Suspected Intel Ponte Vecchio Xe GPUs appear before Trade Committee – Computer – News

I must owe you the answer, but the history of Larrabee and this Aurora (planned for 2018 !!!) is interesting:

2009: Larrabee becomes canceled. The technique is partnered with the Single-chip Cloud Computer project continued in the “Knights” series (Xeon Phi).
2011 First Intel “manycore” commercial product on the market: Knights corner, 22nm FinFET (FinFET was completely new at the time).
2013: Successor “Knights Landing” (14nm) announced; is going to run in the fastest supercomputer in the world.
2014: Successor “Knights Hill” (10nm) announced.
2015: Aurora announced, this would use the 10nm Knight Hill. Planned delivery date: 2018.
2017: Intel admits, that 10nm has failed epically; the Knights Hill platform gets the knife in the back; and it is clear that “Aurora ready in 2018” is a myth.

Meanwhile, there is a new reality:

• If Aurora is going to be delayed for three years, then the performance has to be a factor higher to be able to participate,
• Within HPC there has been a shift from deductive tasks (general rules from which special cases are calculated, eg “natural laws” as general rules and “nuclear explosion” as special case) to inductive tasks (“Machine Learning”: From many special cases via inference try to derive a general rule).

What is a possible problem, however: 10nm GPUs depended on a technology called “Contact Over Active Gate” (COAG).

2018:

• Intel indicates that their first 10nm has been shipped; it will be the only Cannon Lake SKU (normally there are tens, if not hundreds) ever to be sold. The GPU in this chip has had a knife in the back; it is there but is turned off !!!
• Most Likely Reason: COAG is an epic failure.
• Result: It is practically impossible for Intel to make a good-performing GPU on the first 10nm process.

Subsequently, someone has apparently decided that the 7nm process should be used for the Aurora for 2021; and that the technology needs to be adapted from the “GPGU” target originally planned in 2015 towards “Machine Learning”.

In other words, due to the epic failure of Intel 10nm and the 3 year delay; the task that the Aurora computer will perform has changed drastically. AI (Machine Learning) has become the newest (and most important?) Task for exa-scale supercomputers; so Xe will be adapted to that.

What may not be the same for everyone (or maybe I’m wrong?) Is that in the Greek alphabet the letter “Xi” (Vecchio Xe !!!) after comes after the letter Phi. In this sense, Xe is thus the successor to Phi; while Phi was Larrabee’s successor.

2020: The first gossip appear onlinethat the Aurora is a “flaming mess”. If so, the question is whether 2021 will be achieved; as described by T.net above. If not met, the Frontier (AMD Epyc) may become the first Exascale supercomputer.

[Reactie gewijzigd door kidde op 25 april 2020 19:39]

–