Always seems to be two years away, like the Aurora supercomputer at Argonne.
So basically what I'm saying is - "same" :D
After it was delayed, Intel said that 2020 was when they'd be ready. Spoiler alert: they aren't: https://www.datacenterdynamics.com/en/news/doe-confirms-auro...
We're now looking at 2022 as the new "deadline", but we know that Intel has enough clout to force a new deadline as necessary. They've already slipped two deadlines, what's the risk in slipping a 3rd time?
---------
I don't like to "kick Intel while they're down", but Aurora has been a disaster for years. That being said, I'm liking a lot of their OneAPI tech on paper at least. Maybe I'll give it a shot one day. (AVX512 + GPU supported with one compiler, in a C++-like language that could serve as a competitor to CUDA? That'd be nice... but Intel NEEDS to deliver these GPUs in time. Every delay is eating away at their reputation)
Aurora was originally slated to use Phi chips, which are an unrelated architecture to these GPUs. The delays there don't say much about problems actually getting this new architecture out. It's more that they were halfway through making a supercomputer and then started over.
I could probably pin the biggest share of the blame on 10nm problems, which are irrelevant to this architecture.
As far as this architecture goes, when they announced Aurora was switching, they announced 2021. That schedule, looking four years out for a new architecture, has only had one delay of an extra 6 months.
I doubt that.
If Xeon Phi were a relevant platform, Intel could have easily kept it... continuing to invest into the platform and make it into 7nm like the rest of Aurora's new design.
Instead, Intel chose to build a new platform from its iGPU architecture. So right there, Intel made a fundamental shift in the way they expected to build Aurora.
I don't know what kind of internal meetings Intel had to choose its (mostly untested) iGPU platform over its more established Xeon Phi line, but that's quite a dramatic change of heart.
------------
Don't get me wrong. I'm more inclined to believe in Intel's decision (they know more about their market than I do), but its still a massive shift in architecture... with a huge investment into a new software ecosystem (DPC++, OpenMP, SYCL, etc. etc.), a lot of which is largely untested in practice (DPC++ is pretty new, all else considered).
--------
> As far as this architecture goes, when they announced Aurora was switching, they announced 2021. That schedule, looking four years out for a new architecture, has only had one delay of an extra 6 months.
That's fair. But the difference between Aurora-2018 vs Aurora-2021 is huge.
It is, yes. But none of that difference reflects badly on this new product line. It just reflects badly on Intel in general.
Xe hasn't existed long enough to be "always 2 years away". It's had a pretty steady rollout.