“At IEDM 2021, Intel demonstrated the world’s first experimental realization of a magnetoelectric spin-orbit (MESO) logic device at room temperature, which showed the potential manufacturability for a new type of transistor based on switching nanoscale magnets.”
I think this must be the silver bullet for zettascale by 2027…
]]>Intel has viewed reality thru the lense of marketing for far too long.
]]>Keeping it running is going to be very expensive
]]>The point must be that Intel wants to stay relevant in HPC, which means having some sort of accelerator. They must believe that Xe has merit in the HPC business, or at least that the successor will have merit. Bailing on the contract now will not only kill Intel as a prime contractor, but it will make customers unlikely to trust Intel’s product roadmap for any non-prime deals. Why would HPE/Dell/Atos/etc propose a machine based on Intel’s accelerator instead of Nvidia or AMD? Intel has to deliver a better product (price/performance) and engender trust that they can actually deliver what they promise. Intel already walked away from their previous HPC accelerator architecture, which really was the result of walking away from a prior accelerator that never made it to market. If they walk away from another, and leave the customer holding an empty bag, they will lose a lot of good customer credibility.
]]>The whole point of saying 2 EF is so they can deliver 1EF usable performance. Intel delivered nothing yet as of today – What relevant software are you referring to? Who cares about oneAPI? They only state it because now they want to sell just another GPU system which others have been doing for more than 10 years. Where is the innovation? The original interconnect is dead, DDR5 and PCIE5 are industry standard, DAOS is blah … whats the point of it than to try and sell more NVMe … if they really want to derisk they need to just get a Lustre file system and call it a day, focus on the silicon. All this nonsense while ORNL and China already deployed Exascale systems with much higher efficiency and working perfectly. Intel keeps shooting itself in the leg and asks “why , me?”. Really, we are aiming for ~50% efficiency for a leading system in 2022? Who are they kidding? If Pat knows whats good for Intel, bow out of Aurora, pay ANL its money back and let them buy a machine from HPE/Cray. Enough of this long drawn mess and dragging ANL/DOE along. Only a fool keeps doing the same thing and expects a different result – both Intel and ANL are doing the same thing over and over since the past ~6+ years and expecting a different outcome.
]]>Ben – The DOE typically does list performance requirements for a bunch of their preferred codes, typically in terms relative to prior machines. The linpack number is for the press, and for the government bureaucrats who need a single number to look at. When I say press, I don’t mean nextplatform or hpc wire, but more general interest press that doesn’t have the time or expertise to dig into anything other than “Big machine X times faster than last year’s big machine!”
]]>All this noise will have negative impact on Intel. Patrick Gelsinger should behave like a CEO instead of marketing executive or a car salesman.
]]>Optane, DAOS, CXL, DDR5, AMX, PCIE5, DSA, foveros 3D manufacturing at a scale unseen before … all leading edge features. Sapphire Rapids was sampling in Nov 2020, but this HBM version appears to have been a late addition. Quite a list of features that are unmatched by the competition.
Intel’s delivered a huge amount of software as well… all that oneAPI code.
]]>