Chip fab Intel said to be using better chip fab TSMC to make 5nm Core i3 processors, 20% of its non-CPU parts

Ahead of Intel publishing its latest quarterly financial results on Thursday, let's have a quick reminder of who is actually making chip manufacturer Intel's chips these days.

Market analysts at TrendForce claimed this month TSMC will handle the mass production of Intel's 5nm Core i3 processors in the second half of this year. Chipzilla's mid- and high-end processors parts will be built by the Taiwanese foundry using a 3nm process node in the second half of next year, it is also thought.

And TrendForce reckoned "Intel has outsourced the production of about 15 to 20 per cent of its non-CPU chips, with most of the wafer starts for these products assigned to TSMC and UMC," citing its own "investigations." By non-CPU, it means some of Intel's graphics processors and other components as well as sub-components.

But that’s not all. Intel is said to be in talks with TSMC for “five projects,” some of which include GPUs, AI accelerators, and system-on-chips, according to Nikkei Asia. Beyond these partnerships, Intel will, presumably, fabricate the rest of its product lineup by itself as usual.

Unlike most semiconductor outfits, Intel has the facilities to both design and manufacture its own processors in-house. It used to be at the leading edge in the chip fab world, yet over the past five years, it has suffered set backs – delays to its 10nm and 7nm process nodes, primarily, while stumbling with cellphone chip and cellular modem designs – and lost ground to rival foundries TSMC and Samsung.

Mass production of 7nm processors is well underway at TSMC, and it’s already rolling out 5nm and soon 3nm parts. Chipzilla’s fabless competitors, such as AMD and Nvidia, have tapped up TSMC to make their silicon, and thus have stolen a march on Intel. And Apple, a long-time customer of Intel's x64 processors, is touting MacBooks powered by its astonishing homegrown Arm-compatible TSMC-fabricated 5nm M1 chip. It's said Apple is TSMC's primary customer right now, responsible for an estimated 20 per cent of its revenues.

When it comes to PCs, AMD has been snapping at Intel’s heels for the past few years with its Ryzen family. Last week, AMD CEO Lisa Su launched the Ryzen 5000 series of microprocessors, a line-up of 7nm TSMC-made chips using AMD's latest Zen 3 architecture. The first notebooks containing the new hardware are expected to reach markets as soon as next month.

Where’s Intel? Well, last time we heard from Chipzilla, a six-month delay in its 7nm technology meant parts made with that process node won’t actually be available in devices until late 2022 or early 2023. Server-grade components using the process node are expected to arrive the first half of 2023. That may well explain why Intel would shortcut to 5nm with TSMC with its desktop Core series, starting with its entry-level parts this year.

To put it all in perspective, TSMC's 7nm is on a par with Intel's 10nm node, in terms of transistor density, and TSMC's 5nm with Intel's 7nm. While we're still waiting about a year or more for Intel's 7nm, TSMC is already shipping 5nm. Similarly, TSMC was shipping its 7nm while Intel was struggling to get its 10nm out of the door.

Then there's the rise of Nvidia. The boom in AI and machine learning has reinvigorated the semiconductor world, making GPUs good for more than graphics and gaming. They’re used to crunch through massive amounts of data to train neural networks and to run inference algorithms for all sorts of applications, from smart speakers to self-driving cars. Intel has tried to elbow its way into this emerging market for better or worse.

First, Intel snapped up startup Nervana Systems in 2016 for $350m. Nervana was tasked with building a software and hardware stack fully optimized for deep learning. But after Nervana was swallowed up, it was burdened with bureaucratic issues and it, too, suffered numerous setbacks.

Nervana repeatedly overpromised and underdelivered, and, Intel eventually decided to scrap the whole effort before the team could ship its AI chip for training. Intel’s solution? Stumping up $2bn to buy another AI chip startup, Habana. All while its fabs floundered on the new process nodes.

Second, Intel decided it would have a crack at building its own standalone GPUs again, with gaming, AI, and supercomputers in mind. Dubbed Intel X^e, this technology is gradually making its way into systems. The architecture was due to appear in 7nm chips code-named Ponte Vecchio, destined for the US government’s 1-exaFLOPS Aurora supercomputer in 2021 though it may be pushed back to 2022. It's believed TSMC is producing parts of the Ponte Vecchio packages with Intel, and is part of that 15-to-20-per-cent non-CPU figure quoted by TrendForce.

Intel's then-CEO Bob Swan revealed mid-2020 that the processor giant was considering calling in the help of external foundries to meet its deadlines. “To the extent that we need to use someone else's process technology, we will be prepared to do that," he previously said.

One area that Intel has reigned supreme is with Xeon processors for data center compute servers, where the x86-64 architecture remains dominant. But, again, its competitors are ramping up efforts in this space and winning contracts.

AMD touts Zen-based x86-64 Epyc processors aimed at servers and high-performance computing, and its latest chips are going into European supers that are being built over the next few years. The world's fastest publicly known supercomputer is Japan's Fugaku that uses Fujitsu's Arm-based A64FX processors. Europe is eyeing up a mix of Arm and RISC-V cores for future super designs. Nvidia’s big cheese Jensen Huang hinted his corp will be focusing on Arm server CPU designs when it gets its hands on Arm, if the merger completes. Ampere has a pretty decent 7nm 80-core Arm-based server processor in the wild, too. Then there's Amazon Web Services' 7nm Arm-based Graviton2 servers. And, yeah, you get the point.

As Intel scrambles to get its mojo back, it makes sense to offshore some of its manufacturing to foundries like TSMC. That way it will have a better chance of keeping up with its competitors and supply demands – TSMC capacity willing – while it gets its breath back and attempts to iron out the kinks in its own processes. A new CEO, Pat Gelsinger, once Chipzilla's CTO, was ushered in this week to lead its renewed efforts.

Intel declined to comment, though we note it is in a mandatory quiet period ahead of its earnings release. A TSMC spokesperson told us: "Intel is our long-standing customer and we do not comment on individual customers." ®