Excellent breakdown on why the PU acronym game has gotten out of hand. The host relationship framework is particularly useful becaus it cuts through so much marketing noise. I've been working with systems where the differenc between co-processor and integrated IP determines whether a workload even makes sense, and that's never captured in benchmarks. One thing I'd add is power budget constraints as a fifth axis, especially for mobile and edge deployments where thermal limits can completely reshape which architecture wins regardless of peak throughput.
Thank you, yes, I see host relationship as a fundamental differentiator but if left out way too often in these discussions.
Completely agree on power too. It might be difficult to compare across different domains (like an AI accelerator vs a CPU) but for the same domain, power budget is a must-have.
I am a college student, and reading your Substack posts i feel more enthusiastic and curious about hardware and electronics then i ever did just by attending my college lectures. This one, too, opened my eyes to many different things. Thank you for doing these, and keep continuing.
Thank you. I really appreciate this comment. This is exactly what I intend to do with my content - getting more people curious about chips and pushing them to ask more questions. I'm glad it's heading in that direction.
Didn't expect this take on processors! Reminds me of your node naming insights. Very clever.
Thank you, yes it seems to be going the same direction as process nodes
Excellent breakdown on why the PU acronym game has gotten out of hand. The host relationship framework is particularly useful becaus it cuts through so much marketing noise. I've been working with systems where the differenc between co-processor and integrated IP determines whether a workload even makes sense, and that's never captured in benchmarks. One thing I'd add is power budget constraints as a fifth axis, especially for mobile and edge deployments where thermal limits can completely reshape which architecture wins regardless of peak throughput.
Thank you, yes, I see host relationship as a fundamental differentiator but if left out way too often in these discussions.
Completely agree on power too. It might be difficult to compare across different domains (like an AI accelerator vs a CPU) but for the same domain, power budget is a must-have.
HE IS BACK AND IT'S BETTER THAN EVER.
You're too kind. Thank you.
I am a college student, and reading your Substack posts i feel more enthusiastic and curious about hardware and electronics then i ever did just by attending my college lectures. This one, too, opened my eyes to many different things. Thank you for doing these, and keep continuing.
Thank you. I really appreciate this comment. This is exactly what I intend to do with my content - getting more people curious about chips and pushing them to ask more questions. I'm glad it's heading in that direction.
A better comeback than Akshay Khanna, iykyk.