Jim Keller - Moore's Law in the age of AI Chips
Вставка
- Опубліковано 4 лют 2025
- For more talks and to view corresponding slides, go to scaledml.org, select [media archive].
Presented at the 5th Annual Scaled Machine Learning Conference 2020
Venue: Computer History Museum
scaledml.org | #scaledml2020
I graduated computer hardware / ECE in 2007 and went into software because of exactly this erroneous thought 'what jobs will there be in 10 years when Moore's law is dead'. If any young person is watching, listen to Jim.
What are some cool places/companies in this space that you know of?
companies only very recently pay HW engineers close to SW, so still might not have been a bad idea to go into Software, unless you're super interested in designing hardware.
Jim Keller really loves Comic Sans since he seems to use it in every presentation
*My takeaways：*
1. Compute models: Scalar, Vector, Matrix and Spatial 13:50
2. Data models: global memory vs local memory 16:15
3. It is extremely difficult to write SIMD program 18:25
4. GPU is not SIMD 19:31, GPU running a vector of scalar programs that share a global memory 20:08
5. AI chips 21:20
those are not takeaways, those are time stamps
There was a talk from a guy who used to work for Intel, he mentioned that he was benchmarking a vector of scalar program doing the same thing as the SIMD program optimized by Intel's internal compiler people, and the vector of scalar beats it because it does not have to wait around each other like the SIMD does.
I learned more from this video than any hype out there from individual companies. I mean, I'm more excited from this video for future technologies than anything I've seen in the past.
Jim Keller? that's a like from me.
Man I wish I had another 1,000 likes to pile on that. How many of the top emerging CPUs did he NOT have a hand in?
Apple's M1? Swift and Cyclone were his babies. Ryzen? Check. If Tesla ever makes self-driving happen, it will have been on his architecture. If Microsoft follows Apple and hands AMD a fat stack of cash for K12, that will be a twofer for 'ole Keller. I'm excited to see his second AI architecture come out.
Comp Arch 101
Jim Keller made ,me remember my entire comparch course from over 14+ years ago
Very insightful! Thanks!
Isnt 2^60 more like an exa byte?
Great talk thanks
I'd like Jim to revisit this since its ironic for him to defend Moore's Law until leaving Intel
You should watch the lex fridman podcast
I really need some Korean subtitles.. Can somebody help me..?
Google how to download generated English subtitles, then write an automated script to Google translate each line into Korean.
If you turn on automatic subtitles, you can also select "auto-translate" for Korean.
If you don't know this guy he may come across like a wheat farmer from Alabama. Then you realize he is one of the smartest people in the world.
0:33 technology optimism talk by who?
Riva Tez
@@jakovmarinvezic2364 thank you
What a great talk. Can't wait to see Jim do to Intel what he did to AMD.
not anymore :(
@@kmolnardaniel intel is done
Jim Keller can only help those who are willing to listen and do what he says needs to be done. No doubt he's helped the architecture team, but I don't think Intel was as willing as AMD, Apple, or Tesla to listen.
@@Jaker788 Yup.
He's not at Intel... they burned that chance
Comic sans
anyone betting against intel in recent news is betting against keller's brilliance
This did not age well.
@@M.-.D what do you mean? Intel has continued to set record revenues and beat expectations, and products bearing raja and keller's impact won't even release until later this year. EASY money.
@@christopherchang6378 it might have to do with Keller leaving Intel.
@@christopherchang6378 found the intel shareholder
@@ttb1513 yah, cause his bro-in-law was dying. his stints are usually longer than a couple of years anyway.
If Moore's Law is not dead, then why can't you buy a 64x more transistors and 64x faster CPU than i7-920 from 2008 (which costed only $284 for 4 cores, 8 threads and 8 MB of L3 Cache at 2.8 GHz all-core with good OC capabilities)? That would be 46 784 million transistors. Far more than 10900K which costs $488 and has a useless iGPU.
Core 2 Quad Q6600 came out in Q1 2007 for $851 and was coming down in price with every passing month. Core i7-920 came out in Q4 2008 for $284 and is ~68% faster at maximum OC on both.
These are the changes I was accustomed to. If you compare $999 i7-5960X from Q3 2014 and the newest $237 i5-10600KF, they have very similar performance. So less improvement and innovation in about 6 years than earlier in less than 2 years. Very frustrating. So little has changed since i7-920 was released back in 2008... 10900K has only 2.5x more cache.
@@Kynareth6 yeah I had X3440 which could be overclocked to 3.8GHz. it was great for over 6 years.
2990wx is >20B transistors, 3990x is >32B transistors. Personal PC chip has stalled but workstations and servers are still going. Back in the core 2 times server & pc chips were on same line
@@teamrocket9684 Intel workstation CPU in 2010 - 6 cores. In 2020 - 18 cores. Only 3x for the same money after 10 years. This is also stagnation. Threadripper 3990X costs 4x much as 6 cores in 2010 so does not count.
its not dead, its just slowed down!
A.I is the schizophrenia of computing
who said bums cant learn to code