Gelsinger holds 20A Arrow Lake wafer

How Dire Are Intel’s Problems?

You’re crazy for this one, Rick.

Steadily increasing over the past several years, criticism of Intel got louder with its 2Q24 earnings. It wasn’t a good quarter, but it wasn’t dire, as we discussed. Revisiting past decisions and the current state of the company shows some attacks to be too pointed. Nevertheless, the next six quarters or more will entail difficult changes.

Missed Opportunities

Intel has been castigated for missing the handset wave, particularly the smartphone transition, and the AI revolution. These criticisms are ultimately true, but it’s more a story of bets that didn’t pay out.

Cellphone Chips

  • Walking away from the iPhone probably made sense at the time. Intel had fat profits and Rubenesque staffing; an ASIC-style deal would’ve diluted margins. Moreover, Nokia dominated the handset market to a degree no OEM has since matched, and the tech savvy and hyperconnected favored Blackberry for its revolutionary email abilities.
  • Intel got into cellular modems in 1999 by acquiring DSP Communications (DSPC). In that era, Intel’s record of selecting the right targets and growing them was poor, and DSPC was no exception. Nevertheless, it’s unfair to say Intel didn’t see value in the handset market. This deal was the first of several for handset-chip technology.
  • Intel supplied Blackberry. Intel supplied the then-hot company with application processors integrating the cellular modem. For example, the PXA901 employed Intel’s modem technology and Arm-compatible XScale CPUs. The latter came from acquiring the StrongArm business from Digital Equipment (DEC).

In summary, Intel didn’t capitalize on mobile-phone growth, but it’s not for lack of trying or foresight. Agreeing to supply Apple would’ve put it in a different place, but that’s only obvious in hindsight.

Artificial Intelligence

It’s too early to conclusively say Intel missed the AI revolution. Even for data-center accelerators (GPUs/NPUs), the company isn’t dead, only doomed. As with handset chips, Intel’s acquisitions tell a story.

  • Intel acquired Nervana in 2016 and for a while looked like it would land a big Facebook win. As with most Intel acquisitions, Nervana went nowhere.
  • Intel bought Movidius at about the same time. This deal has borne fruit; the acquired technology has evolved and forms the basis for Meteor Lake’s and Lunar Lake’s NPU. The undersized Meteor NPU didn’t garner Microsoft’s blessings, but the Lunar NPU has.
  • Habana folded into Intel in 2019, and its Gaudi data-center NPUs aren’t bad. However, even Intel CEO Pat Gelsinger concedes it’s behind Nvidia, hyperscalers’ internal designs, and AMD. Moreover, with a separate data-center GPU operation, Intel had too many overlapping design efforts. Future products (e.g., Falcon Shores) will add some Gaudi capabilities to the GPU architecture, but this technical reset further delays the GPU/NPU business, likely dooming it.

Intel is at competitive parity in PC-based acceleration. In servers, its CPU’s AI instructions could prove relevant for lightweight workloads. For server-side acceleration, the company recognized the opportunity by 2016 but won’t have much to show even ten years later—possibly ever.

5N4Y

The biggest concerns about the company center on manufacturing and the Intel Foundry Service (IFS). Confusion about the final two nodes in the Intel 5N4Y plan exacerbates investors’ agita.

  • The first node doesn’t count. Intel 7, originally called the 10 nm Enhanced SuperFin process, predates Gelsinger’s reign.
  • Intel 4 is only for Meteor Lake’s computing die. It lacks the collateral support to be used for much else, rendering it less than a full node in operational terms. It’s important evidence showing that Intel was regaining its technology and manufacturing prowess.
  • The 20A node, likewise, is only a technical milestone. Recent public statements indicate that Intel will outsource all active Arrow Lake wafers to TSMC, leaving 20A without a product.
  • Intel claims 18A is on track, but other sources say it isn’t. At a minimum, Intel must fill 18A fabs with its PC and server processors. The company states the Lunar/Arrow follow-on called Panther Lake and the E-Core Xeon called Clearwater Forest will employ 18A but is mum about the next P-Core Xeon, Diamond Rapids.
  • Landing external IFS 18A customers is important but secondary to internal use. Gelsinger’s public statements acknowledge the company lacks committed foundry orders. We can’t blame customers for being skeptical that Intel will deliver the yields, performance, power, and area (PPA) they require. Moreover, they need EDA support, physical IP such as cell libraries, and design IP. Early 18A yields and the process-design kit may be on track from Intel’s perspective. However, external customers require confidence that yields will improve and the technology will meet PPA targets with their design flows.
  • Putting five nodes in production in four years was never going to happen. Intel had to prove it got its technology and manufacturing mojo back, but it couldn’t risk cascading failures. Therefore, it promised five nodes in four years but decoupled Intel 4 from Intel 3 development and 20A from 18A. However, fleshing out each node with all the required complementary technology and ramping up production sufficiently to adequately amortize the fixed costs was never possible.
    • Intel 4 and 3 are similar enough to share a fab, and the company has moved them from the Oregon technology-development (TD) plant to the production-oriented Ireland facility. The company, however, is outsourcing most PC-processor silicon to TSMC, leaving Ireland underutilized in the short run. Ireland alone probably isn’t big enough to properly amortize both nodes’ costs, anyway.
    • The 20A node could share production with 18A, but building 20A capacity and the complementary technology for a single computing die is costly. It’s better to allocate resources to 18A, employ the process instead of outsourcing to TSMC for all PC and server processors, and invest in winning external customers.

