It is foreseeable that Intel will lose the exascale race in the supercomputer against AMD, Intel confesses the defeat in diplomatic form. Officially, it is said: Intel does not want to be at the top for the time being and to be the figurehead – that is the complete opposite of the previous maxim.
Admission through the back door
Admitting defeat is not easy for Intel in this situation either, which is why the manufacturer takes the indirect route. The recently newly appointed boss Jeff McVeigh puts it diplomatically towards The Next Platform:
Honestly, we are not looking to prime, to be the front facing company, but to work with our OEM partners in the channel to make that happen. We are interested in bringing components together into balanced systems that address the customer needs. It's not just an individual component, but how do those things work well together, and then also the software stacks on top of that.
Intel doesn't want any more be the prime contractor, i.e. the first point of contact for large projects in the HPC area. That was last Intel in decades. But things haven't been going really smoothly for longer than just the last few years.
Intel's supercomputer business has been paralyzing for a long time
After its initial success, the Knights series quickly led to a dead end, although Intel relied a lot on these accelerator cards – too much. Ultimately, lighthouse projects such as Aurora collapsed as a first attempt: Intel was supposed to deliver CPUs and Knights Hill accelerators in 10 nm for a start of the supercomputer in 2018, which was announced in 2015. The big swan song followed: Knights Hill and the whole family were hired.
But Intel wanted to keep the prestigious contract, in 2019 a completely redesigned Aurora system with Xe GPUs was in prospect for 2021/2022. Aurora was thus more of a Prime project than ever before, because the system was supposed to be the first exascale supercomputer and underpin the USA's claim to top performance in the field.
But all the shifts in manufacturing technology, in the processors and also in the GPUs made the topic a nightmare for Intel, even contractual penalties are now due: In the quarterly report Q2/2021, Intel told analysts that at the end of the year it was 300 million US dollars to have to put aside for this.
Just wanted to clarify, George, the onetime charge in Q4, roughly $ 300 million, can you give a little more color on that?
[..]
Without going into too much detail, it is related to our high-performance compute activities through our Intel federal.
Intel and its partners will deliver the Aurora supercomputer in 2022. It will be some time before it is ready for use; with a bit of luck it will be by the end of 2022. The first exascale system will not come from Intel, but will be, of all things, Frontier from the small competitor AMD. Four years delay for a supercomputer is simply too much, the new CEO Pat Gelsinger wants to save himself the drama including bad PR and thus also the newly established department in the future.
AMD's Frontier on the home straight
The supercomputer Frontier is currently being implemented, the Oak Ridge Leadership Computing Facility (OLCF) confirmed this three days ago Dates for delivery (2021) and commissioning (2022). The building is currently being completely rebuilt, the old titanium supercomputer has been dismantled, and kilometers of cables and pipes for cooling have been re-laid. In the end, Frontier should deliver up to 1.5 ExaFLOPs of performance thanks to AMD Epyc and AMD Instinct. With El Capitan, the next, then again fastest supercomputer should follow within two years – again with CPUs and GPUs from AMD.