermore the extra 1GB in RAM presents a further hurdle in overclocking.

We were able to get a moderate overclock out of the Toxic, bringing it up from 925/1225 to 960/1250, a 4%/2% overclock respectively. All told this is pretty good, especially since we’re only at 1.2v on the core.

Up next is MSI’s 5870 Lightning, which should be the most capable card among our selection when it comes to overclocking. Between the 12 VRM power phases for the GPU and the dual-8pin power sockets, power delivery should not be an issue for this card – any limit is going to be what the GPU, RAM, and memory bus can accomplish. MSI allows for voltage adjustment up to 1.35v, however under FurMark we exceeded our 95C limit at anything above 1.285v. So for our overclocking efforts we took the card as far as it could go at 1.285v.

Deciding whether something overclocked well or not is going to hinge a great deal on your definition of “overclocked” here. On the one hand we got a 4.5% core overclock, but on the other hand even with this overclock it couldn’t catch up to the stock clockspeed of the Gigabyte Super Overclock. Ultimately anything above 940MHz would cause the card to crash under FurMark, disqualifying any higher speeds. It may be stable for gaming at higher speeds, but it doesn’t meet our criteria for general stability. We suspect that with exotic cooling to handle higher voltages that this card could be pushed further, but we don’t expect too much more given that an extra 0.12v only got us 40MHz more out of the GPU.

The memory clock is a different matter though. As we’ve discussed previously the biggest issue AMD had with GDDR5 was not the RAM itself, but the bus, which is why reference Radeon 5870s ship with 5GHz GDDR5 running at 4.8GHz. An 8% overclock to 5.2GHz effective isn’t anything earthshattering, but it’s better than our core overclock and is all the more impressive given the wall AMD was hitting here.

Finally we have the Gigabyte Super Overclock, the card that straddles the line between a factory-overclocked card and a serious overclocker card. With an enhanced power delivery system it’s more capable than a reference 5870, but it’s not going to be the most capable card in our roundup. With Gigabyte’s OC Guru software we can increase the core voltage up to 1.28v, however we found that the GPU temperature would exceed 95C with that much voltage. For our overclocking attempts we had to dial it down to 1.26v to keep our temperatures manageable.

Frankly, as an overclocker card the 5870 Super Overclock was an abject failure. We couldn’t even get it stable under FurMark at 960MHz core clock with our 1.26v core voltage. Furthermore once we pushed the core voltage above 1.23v thermal protection would kick in for the VRMs, which means even if the card was FurMark stable the VRMs aren’t up to the task. Our impression is that Gigabyte is already heavily binning their Cypress chips to get a card that can run at 950MHz in the first place, resulting in many of these cards already running near their limits. The end result is that at 100MHz over reference, many of these Cypress chips have already given nearly everything they can give. As with the MSI card it may have some more headroom under exotic cooling, but as it stands that would only allow an overclocker to get another 0.02v out of it.

Meanwhile the memory overclock was more of a success story. At 5.3GHz effective it’s a 6% memory overclock on a card that was already running its memory 4% over reference, giving us a combined 10.5% memory overclock. This is made all the more impressive by the fact that Gigabyte doesn’t have any kind of heatspreader/heatsink attached to the RAM chips on the Super Overclock, meaning that this was accomplished solely on whatever air flows over the RAM chips from the card’s 2 80mm fans.

Ultimately none of our cards geared for overclocking proved to be great overclockers. Even with binning by partners there’s still a great deal of variability what core clock speed a GPU can ultimately hit, driving home the fact that no overclock is guaranteed. It’s entirely possible that we received particularly poor clocking samples from Gigabyte and MSI while receiving a well-clocking sample from Sapphire, or it may be some other combination entirely.

It’s always difficult to draw conclusions from a sample size of 1, but we work with what we have. From that the conclusion we can come to is that the average Cypress isn’t good for much more than 950MHz on air (at least, not without something even more exotic like a Thermalright Spitfire), so if this is the case then all of these cards are fairly similar in what they can achieve with core overclocking. More to the point this drives a stake in aftermarket overclocking when you can buy a card that ships at 950MHz in the first place.

Memory overclocking is a different beast however – and it’s the more complex of the two. As we’ve mentioned time and time before, the problem with GDDR5 on Cypress is not the chips or the memory controller, it’s the memory bus. With error detection and fast link retraining, memory overclocking can throw us a curveball if they keep the card stable while at the same time reducing effective memory performance. There are certainly some benefits to be had in memory overclocking which become apparent with the overclock on our Gigabyte Super Overclock, but it suffers from being anything but straightforward. A good memory overclock seems achievable on both of our 1GB cards, but at some unknown point you run the risk of reducing performance more than improving it.