Last week, Intel announced that it would delay its 7nm process node at least six months. According to company CEO Bob Swan, the node is at least 12 months behind Intel’s internal roadmaps. In the short term, that means product delays and probably some refreshed 10nm hardware in place of new 7nm equipment. In the long term, the potential outcomes range from “Intel regains process leadership” to “Intel becomes a fabless company.” Not much pressure there.
Intel’s current situation is historically unprecedented for the giant CPU manufacturer. While Intel shipped 10nm commercial laptop silicon in volume beginning ~12 months ago, the company has yet to transition its desktops or servers to the new node. Production volumes of Ice Lake servers are expected to start shipping at the end of 2020, with the first desktop 10nm CPUs scheduled for H2 2021. It took Intel about five years to move from 14nm Broadwell to Ice Lake (mobile), 5-5.5 years for Xeon depending on which models launch first, and six years for desktop. Rocket Lake may deliver a new architecture on 14nm, but Alder Lake will be the first new node.
While Intel’s 14nm transition woes have occupied a majority of column inches, they aren’t the only headwinds the company is currently facing. Two years ago, I wrote a story titled “Intel is at a Crossroads,” discussing both the company’s manufacturing problems and the bets it was making in fields like AI and 5G. Two years later, some of those bets have either failed or have yet to pay off.
Intel’s 5G business became much narrower after the company sold its modem IP to Apple. Intel’s upcoming Movidius platform, Keem Bay, which reportedly offers much higher efficiencies than competing parts from Nvidia, allegedly integrates a Cortex-A53 CPU rather than an Intel x86 part. In January, Intel announced it was ending its investment into Nervana and moving its product plans to IP created by rival Habana Labs.
Ponte Vecchio, which was supposed to be Intel’s first 7nm chip and set to debut in early 2021, has been pushed to late 2021/early 2022. Immediately after notifying investors that Ponte Vecchio had slipped, CEO Bob Swan stated: “We now expect to see initial production shipments of our first Intel-based 7-nanometer product, a client CPU in late 2022 or early 2023.” The implication here is that Ponte Vecchio is either no longer being built on 7nm, or that the GPU core is no longer being built at Intel. During Intel’s conference call, multiple investors picked up on the fact that the company referred to 7nm as being 12 months behind internally but facing only six months’ worth of delay, and sources we spoke to with some knowledge of Intel’s roadmaps echoed those uncertainties. There is not a great deal of confidence in the industry about Intel’s ability to hit these new dates — not given the tremendous problems the company has encountered to-date.
Intel’s PSG (Programmable Solutions Group), which contains its Altera FPGA business, continues to bump along at ~$500M per quarter — not chump change, but well behind Xilinx. The entire concept of CPU FPGA on the same piece of silicon seems to have taken a backseat to other advances. Intel launched a Xeon with an integrated FPGA with the Xeon Gold 6138P back in 2018 but has not updated the SKU since.
Optane may have long-term potential as a memory technology and 2nd generation should be faster than first, but existing software often has to be rewritten to take advantage of Optane’s characteristics and it’s still at a disadvantage to NAND in terms of cost. Foveros and Intel’s Omni-Directional Interconnect could be foundational breakthroughs for future technologies, but even the best interconnect needs solid components to attach to. Jim Keller’s decision to leave for “personal reasons” doesn’t really pass the sniff test.
Intel has a remarkable habit of making tons of money when it’s in trouble, technologically speaking. Prescott was both the worst CPU Intel ever shipped and a revenue record-breaker when it was new, and Intel’s overall data center and CPU sales have been doing very well the past few quarters. Part of that boost is from COVID-19, but the company was enjoying robust demand even before the pandemic hit. Those of you hoping that 2023 will dawn on a broken Intel, begging for financial relief from the likes of AMD or TSMC had best temper those expectations. Intel will continue to do well for a number of reasons, including continued excellence in specific markets/workloads, familiarity, inertia, multi-year purchase agreements, and being the platform of choice for a lot of software in circumstances where customers buy what their applications officially support, Intel is going to continue to hold on to large segments of workstation, desktop, and server markets.
If this was 2012, the article might end right there. With no ARM or x86 competitors on the horizon, Intel could ride out the next 3-5 years, fix its process nodes and its CPUs, and get back in the saddle without ever facing a serious competitor. But this isn’t 2012, and Intel is facing threats across the spectrum. Its rival x86 manufacturer, AMD, is making a serious play for overall market leadership. At the same time, Intel faces not just one ARM-based competitor, but a number of them: Qualcomm and Apple in laptops and Amazon, Ampere, Nuvia, and all other Neoverse-based products in servers. All we need at this point is for Western Digital to announce a new RISC V-based high-performance processor.
It’s not the near-term risks in 2021 – 2022 that should be keeping Intel investors up at night. It’s the question of how this situation evolves in 2024, 2025, and beyond. At this point, there is little reason to believe that Intel is going to have 7nm ready to ship six months after its original target. Backporting 10nm features into Rocket Lake’s 14nm will buy Intel some time, but the company obviously can’t keep designing chips for the same node forever.
All of these delays are virtually certain to pile into each other, because there are some critical technology introduction windows that have to happen at node transition boundaries. EUV has to go in. Once installed, continuing to scale downwards will require either high-NA EUV devices or the adoption of multi-patterning. Either way, these aren’t fire-and-forget technologies for Intel — they’re technologies that have to be installed and then scale downwards while improving yield as well.
It’d be stupid to count Intel out altogether, given the company’s history and long track record of proven performance — but it’d be just as stupid to pretend everything is going well at the silicon giant. I stand by my estimates from last week — Intel has 2-3 years to demonstrate a fundamental turnaround, or investors are going to start raising serious questions about its future as an IDM.