Thread Rating:
  • 0 Vote(s) - 0 Average
  • 1
  • 2
  • 3
  • 4
  • 5
Jim Keller Interview
#1
https://www.techpowerup.com/270197/jim-k...om-scratch
Quote:Jim Keller on Lex Fridman's AI Podcast shed some light on his thoughts on the microprocessor design fundamentals as he sees them. In a hour-and-a-half-long interview, he approaches Moore's Law and its much lauded - and ubiquitously repeated - death, as well as the need for both iterative and zero-point microprocessor design requirements.

Mr. Keller approaches the usual microprocessor design loop, where a company develops a new design from scratch and then looks at the most fundamental way of adding performance. Usually, he says, easy 10% performance increments can be found by simply looking at a design and increasing execution units - increase a buffer here, increase a cache over there, put in another add processor on this part of the pipeline. However, he also speaks of how this process in itself is limiting, inasmuch as doing this often will eventually guide processor designs towards a bottleneck and the diminishing returns problem, where any more additions made to the design don't seem to increase performance - mostly just adding complexity, area and power requirements, and generally convoluting a given design.
...
Mr. Keller approaches the usual microprocessor design loop, where a company develops a new design from scratch and then looks at the most fundamental way of adding performance. Usually, he says, easy 10% performance increments can be found by simply looking at a design and increasing execution units - increase a buffer here, increase a cache over there, put in another add processor on this part of the pipeline. However, he also speaks of how this process in itself is limiting, inasmuch as doing this often will eventually guide processor designs towards a bottleneck and the diminishing returns problem, where any more additions made to the design don't seem to increase performance - mostly just adding complexity, area and power requirements, and generally convoluting a given design.
...
Of course, then comes Intel, which Keller himself describes as having a microprocessor development mindset that's closer to a 10-year sustained designed rather than the 3-5 year development schedule for a new architecture he favors. Interestingly, in the podcast, Jim Keller approaches this microarchitecture mindset on Intel from a short-term and long-term disaster perspective. According to him, repeating and refining a recipe (like Intel did many years with their Core architecture [author's side-note]) is the most efficient way to go about it: incrementally improving a design, saving money and taking a low-risk approach to processor development, albeit threatened by the diminishing returns equation we mentioned earlier.

The problem, according to Mr. Keller, is that managing quarter to quarter means that there is fear in hitting a short-term disaster with a rewrite from scratch; companies thus look to "milk" every ounce of profit from a previous design by incrementally improving it. However, this primes companies to hit a long-term disaster, much like we see today with Intel (it's not a disaster for a multi-billion dollar company like Intel, but you get the point): its architecture, which didn't go through a from-scratch design phase for years, was superseded by AMD's new Zen design and its iterations.
...
Closing up this article, which doesn't aim to be a summation of the AI Podcast, but aims to highlight some interesting tidbits present there, are Mr. Keller's thoughts on team management for a technological project. According to him, he sees the existence of abstraction layers in microprocessor design teams. Where a team of 10 humans works well together, and a team of up to 100 people may be able to function properly under a single supervision, any more than that and teams have to be divided, with organizational boundaries having to be set - and here too appears the diminishing returns equation. Jim Keller says some very interesting things regarding this, as in, that humans in general aren't getting smarter - so there is a fundamental limit to how much "processing power" you can have in a team working on a set project, considering team and size and communication capacity caps that derive from the fact that we are, well, humans.
...
TL;DR: Jim Keller is clearly an extremely accomplished microprocessor designer, but also a project leader, and has very clear ideas regarding the industry and his field of work. You should read the entire article and then move onto the podcast.
Reply


Forum Jump:


Users browsing this thread: 1 Guest(s)