Comment by wood_spirit
Comment by wood_spirit 16 hours ago
Unlike railroads and fibre, all the best compute in 2025 will be lacklustre in 2027. It won’t retain much value in the same way as the infrastructure of previous bubbles did?
Comment by wood_spirit 16 hours ago
Unlike railroads and fibre, all the best compute in 2025 will be lacklustre in 2027. It won’t retain much value in the same way as the infrastructure of previous bubbles did?
Unless you care about FLOP/Watt, which big players definitely do.
> Unlike railroads and fibre, all the best compute in 2025 will be lacklustre in 2027.
I definitely don't think compute is anything like railroads and fibre, but I'm not so sure compute will continue it's efficiency gains of the past. Power consumption for these chips is climbing fast, lots of gains are from better hardware support for 8bit/4bit precision, I believe yields are getting harder to achieve as things get much smaller.
Betting against compute getting better/cheaper/faster is probably a bad idea, but fundamental improvements I think will be a lot slower over the next decade as shrinking gets a lot harder.
>> Unlike railroads and fibre, all the best compute in 2025 will be lacklustre in 2027.
> I definitely don't think compute is anything like railroads and fibre, but I'm not so sure compute will continue it's efficiency gains of the past. Power consumption for these chips is climbing fast, lots of gains are from better hardware support for 8bit/4bit precision, I believe yields are getting harder to achieve as things get much smaller.
I'm no expert, buy my understanding is that as feature sizes shrink, semiconductors become more prone to failure over time. Those GPUs probably aren't going to all fry themselves in two years, but even if GPUs stagnate, chip longevity may limit the medium/long term value of the (massive) investment.
> changing 2027 to 2030 doesn't make the math much better
Could you show me?
Early turbines didn't last that long. Even modern ones are only rated for a few decades.
there is a difference between a few decades and half a decade though? Or his time in general accelerated so much that it's basically very similar
“Within Parsons' lifetime, the generating capacity of a [steam turbine] unit was scaled up by about 10,000 times” [1].
For comparison, Moore’s law (at 2 years per doubling) scales 4 orders of magnitude in about 27 years. That’s roughly the lifetime of a modern steam turbine [2]. In actuality, Parsons lived 77 years [3], implying a 13% growth rate, so doubling every 6 versus 2 years. But within the same order of magnitude.
[1] https://en.m.wikipedia.org/wiki/Steam_turbine
[2] https://alliedpg.com/latest-articles/life-extension-strategi... 30 years
[3] https://en.m.wikipedia.org/wiki/Charles_Algernon_Parsons
Unfortunately the chips themselves probably won’t physically last much longer than that under the workloads they are being put to. So, yes, they won’t be totally obsolete as technology in 2028, but they may still have to be replaced.
Yeah - I think that the extremely fast depreciation just due to wear and use on GPUs is pretty unappreciated right now. So you've spent 300 mil on a brand new data center - congrats - you'll need to pay off that loan and somehow raise another 100 mil to actually maintain that capacity for three years based on chip replacement alone.
There is an absolute glut of cheap compute available right now due to VC and other funds dumping into the industry (take advantage of it while it exists!) but I'm pretty sure Wall St. will balk when they realize the continued costs of maintaining that compute and look at the revenue that expenditure is generating. People think of chips as a piece of infrastructure - you buy a personal computer and it'll keep chugging for a decade without issue in most case - but GPUs are essentially consumables - they're an input to producing the compute a data center sells that needs constant restocking - rather than a one-time investment.
There are some nuances there.
