I am no expert in Intel, but in my view, Gelsinger lost the faith of many by being unrealistically optimistic. Of course a CEO needs to be optimistic, but he promised (in 2021) zettaflop systems by 2027 (the worst example I remember). Did anyone believe that could happen?

His over-optimism gave the whole "5 nodes in 4 years" supposed path to leadership a weird flavor, like it must be somehow a bit of a con even if it gets technically achieved.

I mean, if I thought I had a plan to be the guy who saves Intel from it's own mistakes, I'd be optimistic too.

Also, I looked into the claim when he had said it, apparently he was being intentionally misleading about it, and the press tried to ask what he meant: he was speaking tensor performance on future enterprise Arc card products at datacenter scale, ie, AI bait.

In early 2021, Nvidia's compute flagship was the A100, 19.5 FP32 MAD TFLOPs, but the misleading number they quote in marketing is the tensor performance of 312 TFLOPs of FP16 accumulates. That would be about 3.2 million of these at tensor perf.

Skipping H series, in late Nov of last year, their new flagship is the B200. 124 FP32 MAD TFLOPs, 2250 tensor FP16 accumulate TFLOPs. That is now 445k cards to reach zettascale if using tensor cores. You won't be fitting ~1400 GPU-laden machines in a single datacenter, but the number is becoming more manageable.

They improved, in 3.5 years, 7.2x.

Lets say Nvidia does this again. 3.5 years, again, would put you in early 2028, and they manage another 7.2x win: that could be 62k cards across 7.7k. That absolutely is doable in a single datacenter.

The problem is, and this is where the prediction actually falls apart, not that its impossible: We don't know what future Arc cards look like, nor enterprise ones. Battlemage is an improvement over Alchemist, so the tech *is moving forwards at, but either Celestial or Druid was supposed to introduce the enterprise compute card variants, but that seems to be dead, and no indication either of those lines will even see the light of day now. The new CEO seems to be hard set on making Xe for iGPU only.

I can't find any hard numbers on Intel's tensor units, but apparently they're actually competitive. I can find the normal FP32 MAD numbers, and it ends up that Intel is 13.5w per TFLOP and Nvidia is 8 and both companies have equal efficiency in transistor usage. Assuming Intel made a B200 competitor, and assuming the higher power usage is due to voltage (Intel B series voltage is similar to Series 40's voltages, which is a lot higher than equivalent enterprise/pro series cards), Intel could be making a card that's somewhere in the ballpark as 2/3 as good for the same power usage.

So, in the end, yes, I don't agree with his claims of future Zettascale at Intel by 2027. I don't think he was wrong for the industry as a whole, however. If he would have said, say, 2030, I don't think we would be discussing this, that certainly would have been doable if he was at the helm and they kept doubling down on Arc every gen and everything went according to plan.