News Intel's CEO Admits to and Details Companies Three Biggest Mistakes

[PiranhaTech]

Some of those might be "hindsight is 20/20". Larrabee was seen as kind-of bad, though it was smart of Intel to at least research it. Intel's GPU really didn't take off until they hired Raja Koduri in November 2017.

Arc is doing pretty well. It's hard to get GPUs up to that level in the time they had.

 

[bit_user]

Some of those might be "hindsight is 20/20". Larrabee was seen as kind-of bad, though it was smart of Intel to at least research it. Intel's GPU really didn't take off until they hired Raja Koduri in November 2017.

It's hard to know how much of Alchemist was because of Koduri or in spite of him. I got the sense that his departure from AMD was something of a mutual decision, and I certainly think he wore out his welcome at Intel.

It was widely known that Intel was working on a dGPU prior to bringing Koduri on board. He certainly will have had a few insights about scaling that their existing GPU team wouldn't, but he certainly doesn't deserve all the credit for what Alchemist has accomplished, nor do I think he deserves all of the blame for its shortcomings.

Probably the dumbest thing Intel did was trying to build the incredibly ambitious HPC-focused Ponte Vecchio at the same time as ARC. Had they not done that, maybe Alchemist would've launched during the end of the crypto boom and been a lot more financially successful for them. That would've done a lot to ensure its future.

 

[TerryLaze]

Probably the dumbest thing Intel did was trying to build the incredibly ambitious HPC-focused Ponte Vecchio at the same time as ARC.

You are the one that is confused here...
Intel made ponte vecchio for the super computer(s) mainly to get aurora to the amount of computing they asked for, that's the one that was and still is the main focus, that's what makes them the money, intel made ARC at the time time to take advantage of what they learned from and had to make for ponte anyway.
Just like optane ssds were just a side project to ease the cost of financing the optane dimm modules.

Intel's First Discrete GPU For Supercomputers Is 7nm Ponte Vecchio Leveraging Foveros

A 7nm chiplet architecture along with PCI Express 5-base CXL fabric will enable Intel's first discrete Xe GPU for HPC and AI applications.

hothardware.com

 

[bit_user]

Intel made ponte vecchio for the super computer(s) mainly to get aurora to the amount of computing they asked for, that's the one that was and still is the main focus, that's what makes them the money, intel made ARC at the time time to take advantage of what they learned from and had to make for ponte anyway.

They definitely did not make Ponte Vecchio just for a single customer. They made it as a replacement for Xeon Phi - which was a good decision, but too late.

Ponte Vecchio was, itself, way too ambitious. That surely had a lot to do with why it was so late. Whoever ultimately green-lit such an architecture messed up big time. It should've had 1/10th as many chiplets, for instance.

Regardless of whether that was the main focus or ARC - trying to do both at once was clearly a bad decision. It seems like they got a reality-check, when it came to Rialto Bridge, which they cancelled. And then Falcon Shores got delayed + simplified, probably in the same ways & for the same reasons as Ponte Vecchio should've.