- Most big tech companies are investing in data centers using operating cash flow, not levering it
- The hyperscalers have in recent years been tweaking the depreciation schedules of regular cloud compute assets (extending them), so there's a push and a pull going on for CPU vs GPU depreciation
- I don't think anyone who knows how to do fundamental analysis expects any asset to "keep chugging for a decade without issue" unless it's explicitly rated to do so (like e.g. a solar panel). All assets have depreciation schedules, GPUs are just shorter than average, and I don't think this is a big mystery to big money on Wall St
Do we actually know how they're degrading? Are there still Pascals out there? If not, is it because they actual broke or because of poor performance? I understand it's tempting to say near 100% workload for multiple years = fast degradation, but what are the actual stats? Are you talking specifically about the actual compute chip or the whole compute system -- I know there's a big difference now with the systems Nvidia is selling. How long do typical Intel/AMD CPU server chips last? My impression is a long time.
If we're talking about the whole compute system like a gb200, is there a particular component that breaks first? How hard are they to refurbish, if that particular component breaks? I'm guessing they didn't have repairability in mind, but I also know these "chips" are much more than chips now so there's probably some modularity if it's not the chip itself failing.
I watch a GPU repair guy and its interesting to see the different failure modes...
* memory IC failure
* power delivery component failure
* dead core
* cracked BGA solder joints on core
* damaged PCB due to sag
These issues are compounded by
* huge power consumption and heat output of core and memory, compared to system CPU/memory
* physical size of core leads to more potential for solder joint fracture due to thermal expansion/contraction
* everything needs to fit in PCIe card form factor
* memory and core not socketed, if one fails (or supporting circuitry on the PCB fails) then either expensive repair or the card becomes scrap
* some vendors have cards with design flaws which lead to early failure
* sometimes poor application of thermal paste/pads at factory (eg, only half of core making contact
* and, in my experience in aquiring 4-5 year old GPUs to build gaming PCs with (to sell), almost without fail the thermal paste has dried up and the card is thermal throttling
Believe it or not, the GPUs from bitcoin farms are often the most reliable.
Since they were run 24/7, there was rarely the kind of heat stress that kills cards (heating and cooling cycles).
Could AI providers follow the same strategy? Just throw any spare inference capacity at something to make sure the GPUs are running 24/7, whether that's model training, crypto mining, protein folding, a "spot market" for non-time-sensitive/async inference workloads, or something else entirely.
I'm not sure.
Number of cycles that goes through silicon matters, but what matters most really are temperature and electrical shocks.
If the GPUs are stable, at low temperature they can be at full load for years. There are servers out there up from decades and decades.
Yep, we are (unfortunately) still running on railroad infrastructure built a century ago. The amortization periods on that spending is ridiculously long.
Effectively every single H100 in existence now will be e-waste in 5 years or less. Not exactly railroad infrastructure here, or even dark fiber.
> Effectively every single H100 in existence now will be e-waste in 5 years or less.
This is definitely not true, the A100 came out just over 5 years ago and still goes for low five figures used on eBay.
> Effectively every single H100 in existence now will be e-waste in 5 years or less.
This remains to be seen. H100 is 3 years old now, and is still the workhorse of all the major AI shops. When there's something that is obviously better for training, these are still going to be used for inference.
If what you say is true, you could find a A100 for cheap/free right now. But check out the prices.
Yeah, I can rent an A100 server for roughly the same price as what the electricity would cost me.
> Yep, we are (unfortunately) still running on railroad infrastructure built a century ago.
That which survived, at least. A whole lot of rail infrastructure was not viable and soon became waste of its own. There was, at one time, ten rail lines around my parts, operated by six different railway companies. Only one of them remains fully intact to this day. One other line retained a short section that is still standing, which is now being used for car storage, but was mostly dismantled. The rest are completely gone.
When we look back in 100 years, the total amortization cost for the "winner" won't look so bad. The “picks and axes” (i.e. H100s) that soon wore down, but were needed to build the grander vision won't even be a second thought in hindsight.
> That which survived, at least. A whole lot of rail infrastructure was not viable and soon became waste of its own. There was, at one time, ten rail lines around my parts, operated by six different railway companies. Only one of them remains fully intact to this day. One other line retained a short section that is still standing, which is now being used for car storage, but was mostly dismantled. The rest are completely gone.
