Did TSMC’s 20nm process harm Qualcomm’s Snapdragon 810?

Did TSMC’s 20nm process harm Qualcomm’s Snapdragon 810?

There’s been a debate brewing over the last few days over Qualcomm’s Snapdragon 810, its performance characteristics, and whether the SoC was harmed by moving to 20nm at TSMC. We’ve been investigating this issue for the past few days, and reached out to Qualcomm for additional information.

According to Qualcomm marketing executive Michelle Leyden Li, the Snapdragon 810 doesn’t have a problem. The official response from Qualcomm is that Snapdragon 810 has numerous design wins and will continue to roll out in multiple devices across the next few months. This reiterates comments from Qualcomm’s Steve Mollenkopf. During last week’s conference call, Mollenkopf told investors, “We are not seeing a change design share or the competitive environment, but rather a change in timing of some 810 designs…. In the premium tier, we are very pleased with the design traction on the Snapdragon 810 with over 60 designs having won the key premium design slots with the exception of Samsung.”

Officially, there is no problem. What about unofficially?

TSMC, Samsung, and 20nm: Known unknown

Here’s one thing we know — according to official metrics from both TSMC and Samsung, Samsung’s 20nm is more dense than TSMCs. Semiwiki compiled this informationbetween all three companies; consulting the charts shows that Samsung’s overall density (measured in terms of gate pitch and metal pitch) is significantly better than TSMC’s. In the old days — ten years ago — we might have concluded that this gave Samsung a power advantage, since dense transistor stacks required less power. Unfortunately, the breakdown of Dennard scalingmeans we can’t automatically conclude that this is true — and even if it is, we can’t really predict the degree of improvement. More information is needed.

The Snapdragon 810 incorporated a number of firsts for Qualcomm, and new technologies

Another thing we know: This is Qualcomm’s first octa-core big.Little chip that combines Cortex-A57 and Cortex-A53 cores and the first Qualcomm SoC to implement the Global Task Schedulingvariant of big.Little. If Qualcomm stumbled on implementing this technology, it could explain why the Snapdragon 810 spends comparatively little time in “little” mode. It wouldn’t be the first time that a major manufacturer had trouble with an aspect of big.Little, either, as Samsung’s Exynos 5410 was completely broken in this regard.

Apple’s apparent success with TSMC’s 20nm node isn’t really evident one way or the other. The Apple A9 SoC is a very different animal from Qualcomm’s Snapdragon 810. It’s a dual-core processor where Snapdragon 810 is an octal-core chip, and it runs at a much lower clock speed than the Snapdragon hardware. It’s also based on Apple’s custom architecture, whereas the 810 is a standard ARM core.

There are other design, yield, and software factors that could explain the issue. Twenty-nanometer Samsung chips (the Exynos 5433 in the Galaxy Note 4, specifically) don’t appear to suffer from the same issues, but AnandTech picked up evidence that under heavy CPU load, the Galaxy Note 4 Exynos version (20nm Samsung) performed markedly worse than the Galaxy Note 4 that used a Snapdragon 805. Poor big.Little supportwas the apparent culprit. It’s even possible that the GPU is partially to blame — the Adreno 430 is a new core for Qualcomm.

The importance of software support simply can’t be understated. According to AnandTech, a late software patch for the HTC 9 nearly doubled the phone’s battery lifewhen testing graphics, from 1.73 hours to over three. It’s possible that some of the issues detected to date have been caused by poor software or firmware updates — an issue that falls squarely on the OEM.

Right now, there’s no evidence to suggest that TSMC dropped the ball at 20nm in a way that harmed Qualcomm. The node’s timingand TSMC’s struggle to catch up to Samsung is a separate topic to whether or not the 20nm node is yielding well. There’s also no evidence of a crippling issue with the Snapdragon 810 that fundamentally breaks the processor. Sometimes a chip doesn’t hit its performance or power consumption targets because of several small issues rather than one large one.

One point I will stand firm on: Regardless of any trouble that Qualcomm, TSMC, or any other SoC vendor might have with a design, it’s the responsibility of the OEM to determine which chips do and don’t deserve to be in certain products. If Vendor X uses a chip from SoC designer Y, and the result is a substandard device, that’s on Vendor X’s head. Better to secure a different chip than to ship a device that’s not going to deliver a decent user experience.