The Thermal Price of Silence
Peak Power, Revisited
The Shift was stable in every benchmark we ran and never crashed in any single test, including tests of Prime95 and FurMark provided they were run separately. That said, the system consistently crashed in our Peak power benchmark at or near the one hour mark. One of the main reasons we run that particular test is to ensure a system remains stable under a worst-case thermal scenario; when we started measuring the Shift's CPU temperatures some of what we found was disturbing. Let's start with a look at the Shift's temperatures in our Peak test. One thing we realized once we started poking at the system's thermals is that the presence or absence of the Shift's port-covering grate has a substantial impact on CPU temperatures.
95'C isn't necessarily the highest temperature the Shift's overclocked Core i7-980X would reach, but Prime95 became unstable at that point and began shutting down worker threads (and lowering the CPU load). With the grill removed, the CPU topped at 89'C without generating any errors in Prime95. When we discussed the high CPU temperatures with MainGear, they pointed out that our Peak power test is an absolute worst-case scenario that consumes far more power than any other test or program we've ever measured. That's a fair point, so we headed back to the Shift for another round of testing. This time around, we focused strictly on the CPU. We ran 12 threads of Prime95 using that program's maximum power consumption test. In order to establish a baseline for comparison, we also reset the BIOS for stock CPU speeds and voltages. The only options we didn't change back to defaults were the power management and throttling settings that MainGear left off when it shipped the system. Everything else—uncore, RAM, QPI links, etc—were all set default.
The plot thickens. We continued to see the same 6'C difference between grate-on and grate-off measurements. At stock speeds, even with the grate installed, the CPU got no hotter than a modest 54'C. At 4.2GHz (and MainGear's overclocking voltages), Intel's 980X was far, far hotter. The CPU was stable at this temperature—Prime95 reported no errors even after several hours. Modern processors, both CPU and GPU, are capable of running at temperatures hat would've fried processors from years ago, but we're not comfortable with seeing CPU temps pushing over 80'C. We checked in with Intel for additional information on what constitutes a safe operating temperature. Here's what we were told:
Without knowing the device specifics or environmental conditions, my guess would be that 91C is probably fine. I believe the Tj Max for the 980X is in the range of 100 C, so under a heavy load 91 C sounds reasonable.It is no secret that running a device at a high temperature for a long time will impact its long term lifespan...Voltage has a bigger influence than temperature and temperature has a larger influence than frequency...If the device reaches 91 C for relatively short periods of time (I'm talking hours, days or weeks, not months) then the impact on the lifespan will be negligibleApplying this information to our results, we see that the MainGear's Peak test temperature of 89'C (w/o the grill installed) falls within the upper range of temperatures Intel classifies as safe. The 95'C temperature, in this case, does not—the CPU has become hot enough to begin throwing errors, which renders questions of safe operation moot.
The one issue left on the table was whether or not the radiator pump on the Asetek CPU cooler was functioning properly or not. This was a possibility MainGear had suggested when we first discussed our temperature data; while we thought it unlikely, it was something worth exploring. Since there's a direct correlation between power consumption and heat, we opted to measure the Shift's Gulftown 980X in three distinct states. First, we measured total system power at the wall with the CPU set at stock speeds and voltages. For the second state, we applied all of the overclocked voltages MainGear had used, but kept the CPU and memory at stock speeds. Finally, we returned the system to the same settings MainGear shipped to us.
While these numbers reflect total system power, the Prime95 stress test we used focuses on the CPU near-exclusively. Just increasing the CPU and system voltages to their MainGear-set levels increased power consumption by 89W; pushing the core to 4.2GHz from 3.33GHz resulted in a further 55W rise. That's a total of 144W of additional power—if we assume the CPU accounts for even 80 percent of that, it means the Asetek unit is attempting to dissipate 115W of additional power, on top of whatever the CPU was using at stock. Given that Intel defines a 130W TDP for Gulftown, we've nearly doubled the processor's potential TDP without changing the cooling solution one iota.
Based on this data, we see no reason to conclude that the Shift's radiator wasn't working—on the contrary, it seems to be working very, very well. With no reservoir tank and only a relatively small amount of fluid in circulation, the Asetek unit is more dependent on air passing over the radiator to reduce the water's temperature than it would be if it had a half-gallon of fluid moving through the system. Air, meanwhile, is subject to the basic laws of physics—the hotter the air moving over the radiator, the less effectively it cools. That's also why we saw the variation in stock temperature between our Peak test and our Prime95 test. During the Peak test, the two GeForce GTX 480 cards are running at 98-105'C each, which has a significant effect on the internal case temperature.
Turn Up The Airflow Or Turn Down The Voltages
MainGear has multiple avenues from which it can approach the problem, but since the Shift is designed for quiet operation, we favor reducing power consumption as opposed to using louder fans. A quick peek at overclocking databases online shows plenty of folks hitting 4.2GHz on air with Gulftown CPUs almost all of these are using much lower voltages than the overclocked configuration MainGear provided. As it turns out, we were able to lower the Shift's voltages significantly without negatively impacting system stability.
The difference between the voltage values may be small in absolute terms, but their collective impact on the system's operating temperature was substantial. Using these settings, the Shift drew 485W at the wall during our Prime95 test and reached a maximum temperature of 78'C (grate on). While that's still toasty, it's a solid improvement over what we saw using MainGear's default settings.
The other option would be to increase the amount of available CPU cooling. The Asetek is fitted with a pair of Silverstone R121225SL 120mm fans in a push/pull configuration. These fans emphasize silent operation more than airflow and while you wouldn't normally find us saying that a system needs to be louder, the Shift could give up 2-3 decibels at peak and still be just half as loud as Origin's Genesis. Regardless of what path the company chooses, we feel this is an issue that needs to be addressed. Our thermal tests were conducted in an open-air environment; the PC was dust-free, sat on a wood surface to ensure optimal intake clearance, and happened to rest directly below an air vent. Ambient temperature was kept around 65'F. Even under these conditions, the Shift hit temperatures ranging from 85C in Prime95 to 95C in our Peak power test. Toss in a few months worth of collected dust, an ambient temperature of 80F instead of 65, and a space that's merely adequately vented, and it's all too easy to see the Shift becoming unstable on account of heat buildup.