Suk Lee and Micheal chang did not retire ... they were laid off according to my taiwan friends at intel
LBT has done some amazing things on financing and culture change at Intel. But IFS execution to goals since 2021 is not good.
What are the changes LBT has made in Intel that you think are so amazing?
True, they were both were part of the big lay-off last August. Suk turned 65 in December so he is officially retired after 40+ years. I hope he comes back but it is highly unlikely.
Kevin was there not too long before LiBu became CEO, and he was moved out of LiBu direct report in one of the 1st reorg. You can argue he was not given enough time to own the business of IFS. Or he knew something we don't know, and decide QualComm is a better opportunity for him
Recency bias perhaps Dan? He is probably the best CEO in the last decade. Meanwhile, no way to compare him to the 1st 3 CEO. It's like people arguing who is GOAT for Sports
don't disagree on Phil, but just saying for sports, things are complicated.
I can't speak to this one. I haven't seen much difference in the design groups.
Intel has had several, none of which were compelling or successful. The Gaudi inference chips have not done especially well. I suppose you cloud say that Intel client CPUs with NPUs are an AI strategy, but it seems like an also-ran with Apple, Qualcomm, and AMD. Their stand-out AI software strategy is... OpenVINO? Tiber and Geti? Do most people in the technical community even know what they are? So I would answer your question with "Not really." Shameful. But then LBT says he going to start GPU development? That strikes me as several years late.
I don't have any visibility into the inner-workings anymore, but I'm wondering what the "focus" looks like. New org structures with lead engineers calling the shots, and not directors and VPs who mostly go to meetings? Has program management been reorganized to stop being a process-for-process's-sake function? I'd be thrilled if I heard this was the case.
This makes me more hopeful. I still think the most valuable future of Intel is as a foundry, combined with a back-end development business like Broadcom's.
The TACO effect at its worst, or best, or at least most visible. At least we know LBT can manage Trump to some degree.
and it sucks cause OpenVino is used to compress models also OneAPI is nice i have used it and i am telling it from my personal experience i use the compiler and profiler part a decent bit w
the issue is he has to compete with Robert Noyce/ Gordon Moor / Andy Grove so i guess he can go down as the 4 the best CEO if IFS is a sucess
what do you mean by this?
I seem to be an outlier lately; I don't consider OneAPI an AI software strategy no matter how much I admire it as a technology. Perhaps you can convince me.
I think the program management function (especially at Intel, but it seems the same no matter which large company I look at) is little more than a slideware and meeting generator which adds to costs and disempowers engineering leadership. IMO, it was created because in big companies senior management became too non-technical to judge engineering execution methods, roadmaps, and progress, engineering leaders were not groomed to have big-picture thinking, and senior management wanted a check and balance mechanism with a separate reporting structure to make themselves feel more confident about engineering development progress without actually having to be in the active management chain. I'd eliminate the program management function completely, and assign the engineering leaders to get the big picture or get out.
well i can try cause oneAPI is used a lot especially for model compression vLLM uses it basically it allows you to compress model into GGUF file and than you can run them with llama.cpp you can use SYCL to write cross platform code that you can run on other GPUs.
Interesting. I didn't know that. One API still looks tangential to a strict AI software strategy, but I do like the one OneAPI strategy. IMO, it is some of Intel's best software work.
I'm not convinced GPU architecture is the right way to go. Nvidia just fell into it, so to speak, because GPUs are SIMD processors and AI is largely about dataflow. I like Google's approach with TPUs for efficiency, at least for training. Inference is a different animal. The new reconfigurable hardware architectures are certainly technically interesting (SambaNova, NextSilicon), though their marketing tactics - comparing what they will have in the future with what Nvidia and AMD have today - makes me want more proof of a sustainable advantage. SambaNova's tiered memory approach, for example, also sounds a lot like the Cerebras off-chip memory strategy. I also wonder if Intel would have chosen SambaNova as their technology partner if LBT wasn't an investor and chairman of the BoD.
the SIMD is mostly don't getting used for AI it's the Matmul hardware or the tensor core part that is getting used. TPU is just a large matmul accelerator (gross simplification)
Agreed it would result in unnecessary cost and issues unless you plan to only run a particular model for like long term it doesn't make sense and the issue is models are getting good continuously so you are outdated.