How long did it take for 9 out of 10 of those rail lines to become nonviable? If they lasted (say) 50 years instead of 100, because that much rail capacity was (say) obsoleted by the advent of cars and trucks, that's still pretty good.
> How long did it take for 9 out of 10 of those rail lines to become nonviable?
Records from the time are few and far between, but, from what I can tell, it looks like they likely weren't ever actually viable.
The records do show that the railways were profitable for a short while, but it seems only because the government paid for the infrastructure. If they had to incur the capital expenditure themselves, the math doesn't look like it would math.
Imagine where the LLM businesses would be if the government paid for all the R&D and training costs!
Much like LLMs. There are approximately 10 reasonable players giving it a go, and, unless this whole AI thing goes away, never to be seen again, it is likely that one of them will still be around in 100 years.
H100s are effectively consumables used in the construction of the metaphorical rail. The actual rail lines had their own fare share of necessary tools that retained little to no residual value after use as well. This isn't anything unique.
H100s being thought of as consumables is keen - it much better to analogize the H100s to coal and chip manufacturer the mine owner - than to think of them as rails. They are impermanent and need constant upkeep and replacement - they are not one time costs that you build as infra and forget about.
At the rate they are throwing obstacles at the promised subway which they got rid of the 3rd Ave El for maybe his/her grandkids will finish the trip.
> Yep, we are (unfortunately) still running on railroad infrastructure built a century ago. The amortization periods on that spending is ridiculously long.
Are we? I was under the impression that the tracks degraded due to stresses like heat/rain/etc. and had to be replaced periodically.
I am really digging the railroad analogies in this discussion! There are some striking similarities if you do the right mappings and timeframe transformations.
I am an avid rail-to-trail cycler and more recently a student of the history of the rail industry. The result was my realization that the ultimate benefit to society and to me personally is the existence of these amazing outdoor recreation venues. Here in Western PA we have many hundreds of miles of rail-to-trail. My recent realization is that it would be totally impossible for our modern society to create these trails today. They were built with blood, sweat, tears and much dynamite - and not a single thought towards environmental impact studies. I estimate that only ten percent of the rail lines built around here are still used for rail. Another ten percent have become recreational trails. That percent continues to rise as more abandoned rail lines transition to recreational use. Here in Western PA we add a couple dozen miles every year.
After reading this very interesting discussion, I've come to believe that the AI arms race is mainly just transferring capital into the pockets of the tool vendors - just as was the case with the railroads. The NVidia chips will be amortized over 10 years and the models over perhaps 2 years. Neither has any lasting value. So the analogy to rail is things like dynamite and rolling stock. What in AI will maintain value? I think the data center physical plants, power plants and transmission networks will maintain their value longer. I think the exabytes of training data will maintain their value even longer.
What will become the equivalent of rail-to-trail? I doubt that any of the laborers or capitalists building rail lines had foreseen that their ultimate value to society would be that people like me could enjoy a bike ride. What are the now unforeseen long-term benefit to society of this AI investment boom?
Rail consolidated over 100 years into just a handful of firms in North America, and my understanding is that these firms are well-run and fairly profitable. I expect a much more rapid shakeout and consolidation to happen in AI. And I'm putting my money on the winners being Apple first and Google second.
Another analogy I just thought of - the question of will the AI models eventually run on big-iron or in ballpoint pens. It is similar to the dichotomy of large-scale vs miniaturized nuclear power sources in Asimov's Foundation series (a core and memorable theme of the book that I haven't seen in the TV series).
Neato. How’s that 1999 era laptop? Because 25 year old trains are still running and 25 year old train track is still almost new. It’s not the same and you know it.
Except they behave less like shrewd investors and more like bandwagon jumpers looking to buy influence or get rich quick. Crypto, Twitter, ridesharing, office sharing and now AI. None of these have been the future of business.
