When Intel launched the Core i9-9900K, it established a new best-in-class performance record with a price tag to match it. One of the questions raised in the wake of the launch, however, was why Intel was calling the Core i9-9900K a 95W TDP CPU when it used far more power than that under full load. General power consumption measurements on the CPU came back in the 160-180W range — far more than 95W nominal TDP that Intel actually specifies.
We’ve touched on the explanation for this behavior before, but Anandtech has written an excellent article delving into how Intel CPU TDP is calculated these days and the guidelines that determine overall Intel CPU power draw. The report gathers some important information and I recommend reading it.
TDP, in its most narrow sense, refers to the ability of the CPU cooling solution to dissipate heat. A CPU with a 65W TDP requires a cooler that can handle a 65W TDP processor. A CPU with a 95W TDP requires a 95W cooler, and so on. But Intel’s rated TDPs are always calculated at the base clock, with no turbo frequency enabled at all. In other words, a CPU cooler capable of handling a 95W TDP is suitable for handling a 9900K running at stock frequency. As soon as Turbo Mode kicks in, all bets are off in terms of overall power consumption.
This has become a problem now, when it wasn’t in the past, because CPUs are now much more likely to slam into the TDP limits Intel defines. In the old days, it wasn’t unusual for a quad-core CPU to sit at or even below its defined TDP. In our power consumption tests back when the Core i7-4770K launched, for example, the difference between idle and peak system-level power consumption came out to 84W. Granted, that’s a test of system power consumption (and the CPU still drew some power at idle) but the 84W difference between idle and peak power consumption was a sign that the CPU’s 84W rated TDP was in the right ballpark. Today, a chip with a 95W TDP might plausibly pull 150 – 180W at the wall under full load, due to the impact of additional cores and higher burst frequencies. TDP was never intended to stand in as a metric for how much power a CPU would use, and Intel has always been clear on that point, but it’s now much less accurate than it used to be. Unfortunately, this is where the complications start.
Intel now defines two distinct power levels (PL1, PL2) and a time limit, tau. PL1 is typically analogous to the rated TDP — it’s the long-term power consumption expected over time. PL2 is the CPU’s burst power consumption, while tau is the amount of time the CPU is allowed to run at that increased power consumption level. But motherboard companies are allowed to modify the settings that control how these features are implemented, and the results can change the expected behavior of the CPU dramatically. It’s possible, for example, to put a system into “PL2 forever” mode, where the CPU simply runs all-out, at maximum frequency. This tends to dramatically improve benchmark scores but not every CPU is stable in this configuration (ExtremeTech tests CPUs with this all-core boost behavior disabled, in all cases).
I don’t want to cut too deeply into Anandtech’s reporting, but the site’s findings show that performance on a high-performance Intel chip can be dramatically different depending on which coolers and motherboard settings are in play. Performance can shift by as much as 29 percent, depending on which configuration is used — and some of these variables aren’t well explained in menus or particularly well-detailed.
Anand calls for Intel to disclose two values: A TDP (Peak) corresponding to its PL2, and a TDP for sustained power for PL1. Fixing the issue on consumer motherboards, where pushing the envelope is often a great way to one-up other vendors in benchmarking tests, is going to be tougher. But if you’ve ever wondered why power consumption in modern chips seems tenuously connected to the numbers on the box, issues like this are why.
Now Read: Intel Announces 48-Core Cascade Lake Xeon CPUs With 12 Memory Channels, Intel’s 9th Gen CPUs Are Still Hard to Find at Retail, Intel Smashes Earning Targets, Denies 10nm Problems