Making Sense Of Turbo CORE
Turbo CORE is, without question, the one feature that took me the longest to dig into for this piece. First described by AMD as a deterministic capability, I took that to mean Turbo could tell when load was being applied to three or fewer cores, at which point it’d spin down three idle cores (dropping voltage and frequency) and overclock the other three, incrementing multipliers and voltage settings.
In my preview, I said:
“As far as we’ve been able to determine, Turbo CORE isn’t as granular as Turbo Boost. It operates based on operating system P-states, so when three or more cores are at low utilization levels and the active cores are in P0 (for more on P-states, check this out), the CPU capitalizes on TDP budget to increase performance by a flat 400 or 500 MHz.”
This isn’t the case, though. In practice, Turbo CORE really just works like Cool’n’Quiet in reverse. The processor can take a default 16x multiplier (on the 1090T) and increase it to 16.5x, 17x, 17.5x, or 18x, depending on activity. At the same time, you might not even see the other cores slow down at all.
Perhaps the best illustration I can give you is a couple of quick videos. In the first video, I’m running our iTunes benchmark, converting a WAV into an AAC file. This is a single-threaded workload, so we’d expect Turbo CORE to be doing its thing. You see the thread jumping from one core to the next in AMD’s OverDrive utility. At no point do we see any of the idle cores drop below 3.2 GHz. However, the active cores do hit 3.3 and 3.4 GHz.
Now, I fire up a threaded workload like MainConcept, we see something different. For the most part, all of the cores operate, fully-loaded, at 3.2 GHz. Strangely, we actually do see certain cores peaking at 3.6 GHz with increased 1.475V spikes.
With none of our test workloads demonstrating three cores throttling down in response to higher clocks and voltages from the other three, I was worried that power might be suffering accordingly. If you remember back to my coverage of Core i7-920XM, I was concerned that Turbo Boost would keep the CPU at a higher frequency more of the time, and consequently burn through battery life faster. As it turned out, I was right, and an affinity for getting more done faster meant that Clarksfield was harder on energy use.
Anticipating something similar, I logged back-to-back PCMark Vantage runs on the Phenom II X6 1090T and X4 965 Black Edition:
These two actually map very closely. You’ll notice that the hexa-core chip does tend to peak at higher power numbers. Also, the lower red lines suggest that the quad-core contender drops to a lower idle power (also quite logical). But when you average out total consumption, the Phenom II X6-based system sits at 141W, while the X4-based machine uses 135W. That’s not bad.
At the end of the day, think of Turbo CORE as compensation for losing clock rate in the move to six cores. Just because the technology can hit 3.6 GHz on the Phenom II X6 1090T doesn’t mean you’ll get the equivalent of 3.6 GHz in single-threaded applications. Fortunately, at the same time, you’re also not paying a huge power price for the small bump in performance.
Of course, there’s a caveat. As was the case at first with Intel’s Turbo Boost technology, it’s entirely possible that AMD’s monitoring tools aren’t capturing the true nature of what’s going on here. The refresh in OverDrive is quite slow, and processor states can change many times per second. AMD has assured me that the behavior I’m seeing sounds right, but there’s always that chance…