AI Hardware w/ Jim Keller
HTML-код
- Опубликовано: 30 сен 2024
- “Our mission is to help you solve your problem in a way that is super cost-effective and available to as many people as possible.”
Listen to Tenstorrent CEO Jim Keller discuss why AI models are graphs and how we built our software and hardware to give you a fundamental advantage at this year's AI Hardware Summit.
Our international team is always looking for new talent, learn more at www.tenstorrent.com/careers
#ai #artificialintelligence #machinelearning #compute #JimKeller #tenstorrent #hardware 
Наука
![[24] Jim Keller Interview 2: AI Hardware on PCIe Cards](http://i.ytimg.com/vi/fOUB73dZXEo/mqdefault.jpg)
Babe wake up, new Jim Keller video just dropped, and it has good audio this time!
The fact that we got a good mic on Jim Keller is a miracle!
Dude haha right tho!?
200 iq and a billion dollar hardware company.
Worse microphones than a russian csgo player.
😂😂😂 I was about to ask Sennheiser to provide the guy with a microphone for the interest of mankind. Spot-on comment!
Jim is a *living* legend
As a software guy I'm down to Jim trashing me all day
"In Matrix multiply, there's a little term which is: You do n cubed operations for n squared data."
How can I invest in Tenstorrent?
"Everything is graph. Hardware guys make graph go fast. Yay hardware." -- Jim is grug confirmed.
One of the greatest and most unique minds of our generation.
I really want to write a compiler that targets this hardware.
When will I be able to purchase it?
Compilers already exist, it's just many many risc-v cores connected to a fabric...
@@kayakMike1000 Yes, but programmers love to re-invent the wheel. Often (naively?) hoping to do it better, and later discovering unforeseen complexity.
Jokes aside, There seems to be some interesting opportunities for exploiting the parallelism in this architecture. Somewhat analogous to how out of order execution exploits the partial ordering of a regular single threaded program.
You won’t be able to write the compiler for it, you give it a graph
Write an efficient place and route algorithm. Wait! You can use AI to optimize that !!
Jim Keller is a personal hero of mine.
Me too
Yep
Where's the buy button?
Where are the retail cards and are any comparable in price/performance vs consumer Nvidia GPUs? I guess everything is CUDA dependent these days...
Moore's law isn't dead!
You dont have an office in Warsaw. Big mistake.
Sounds like something George Hotz would like.
So what are the numbers how much compute and memory per chip ? how big of a model i can lay out in a small box ?
How can I invest in Tenstorrent?
im looking for it too
You probably can't, unless they ever decide to go public, which I hope they don't for their own sake
Trying to solve optimized graph problems (with Buda/PyBuda) sounds like it can run into P vs NP (a major unsolved problem in computer science). Bioinformaticians sometimes hit that boundary when mapping problems to graphs.
Is this a computing-in-memory architecture？
From the diagram he showed does it look like compute in memory?
@@regicoco904 no
Search for "the cost per transistor stopped dropping a decade ago at 28nm", Moore's Law was originally about minimum component cost, not just raw transistor count. I'm not saying it's "dead" or anything, but we're not on the same pace as before.
pytorch docker image for TT AI server?!?
Anyone have any advice on becoming a chip designer? (Books, topics, app notes, classes etc) Thanks in advance
Try building Ben Eaters 8-bit breadboard computer by following his video series. He sells a kit on his website.
You'll also understand how to build the simplest possible CPU.
If actually follow along you will also get excessive experience points in the wire stripping skill.
So you may want to consider emulating the design with PCB etching rather than a breadboard and wires.
I would search what classes are in an Electrical Engineering bachelors degree, then mimic that. Seek out the text books, find Udemy and Coresera courses, read the books, watch online lectures. If I wanted to do what you're seeking to do, that would be my approach.
You have to start with a Digital Design course and you can use "Digital Design with introduction to verilog HDL by Morris Mano", at the end of the course and as you follow along with the textbook you will learn the concept Hardware Descriptive Languages(HDLs) which are languages used to describe the relation between digital components and behavior of digital circuits.
the next course you have to study is computer organization and architecture. some universities teach this course into two separate courses, one named computer organization and the other one is architecture. you learn in this course the internal structure of basic CPU, like ALU and how components inside the CPU communicate with each other and how the CPU communicate with other parts like RAM for instance.
while you are studying, try to implement and simulate what you learn using Verilog for example as a strong foundation in it is a must.
This video should have millions of views. Jim Keller is a legend!
W
This looks similar to the Tesla AI chip, maybe a bit more generalized.
Jim made that one
Oops, posted to wrong Jim Keller interview!….. Fantastic interview of an amazing man! Great video. Re discussion of regulation. He is correct, you can not regulate technology. But regulation is critical. Transparency and openness is needed for society to flourish , not just technology and industry. Regulation needs to be ethics based. Human lives are already impacted by machine based decisions. Organizations using these machines must be required to have mechanisms for meaningful and timely appeal. Already too many examples of people cut off from life preserving care through algorithmic errors. Those are easily documented. How many others in other domains go undetected? The opportunity cost to society is too great to not require fundamental regulatory guardrails. See blogpost AI and Power: The Ethical Challenges of Automation, Centralization, and Scale by Rachel Thomas.
why include the nonsensical brain analogy when it amounts to nothing more than "i can draw an abstract representation of what i'm doing that resembles this image of a rat's brain"?
that's ur best argument for why people should give you money? presumably?
Why complain about a single slide in a 30 minute presentation? If that's all you can take away from it then Jim Keller doesn't need your money
Why make this comment when it amounts to nothing more than "I don't understand anything said there for it's wrong"?
@@asandax6 jim keller is a smart and well-educated man. he is not mistaken. he is deliberately attempting to mislead his audience.
when i see bs im going to call it out. it's ok if you dont like that.
@@ChrisJackson-js8rd He is making an analogy of the mechanism of Neural Networks which is structured kinda like how the brain is structured. What misleading claim did he make?
@@asandax6 the rats brain has no distinction between training and inference. the product he's selling does. behaviourally these two systems will be in no way analogous. despite the visual similarity of the graphical depictions