The Obsidian Series 750D "Yamamura" Custom Liquid Cooling Build: Overclocking

(This is the third part of a multi-part blog. The first part talks about component selection and is here, and the second part details the assembly and is here.)

With the Obsidian Series 750D “Yamamura” built and fully operational, we have a powerful Intel Core i7-4790K and two NVIDIA GeForce GTX 980s all under water and a mountain of cooling capacity at our disposal. Certainly it’s completely unreasonable to keep these incredibly fast components just running at stock speeds – they need to be pushed to their limits.

Well, within reason. Specifically, they need to be pushed as far as they’ll go for 24/7 use. And at the same time, it doesn’t make sense to leave the fans cooling two 240mm radiators and a 360mm radiator running at full bore all the time.

The Intel Core i7-4790K we’re using is an engineering sample, and it had already been running at a 4.7GHz overclock under my last loop. This time I have access to an arguably better waterblock, more cooling capacity, and a more efficient pump. Yet 4.8GHz remains elusive and ultimately out of the reach of this chip. This i7-4790K will do 4.7GHz all day at 1.3V on the core and 1.9V on the VRIN, but that’s already a healthy jump from 1.225V on the core just for 4.6GHz.

The essential problem the chip ran into is heat. Devil’s Canyon does a far better job of conducting heat off of the die than conventional Haswell did, but you’re still looking at a combination of mitigating factors: high heat density stemming from a small die having to travel through Intel’s TIM and heatspreader into more TIM and then the EKWB Supremacy EVO block. There’s only so much heat that can be efficiently removed from that chain and unfortunately, the extreme voltage required to flirt with 4.8GHz just generates heat faster than it can be safely removed from the die. So we remain at a still speedy 4.7GHz.

NVIDIA’s GeForce GTX 980 is an odd duck when it comes to overclocking. The 980 is an extremely efficient chip, but NVIDIA tuned its stock design to maximize that efficiency. This means a very restrictive power cap and a reference PCB design that is unfriendly to high overclocks.

By watercooling the GM204 die, VRAM, and arguably most importantly, the power circuitry, we can circle around and eke out more performance. My experience with overclocking watercooled GTX 780s and 680s was that you’d be able to find the top speed the GPU could hit even under air, but watercooling ensured the chip spent the majority of its time at that top speed. You can ramp the fan on the stock air cooler to keep the thermal limits in check, but power limits were more likely to be hit, and the GPU clock would bounce between a few speed bins.

The unsung hero of watercooling graphics cards is the VRM cooling: by keeping the power circuitry running at substantially lower temperatures, it will operate far more efficiently. This actually ekes out headroom at the top of the power curve, where NVIDIA’s TDP limits come into play. Where the GTX 980 differs is that this area is still restricted, even with the 125% power limit applied. That top speed stabilizes a bit more, but the 980 still smashes its idiot head into the power cap.

Our only recourse is to modify the BIOS of the graphics card, which is something that should be done with extreme caution and care. The modified BIOS I used raises voltages along with the power cap, but in practice was not able to unlock higher speeds overall for the GTX 980. 1540MHz continued to be the highest speed the GPU core would run at, but the raised power cap further stabilized operation at this high speed. Incidentally, the modified BIOS also allowed VRAM speed to hit a staggering 8GHz, up from the 7.8GHz overclock on the stock BIOS. This is primarily academic, but 8GHZ on GDDR5 looks cooler than 7.8GHz.

You can see we can get pretty substantial increases across the board, but remember that the GTX 980 SLI overclocking doesn’t tell the whole story: the modified BIOS and increased power cap also help the GTX 980s sustain that 1540MHz peak.

Performance improvements when overclocked are absolutely staggering. We’re looking at almost 20% across the board, with the lone exception being Handbrake and its more modest 10% improvement. Minimum framerates improve even more; Tomb Raider now never ducks under 60fps. Keep in mind that BioShock Infinite and Tomb Raider were both run at 5760x1200 at maximum settings, with Tomb Raider running both TressFX and 2xSSAA.

While the Intel Core i7-4790K would likely be perfectly served by an all-in-one cooler like a Hydro Series H80i or better and still hit these performance levels, the GeForce GTX 980s pretty much demand to be put under water. A full cover waterblock, or at least a watercooling solution that properly accounts for VRM temperatures (like a forthcoming HG10 model), is key to unlocking Maxwell’s true performance potential.

In the next chapter, we’ll talk about power consumption and heat, and how Corsair Link and the Commander Mini are deployed to smartly optimize noise levels and fan power and keep the Yamamura running at peak performance.