Do you know about why AXG was essentially nerfed and Raja was fired? Supposedly he was lying to the rest of the company and after ARC gpus were finally released a year behind schedule AND having software that was clearly only half-baked, Intel finally decided to pull the plug on funding AXG. All the delays, uncompetitive products, and missed deadlines led AXG to get cut hard.
I don’t have first hand knowledge, but from the technical deep dives I’ve seen on Alchemist, there were a few basic mistakes made on architecture that someone with Raja’s experience should have prevented from happening.
I would need to search hard for the technical articles talking about this, but the basic mistakes were along the lines of — if you take an iGPU and scale it up to a ‘big GPU’ you end up with bottlenecks in some areas that didn’t exist when the GPU was small. They built Alchemist like an iGPU scaled up but didn’t address those bottlenecks, so they’re using a lot more silicon for the same performance than if they made a few changes back in design. You might expect a CA new to GPUs to make some mistakes like this, but Raja having sign off should have caught this early on.
Alchemist under Raja also went for too many software features at once (“check every box that Nvidia and AMD has”) instead of focusing on making a certain set of features work really well. You only get the first impressions “benefit” once, and a more seasoned leader (Raja or under him) should have looked more closely at quality vs. quantity here. (ex: It’s really cool that Intel offered the ability to OC your GPU.. except it didn’t work at first, and the software was very buggy. Maybe this feature should have been skipped for 6 months after launch).
Also, I think he didn’t have enough contingency plans for when things did run late. I saw a few articles that some of the software was being developed in Russia for example, and that delayed maturity of the SW. With the obvious Russia stuff going on after Crimea in 2014, if Intel was developing a critical new product SW stack there, they should have been ready for possible geopolitical scenarios.
You may also see a pattern with Raja of overhyping stuff and under delivering — as he did a few times (“Poor Volta”) when the Vega architecture (which also had a ROPs or some other limitation preventing performance from scaling beyond a certain point) launched.
I don’t know the guy, but from 1 million miles away, he looked a little over his head with the scope of the project. It’s possible that Intel’s culture set him up to fail completely too of course, but when you’re paid that kind of money you really should figure out a way to succeed..
(Note - he had a herculean task to try to launch a 3rd major discrete GPU company, though I think Intel several years ago was in a good shape to do so — engineering talent galore, a cost model that would have allowed sharing with existing iGPUs, and fabs that could be re-used cheaply to get some GPUs out. It’s hard to tell if anyone could have succeeded, but we do know Alchemist under Raja came out very late and underperformed).
AMD is not in Intel's rearview mirror. Nvidia owns AI. Dozens of new chip companies are circling Intel like vultures picking at the carcass. Worst of all, the semiconductor industry is not as healthy as it seems.
