search slide
search slide
pages bottom

Intel launches new Apollo Lake silicon, but new Goldmont CPU architecture may draw significantly more power

Earlier this year, Intel announced that its next-generation Atom architecture, Goldmont, would be confined to netbooks and the low-end of the desktop market. Previously, Goldmont was meant to serve as the anchor for a new range of Intel products and offer the first architectural refresh since Bay Trail launched in 2013.

Unlike Kaby Lake, which got the full launch treatment earlier this year, Goldmont has slipped out the door with almost no acknowledgment or briefing from Intel. Intel has added some features, including full hardware decode for VP9 and HEVC (though not HEVC Main10) support for the S0ix sleep state, Gen9 graphics (up from Gen8), and six PCI Express 2.0 lanes, up from four. The number of Execution Units (EUs) attached to the GPU is also up to 18, from a previous max of 16.

As Anandtech details, however, these relatively modest advances come with a significant increase in TDP. Intel’s previous Atom processors had a TDP of 6.5W across the product stack, while all of the Airmont products unveiled today are 10W chips. That’s a significant jump for the new architecture, though it may not translate into higher power consumption at the wall (we’ll return to this shortly).

Intel has previously positioned Apollo Lake as a cost-saving opportunity for the various OEMs, as the slide below makes clear:


There may be good reason for this. Cutting prices is one of the few ways of encouraging end users to buy, even in a PC market where prices have been rock-bottom for a very long time. Chromebooks have also stolen a significant chunk of the overall market and Intel undoubtedly wants more of these systems to ship with x86 hardware instead of the ARM-based solutions some companies have shipped. Intel’s Gen9 GPU will support the DX12_1 feature set and offers its own memory stream compression, which should help overall performance, though systems at these price points are rarely useful for anything more taxing than a Facebook game.

Intel is also claiming that Goldmont will offer a 30% performance gain on CPU and a 45% gain in GPU performance, though it didn’t offer any slides or details to back that claim up. We still don’t know what Intel changed or how Goldmont differs from Airmont — but this plays into the TDP considerations we mentioned above. Specifically, all of new Goldmont desktop CPUs in the Pentium J4 and Celeron J3 families have roughly equivalent clock speeds to the Airmont CPUs they replace, but are rated for a 10W TDP up from 6.5W for Airmont. The mobile Goldmont CPUs are all rated for 6W — but they also take a serious whack on base frequency, from 1.6GHz down to 1.1GHz.

TDP (Thermal Design Power) is a complex topic and is generally defined as the typical expected power dissipation of a CPU in representative workloads. We don’t know how Intel sets its clock points within its workloads — how much of the total time is spent at max Turbo, how much at lesser speeds, and how much at the base frequency. What we do know is that the CPU base frequency can be thought of as a sort-of floor. Since Intel defines the base clock as the CPU clock you’ll get at all times unless the CPU is overheating, and since TDP is a metric that Intel gives heatsink and cooling solution manufacturers to target for optimal performance, we can safely assume that TDP is calculated to ensure the CPU doesn’t get hot enough to start throttling to protect its own operation.

The fact that Intel has both bumped up the TDP rating on the desktop chips by 1.53x and cut the base clock on the mobile chips by 31% suggests that Goldmont may not have come out the way Intel was hoping it would. There’s enough play in these figures that Goldmont could still prove to be more power efficient than Airmont in certain workloads — again, we don’t know how Intel calculates TDP precisely, and if chips burst up to higher frequencies and complete workloads more quickly, the end result can be lower power consumption over time. We also don’t know if Goldmont is built on Intel’s 14nm process or the 14nm+ it used for Kaby Lake.

Up until now, most interpretations of Intel’s decision to leave the smartphone and tablet markets have focused on the fact that Intel was still shipping products contra-revenue and was less interested in trying to buy its way into that space. To be clear, I think those were still major components of the company’s reasoning — but it’s also possible that Intel looked at how Goldmont was shaping up and realized that whatever the CPU’s merits in mobile and desktop processing, it wouldn’t be an architecture that could take the fight to ARM in smartphones and tablets.

Leave a Reply

Captcha image