Discussion An insightful interview with Jim Keller

[yeshua]

About CPU design and architectures, AI, the meaning of life, physics, philosophy, Moor's law, and many other things.

 

[moinmoin]

Lex Fridman:
"So there's a bunch of levels, abstraction of in organizational I can tell. And in your own vision there's a lot of brilliance that comes in it every one of those layers. Some of it is science, some was engineering, some of it's art. If you could pick favorites, what's the most important, your favorite layer? On these layers of abstraction where does the magic enter this hierarchy?"

Jim Keller:
"Uh, I don't really care."

Ouch!

 

[lobz]

About the CPU design and architectures, AI, meaning of life, physics, philosophy, Moor's law, and many other things.

Man, was this the most fun 90 minutes of my year so far or what? I really disagree with your take on this forum, but thanks for sharing this, because it was awesome to watch, I go to work an hour late now because of you, I just couldn't stand up and leave it.

 

[moinmoin]

Some notable CPU-related quotes in there (plus some unrelated thoughts from me *coughs*):
26:29 complexity gets worse, so rewrites from scratch should happen every 3-5 years. (Zen 3 may be 3 years or more after Zen 1.)
28:41 iterative improvements get into diminishing return curves, choice is the between short term disaster and long term disaster. (can't help thinking of Intel's 14nm being an example for a long term disaster.) quarter by quarter business are terrified by changing everything.
29:40 you optimize the old one while building the new one, but the marketing guys want promises that the new computer is faster on every single thing which is not really feasible
33:49 Moore's law: how small could a switching device be, currently 1000x1000x1000 atoms, gets quantum effects around 2-10 atoms, could imagine transistors as small as 10x10x10 atoms, that's a million times smaller

 

[coercitiv]

you optimize the old one while building the new one, but the marketing guys want promises that the new computer is faster on every single thing which is not really feasible

Jack of all trades, master of... everything.

 

[Det0x]

that's kinda cool © Jim Keller

 

[lobz]

Jack of all trades, master of... everything.

Well you know... anyone who reads a couple of books every week for decades...



clearly has too much time and doesn't work enough

 

[ksec]

Was searching for details on Google about this video and I end up on Anandtech again....

I am not sure how the number of 1000x1000x1000 came up because even if you count silicon atom being 0.2nm, even 22nm FinFET has a smaller number than a 1000. It would be closer to accurate if the number was taken from the smallest atom being 0.1nm, and 22 FintFET, and roundup to whole number.

Then there is height used in this equation, that is about the same as saying you can double the 22nm transistor die area by 3D Stacking another layer on top. I dont think that is a common view of Moore's law, not to mention it will still be expensive.

And this height added in equation sort of made an addition multiplier to the "million times smaller".

 

[lobz]

Was searching for details on Google about this video and I end up on Anandtech again....

I am not sure how the number of 1000x1000x1000 came up because even if you count silicon atom being 0.2nm, even 22nm FinFET has a smaller number than a 1000. It would be closer to accurate if the number was taken from the smallest atom being 0.1nm, and 22 FintFET, and roundup to whole number.

Then there is height used in this equation, that is about the same as saying you can double the 22nm transistor die area by 3D Stacking another layer on top. I dont think that is a common view of Moore's law, not to mention it will still be expensive.

And this height added in equation sort of made an addition multiplier to the "million times smaller".

He explained it pretty well, so I'm lost as to how haven't you heared that.

 

[SAAA]

Very nice interview, totally worth your time.

Was searching for details on Google about this video and I end up on Anandtech again....

I am not sure how the number of 1000x1000x1000 came up because even if you count silicon atom being 0.2nm, even 22nm FinFET has a smaller number than a 1000. It would be closer to accurate if the number was taken from the smallest atom being 0.1nm, and 22 FintFET, and roundup to whole number.

Then there is height used in this equation, that is about the same as saying you can double the 22nm transistor die area by 3D Stacking another layer on top. I dont think that is a common view of Moore's law, not to mention it will still be expensive.

And this height added in equation sort of made an addition multiplier to the "million times smaller".

Common mistake: the node name doesn't correlate with most features today. 22nm node actually has fins only 8-10 nm large btw, but they are well... fins, so thin and high.
The whole point is that while the fins might not get much smaller themselves the space around could be used better, then you could think of better approaches like nanotubes etc that are truly a few nm across at most.
We can build today transistors out of a few atoms, collecting a trillion of them in a chip, working at 5+ GHz, over a billion processors… is much harder.

 

[ksec]

Very nice interview, totally worth your time.



Common mistake: the node name doesn't correlate with most features today. 22nm node actually has fins only 8-10 nm large btw, but they are well... fins, so thin and high.
The whole point is that while the fins might not get much smaller themselves the space around could be used better, then you could think of better approaches like nanotubes etc that are truly a few nm across at most.
We can build today transistors out of a few atoms, collecting a trillion of them in a chip, working at 5+ GHz, over a billion processors… is much harder.

I am calculating using industry's MMP and CPP, not fins.

He explained it pretty well, so I'm lost as to how haven't you heared that.

Like where? That parts starts at 33:49.

 

[lobz]

I am calculating using industry's MMP and CPP, not fins.



Like where? That parts starts at 33:49.

First look up the definition of a transistor and you'll undertand instantly what he said

 

[SAAA]

I am calculating using industry's MMP and CPP, not fins.



Like where? That parts starts at 33:49.

OK I read you post wrong: the actual 1000x1000x1000 is using a simple box of atoms close in size to the transistors we have now. Basically he's saying that we are using a billion atoms to build each individual transistor when we could go down to a 1000 without say many quantum problems. Sort of.

Then you have the multiple layers part, this is different from the way they do NAND for example: he means stacking the individual parts, you can't with fins that we use now (2-3x of them per gate) as they are 2D but with pipes and tubes you could. That alone is another 4-9x increase in density.

 

[ksec]

First look up the definition of a transistor and you'll undertand instantly what he said

My bad. Wasn't thinking clearly enough.