Business looks a lot like what it has throughout history. Building physical transport infrastructure, trade links, improving agricultural and manufacturing productivity and investing in military advancements. In the latter respect, countries like Turkey and Iran are decades ahead of Saudi in terms of building internal security capacity with drone tech for example.
Agreed - I don’t think they are particularly brilliant as a category. Hereditary kleptocracy has limits.
But… I don’t think there’s an example in modern history of the this much capital moving around based on whim.
The “bet on red” mentality has produced some odd leaders with absolute authority in their domain. One of the most influential figures on the US government claims to believe that he is saving society from the antichrist. Another thinks he’s the protagonist in a sci-fi novel.
We have the madness of monarchy with modern weapons and power. Yikes.
Exactly: when was the last time you used ChatGPT-3.5? Its value deprecated to zero after, what, two-and-a-half years? (And the Nvidia chips used to train it have barely retained any value either)
The financials here are so ugly: you have to light truckloads of money on fire forever just to jog in place.
I would think that it's more like a general codebase - even if after 2.5 years, 95% percent of the lines were rewritten, and even if the whole thing was rewritten in a different language, there is no point in time at which its value diminished, as you arguably couldn't have built the new version without all the knowledge (and institutional knowledge) from the older version.
I rejoined an previous employer of mine, someone everyone here knows ... and I found that half their networking equipment is still being maintained by code I wrote in 2012-2014. It has not been rewritten. Hell, I rewrote a few parts that badly needed it despite joining another part of the company.
If OpenAI is a public company today, I would bet almost anything that it'd be a $1+ trillion company immediately on opening day.
A really did few days ago gpt-3.5-fast is a great model for certain tasks and cost wise via the API. Lots of solutions being built on the today’s latest are for tomorrow’s legacy model — if it works just pin the version.
> money on fire forever just to jog in place.
Why?
I don't see why these companies can't just stop training at some point. Unless you're saying the cost of inference is unsustainable?
I can envision a future where ChatGPT stops getting new SOTA models, and all future models are built for enterprise or people willing to pay a lot of money for high ROI use cases.
We don't need better models for the vast majority of chats taking place today E.g. kids using it for help with homework - are today's models really not good enough?
>I don't see why these companies can't just stop training at some point.
Because training isn't just about making brand new models with better capabilities, it's also about updating old models to stay current with new information. Even the most sophisticated present-day model with a knowledge cutoff date of 2025 would be severely crippled by 2027 and utterly useless by 2030.
Unless there is some breakthrough that lets existing models cheaply incrementally update their weights to add new information, I don't see any way around this.
There is no evidence that RAG delivers equivalent performance to retraining on new data. Merely having information in the context window is very different from having it baked into the model weights. This approach would also degrade with time, as more and more information would have to be incorporated into the context window the further away you are from the knowledge cutoff date.
They aren't. They are obsequious. This is much worse than it seems at first glance, and you can tell it is a big deal because a lot of effort going into training the new models is to mitigate it.
But is it a bit like a game of musical chairs?
At some point the AI becomes good enough, and if you're not sitting in a chair at the time, you're not going to be the next Google.
Not necessarily? That assumes that the first "good enough" model is a defensible moat - i.e., the first ones to get there becomes the sole purveyors of the Good AI.
In practice that hasn't borne out. You can download and run open weight models now that are spitting distance to state-of-the-art, and open weight models are at best a few months behind the proprietary stuff.
And even within the realm of proprietary models no player can maintain a lead. Any advances are rapidly matched by the other players.
More likely at some point the AI becomes "good enough"... and every single player will also get a "good enough" AI shortly thereafter. There doesn't seem like there's a scenario where any player can afford to stop setting cash on fire and start making money.
Perhaps the first thing the owners ask the first true AGI is “how do I dominate the world?” and the AGI outlines how to stop any competitor getting AGI..?
The A100 came out 5.5 years ago and is still the staple for many AI/ML workloads. Even AI hardware just doesn’t depreciate that quickly.