During its second-quarter earnings call, Intel announced a one-year delay of its 7nm process technology, which it says will create an approximate six-month shift for its CPU product timing relative to prior expectations. The primary issue is a defect mode in the 7nm process that resulted in yield degradation, said Intel CEO Bob Swan during the July 23, 2020, earnings call.
“We’ve root-caused the issue, and believe there are no fundamental roadblocks,” said Swan. “But we’ve also invested in contingency plans to hedge against further schedule uncertainty. We’ve mitigated the impact of the process delay on our product schedule by leveraging improvements in design methodology, such as die disaggregation and advanced packaging.”
The delay in 7nm potentially puts a kink in Intel’s plans to stand up the Aurora supercomputer at Argonne National Laboratory on-schedule by the end of 2021. Intel, the prime contractor, is building the machine with HPE/Cray, using Intel’s 10nm “Sapphire Rapids” Xeon CPU and its 7nm “Xe” datacenter GPU, codenamed “Ponte Vecchio.”
As the centerpiece of Aurora, the frontrunner of the U.S. exascale program, Ponte Vecchio would herald Intel’s entry into the datacenter GPU market. The GPU was also intended to be Intel’s first 7nm product.
Ponte Vecchio’s release is now slated for “late 2021 or early 2022” and Intel is working with outside fabs on at least some elements of the GPU.
Bob Swan on the Q2 earnings call:
“We will continue to invest in our future process technology roadmap, but we will be pragmatic and objective in deploying the process technology that delivers the most predictability and performance for our customers, whether that be in our process, external foundry process or a combination of both. Our advanced packaging technologies combined with our disaggregated architecture give us tremendous flexibility to use the process technology that best serves our customers. As an example, our datacenter GPU design, Ponte Vecchio, will now be released in late 2021 or early 2022, utilizing external and internal process technologies combined with our world-leading packaging technologies.”
Swan stated the “first Intel-based 7nm product” would be a client CPU in late 2022 or early 2023.
I suppose “first Intel-based 7nm product” can be parsed creatively if not naturally (with “Intel-based” modifying the process not the product), but it’s looking like at launch Ponte Vecchio, at least the GPU die portion, will not be built on 7nm, certainly not Intel’s node since it won’t be ready until a year later. There is speculation that Intel could shift production to TSMC (or Samsung) as part of its “contingency plans.”
Given the Department of Energy’s fixed performance and power targets for Aurora and the timeframe of late 2021 or early 2022, TSMC’s 5nm node is a likely candidate.
In the quote above, Swan references Ponte Vecchio “utilizing external and internal process technologies.” The I/O die and GPU die (and the memory stack) can be implemented on different nodes. AMD does this with its Epyc CPUs; Rome, for example, employs 7nm CPU cores and a 14nm I/O die.
From Swan again:
“Originally the architecture of Ponte Vecchio includes an I/O based die, connectivity, a GPU and some memory tiles, all kind of packaged together. That’s kind of the design of Ponte Vecchio. From the beginning, we would do some of those tiles inside and some of those tiles outside, and again leverage the packaging technology as a proof point of how do we mix and match different designs into one package. So, that was the design from the beginning… that design disaggregation gives us lots of flexibility.
“As we go forward now, we can think about whether we introduce Ponte Vecchio with… I think, I said some of those tiles are inside and outside from the beginning. Now, as we go forward, we can assess whether we swap out one of our tiles for a third-party foundry or not. Again, that’s the beauty and value of this change and design methodology that gives us much more optionality and flexibility. So, in the event there’s a process slip, we can buy something rather than make it all ourselves.”
Swan is putting a positive spin on the GPU’s disaggregated design, but swapping out the GPU compute die, as it seems like Intel will need to do, is not a minor change. The process node directly correlates to the performance and energy targets of the Ponte Vecchio GPU and by extension the Aurora system, wherein the GPU will deliver most of the performance and drive a good portion of the power demand.
Intel reported 10nm Sapphire Rapids is on track to begin shipping in the second half of 2021, and its “Intel-based 7nm datacenter CPU” is on the roadmap for the first half of 2023.
Aurora is central to the United States’ exascale plans. Its current implementation, known as Aurora21, has been positioned as the first U.S. exascale machine, although Intel has not committed publicly to a one-exaflops Linpack target.
Aurora is not the only exascale machine in development in the U.S. Oak Ridge National Lab (with HPE and AMD) is aiming to stand up the 1.5 exaflops (minimum peak) Frontier system along a contemporaneous timeframe (late 2021), and with uncertainties around Intel’s datacenter GPU execution, the odds just increased for Oak Ridge taking the lead in the United States’ exascale rollout. Lawrence Livermore National Lab (also with HPE and AMD) is looking to deploy El Capitan — slated to deliver 2 exaflops peak — one year later in late 2022. All three systems feature HPE’s Cray Shasta architecture.
Aurora was originally conceived as a pre-exascale supercomputer in 2015. The DOE CORAL contract called for a 180 petaflops (peak) machine composed of Intel Xeon Phi Knights Hill processors and second-generation OmniPath fabric technology to be stood up at Argonne in 2018. Plans were scrapped as Intel pulled back on, and eventually canceled, Phi and OmniPath development, and the contract was redefined and expanded.
Announcement of the 7nm delay contrasted with Intel’s strong second-quarter financials. The company’s second-quarter revenue of $19.7 billion was up 20 percent year-over-year. Data-centric revenue grew 34 percent, accounting for 52 percent of total revenue. Profit rose 22 percent to $5.11 billion, as reported in the Wall Street Journal, but stocks plunged 18 percent on news of the delay and lowered Q3 guidance, and at the time of this writing have not recovered.