There's no arguing that Keller is a smart guy, but he doesn't design an entire CPU architecture himself. If you're desperate and say "Okay, we are hiring the smartest guy we can find to build our new CPU, and give everything he needs to make it happen", then perhaps you get the AMD64, Zen or the A4 and A5. If you try to just dump a smart guy into a team as just another engineer, maybe you get nothing, like Intel.
Perhaps AMD, who already knew him, just gave Keller everything he need to build a team that can deliver on a new architecture, even when he's no longer there. Same with Apple. Intel on the other hand may have been unwilling to grant Keller the same level of autonomy and control. Then it also makes sense that he would leave Intel, for personal reasons, those being: "I can't work here, they won't let me do my job".
If the company is growing, there are new X-of-Y positions to move up to.
If the company is stable or shrinking, people start watching out for their own careers with knives out.
AMD possibly avoided this because of size & realization of what needed to be done. Intel's too big & old: I would be very surprised if they weren't much more internally resistant to that sort of change.
And one can only deal with your colleagues throwing up brick wall after brick wall on every bit of minutiae for so long, as least if you're talented enough to have other options.
PS: Nice nickname. Read the book from Th. Mann - over a period of 7 years I think.
I think we are already starting to see fruit of his work. Intel doesn't need Jim Keller for CPU uArch design. Intel has had their uArch roadmap ready, and they were the best in the Industry if it wasn't for the 10nm delay. They also have work in the pipeline all being held back by their process node.
Jim described it in one of his interview ( Sorry I spent 10 min but couldn't find the source, so I may have remembered it wrong ) about not having process node held back your chip design, where he has experience in doing so in AMD and Apple. Being flexible enough to back port your design should anything happen as Plan B. Where previously Intel was just keep waiting for the process guys to fix it. That in itself is a huge workflow changes. It is hard to imagine the amount of work required to push this through especially with all the internal politics at Intel.
And Intel is at least looking at TSMC / alternative paths for some of their product lineup now ( Gaming Focused Large- Die Size GPU ) . Whether that is decided or not is unclear. But at least we have Rocket Lake launching soon which is sort of a half baked Willow Cove ( Used in Tiger Lake ) ported back to 14nm on Desktop. And we have Sapphire Rapid as well as other product roadmap hinting at multiple node ( Shown in Investor meetings notes ). That is at least showing Intel has changed their Internal design to be flexible enough in case of another 10nm like fiasco. And I think Jim Keller has some credit in this transition.
That is of course, having flexible design still doesn't fix their problem if TSMC is 2 years ahead of Intel in leading edge node design, volume and cost. And as I have repeatedly stated, Intel's problem is not design, but their business Model. And It would not surprise me if TSMC have shipped more 5nm wafers last year ( 2020 ) than Intel's entire 10nm production history since 2017.
Just let that sink in for a bit.
Why? Because I never ever see any articles talking about any other interesting employees. Every single time it's Jim Keller.
I'm sure he's good, his interview with Lex Fridman shows that he's knowledgeable and creative, but there's no way he's as exclusive a major force as the media portrays him.
For example, one could round up as many scientists as they could find in 1900, but there is no number that would guarantee the progress made in theoretical physics by someone like Einstein alone.
You're probably going to see a whole lot more of this sort of thing given the limits to process scaling. Keeping things simple and backwardly compatible made sense when you could just throw more transistors at the problem. Now you're seeing more and more specialized circuitry that software people are just going to have to deal with.
Compare that to the non-scalable SIMD instructions that mean you have to rewrite your code to take advantage of them and resultingly people don't bother to use them at all.
AMX allocates a huge die area to GEMM functionality that gets used a lot less in real numerics than you'd gather from reading a linear algebra textbook.
There are other approaches to the problems the industry faces other than 'fill up the die with registers that will never be used', nvidia and apple are going that way and that is why they are succeeding and Intel is failing.
These statements almost seem contradictory. What if instead of "not being able to right that ship", it is instead an example to the contrary?
Was he exclusively working on "that ship" ... wait, I thought you said "chip".
If his successors applied his management style, mobile phones would have an Intel processor in them.
The person can be a key motivator to bring the team forward, or reach decisions they wouldn't otherwise have taken.
Intel started out with memory. They never left and are on by far the cutting edge with memory tech. In particular Optane NVM DIMMs are so fast they basically define a new layer in the performance/cache hierarchy. Intel might see a shift in their focus over time away from CPUs to chalcogenide based persistent memory, where it seems they have held the lead for some time now.