Recapitulating, 5N4Y rallied Intel around a goal and allowed it to show it hadn’t permanently forgotten how to develop new process technology. However, 5N4Y didn’t make sense operationally. The company must ramp up 18A next year without hiccups to restore outsiders’ confidence.

Products

Intel also had to show it hadn’t forgotten how to make new products. Here, too, problems remain.

  • In servers, AMD has been taking share for years. Fortunately for Intel, AMD has captured less than a quarter of the market in unit terms. Unfortunately for Intel, AMD has a 33% revenue share, indicating Intel is losing at the high end. As good as the new Granite Rapids Xeon is, AMD’s forthcoming Turin Epycs will retake the core-count crown.
  • In PCs, Intel still dominates but recently has been ceding share to AMD, and both face challenges from Arm-compatible processors, such as the Qualcomm Snapdragon Elite. The PC market, however, is surprisingly sticky for Intel, and its PC processors aren’t that bad despite their faults. The company could cede share and suffer historically low margins for another year, but it will still dominate the x86 PC market.
    • Intel’s main PC processor, Raptor Lake Refresh, is an uninspired design. Targeting laptops instead of replacing Raptor across the board, Meteor Lake is an expensive amalgam of dice. Lunar is also narrowly targeted but greatly improved. Once an idiosyncratic stopgap, Lunar will play a bigger role in Intel’s product lineup. Another multidie processor outsourced to TSMC but this time with DRAM on package, it’s not great for Intel’s margins.
    • Originally expected to showcase 20A and occupy the entirety of Intel’s PC-processor stack, Arrow Lake is now only more water treading until Panther Lake enables the company to replace its whole PC-processor lineup with 18A chips.

Other Options

  • Altera will be spun off. Intel talks about an IPO, but recent reports suggest another company could buy it. An IPO would likely result in Intel holding most of its shares. A sale would give the company cash. Altera should be able to stand on its own, but an FPGA and gate-array (eASIC) business could complement a fabless ASIC company, particularly one serving the telecom market.
  • Separating IFS is premature. The operation isn’t ready to serve customers other than Intel’s product groups. Capital investment is now being done with financial partners and government grants, mitigating Intel’s risk.
  • Selling off individual product groups, like the PC-processor business (CCG), has been mooted by the press. Particularly with respect to CCG, chaos would result.
  • Selling the whole product side would be less chaotic but wouldn’t fix anything.

Bottom Line

Intel CEO Gelsinger hasn’t righted the ship, and the company isn’t in great shape. He announced in the 2Q earnings call the company will lay off 15% of its staff and cut capital spending. We think these steps should’ve been taken sooner, particularly the always-difficult staff reduction. The data-center GPU/NPU business’s viability is an open question, and we’re not optimistic. However, the Intel 18A process’s uncertainty is the biggest concern, and the company must be more transparent about how 18A compares with peers’ contemporaneous technologies, especially in the context of external customers’ design flows.

A noteworthy industry analyst recently asked us how long Gelsinger has. Our guess is he’s safe through 2025 but could face slings and arrows. Cutting data-center GPU/NPU development would be controversial—perhaps career ending—but could be necessary. A viable foundry business that fabricates hyperscalers’ NPU designs would soften this decision’s impact. Otherwise, there’s no obviously better path than the one he’s laid out, which is to focus on 18A, use it for new products, and attract foundry customers.


Posted

in

by


error: Unable to select