This was a great article! Brings back memories of those days as a fingerling in the industry. One could see a lot of it for sure but hindsight always seems clearer.
I’m curious if you were working at the time and lived through all the craziness?
Look man, I barely understood SO much of it. But I like reading, and I find books my consumption of choice. You can learn so much in a single unified way. Just finish the book!
thank you - i did my best to try to get the whole thing. Saw your post that I missed part of it - but it was hard to decide exactly where to start / end it all.
really great piece of research! Best thing I have read all month. My nnagain list has some good folk on it, and also the internet history list has founding people on it going all the way back to the 60s, if ever you want to talk to original sources: https://elists.isoc.org/mailman/listinfo/internet-history
If you have not read it, the story of iridium is great: "Eccentric Orbits".
If ever you do, there are a few I know that dispute that version of iridium's history somewhat. I have been trying to figure out where starlink will go... (another mailing list of mine)
Just qq on Nvidia roadmap: "I expect each generation to be about a 3x increase in performance for 50% more price." - is this based on historicals or was there some analysis driving this?
I think that's just the rumors going around, and a bit of using A100 to H100 initial release as a guide. Would guess gross margins actually get a bit worse fwiw
I don’t think you grasped the point of this. It isn’t measly at all, but if everyone is putting down supply supporting 100s of percentage of growth when actual demand is much less, then you have a problem lol.
Excellent, Doug. It’s still not clear to me how to monetize AI such that it merits such large caped buildouts. The cost of training are high and hyperscalers are realizing this now. Check out this WSJ that few are talking about:
Great piece Doug. I hear the fiber buildout analogy a lot lately so this dive into it is very timely.
I did have one question. Where can I read more about the NVDA fan out attempts you mention. "Moreover, Nvidia is working hard to increase fan-out ability, meaning the number of GPUs you can have in a fleet." What's limiting it now and what kind of things are they doing to address it?
This was a great article! Brings back memories of those days as a fingerling in the industry. One could see a lot of it for sure but hindsight always seems clearer.
I’m curious if you were working at the time and lived through all the craziness?
Hahahaha oh no, I’m turning 30 next year. In 1999 I was in kindergarten.
Luckily for me, I love research and reading, I read almost 5 books to get up to speed.
Cool. I could not believe that you actually read Patterson and Henessey! If only more industry analysts would emulate something like that.
Look man, I barely understood SO much of it. But I like reading, and I find books my consumption of choice. You can learn so much in a single unified way. Just finish the book!
THANKS!!!! I forwarded this to multiple mailing lists I am on, like internet history an my new nnagain list which has ol farts like me on it: https://lists.bufferbloat.net/pipermail/nnagain/2023-November/000344.html
thank you - i did my best to try to get the whole thing. Saw your post that I missed part of it - but it was hard to decide exactly where to start / end it all.
really great piece of research! Best thing I have read all month. My nnagain list has some good folk on it, and also the internet history list has founding people on it going all the way back to the 60s, if ever you want to talk to original sources: https://elists.isoc.org/mailman/listinfo/internet-history
If you have not read it, the story of iridium is great: "Eccentric Orbits".
I have the book! I have been meaning to get to it, sadly lots on my list.
I would like to chat with you about what the economic effects of fixing "bufferbloat" might be like, one day.
If ever you do, there are a few I know that dispute that version of iridium's history somewhat. I have been trying to figure out where starlink will go... (another mailing list of mine)
What a breathtaking piece.
Just qq on Nvidia roadmap: "I expect each generation to be about a 3x increase in performance for 50% more price." - is this based on historicals or was there some analysis driving this?
I think that's just the rumors going around, and a bit of using A100 to H100 initial release as a guide. Would guess gross margins actually get a bit worse fwiw
"The fake statistic of the internet doubling every 100 days made it seem like the world was improving at an exponential rate"
Doubling once a year is still exponential rate and saying 57% annual growth is measly is just disingenuous
I don’t think you grasped the point of this. It isn’t measly at all, but if everyone is putting down supply supporting 100s of percentage of growth when actual demand is much less, then you have a problem lol.
Excellent, Doug. It’s still not clear to me how to monetize AI such that it merits such large caped buildouts. The cost of training are high and hyperscalers are realizing this now. Check out this WSJ that few are talking about:
https://www.wsj.com/tech/ai/ais-costly-buildup-could-make-early-products-a-hard-sell-bdd29b9f
oh yes - trust me. We are spending for a business to come :wink:
This was a fantastic read. Thank you for sharing your research.
Great piece Doug. I hear the fiber buildout analogy a lot lately so this dive into it is very timely.
I did have one question. Where can I read more about the NVDA fan out attempts you mention. "Moreover, Nvidia is working hard to increase fan-out ability, meaning the number of GPUs you can have in a fleet." What's limiting it now and what kind of things are they doing to address it?
It’s about scaling out. There’s software to split a LLM into pieces for parallel and then there’s a scaling limit of infiniabnd itself
Ah, ok. Thank you for that info
Great write up.
Great stuff Doug, thanks!
Really impressive work here
🔥👍🏽🔥