At their AI ‘festival’ (GTC) this week Nvidia announced their next generation super chip; the RTX 50-series or Blackwell (named after the statistician and game theorist David Blackwell). The basic numbers are predictably huge… 208 billion transistors, 4-30x more performance than the current generation Hopper chip, plus huge boosts to the memory, connectivity and density across the supporting Blackwell datacentre architecture. What’s impressive is that in 2020 we saw the first exaFLOP super computer (an entire building sized machine pumping out more than 1 quintillion floating point operations per second) and with Blackwell, a single rack the size of the fridge in your kitchen is exa-scale. The ultimate expression of this new chip generation is what Nvidia calls the “AI Factory”. 32,000 GB200 super GPUs packed into a datacentre with more internal high-speed bandwidth than the ENTIRE Internet and 600x more power than that ultimate supercomputer from way back in… 2020.
Takeaways: With maybe 100-200 of these AI factories to be deployed over the next few years, what does this mean for us humans? We are about to realise that we have been living in a world with a miniscule amount of computational power. Like our medieval ancestors who would not have been able to imagine the future abundance of food, healthcare, travel and entertainment that would be born from the first industrial revolution, we now face the challenge of adjusting to a future of truly ‘abundant compute’. The difference here is that this will happen in the next few years, not over several centuries.
We’re also going to see a lot more AIs… GPT-4 took 100 days to train back in 2022, it will now take <3 days in an AI Factory. We’ll also see bigger smarter models, Nvidia predict that Blackwell kit will run models of 20+ trillion parameters (nearer to 1/4 the size of a human brain rather than the 1/100th we see today).
And with these AIs and ‘accelerated’ computing capabilities Nvidia’s conference also gives a glimpse of the applications; from new intelligent enterprise software, to autonomous industry, world simulators where time is sped up 1000x to train robots or design new chemistry and biology, to the brains of humanoid machines themselves.
Whilst Blackwell suggests we’re not running out of ‘space at the bottom’ to reference Richard Feynman’s assertion in the 1950s that the nano-scale would provide huge scope for machine miniaturisation, we may be nearing the buffers ‘at the top’. Power grids are straining as we build new datacentres, and the ability to plug in these AI factories, not fabricate the chips, will become the constraining factor. This explains Amazon’s recent step in buying a nuclear power station (at 960MW this could in theory power around 15 Blackwell AI factories). This post provides an excellent analysis of the power challenges, and a concerning note for us Europeans. We are far behind in building AI capable datacentres. Today a simple chatbot doesn’t require nearby compute, but in the next few years, more advanced high bandwidth AI applications may thrive on low latency. The future will be one of intense competition over sovereign AI factory capacity as much as the ingenuity to make use of it.
Our newly updated ExoBrain AI compute model predicts we’ll see a 10x growth by this time next year and 50x by the end of 2026. This is hardware only and does not include the huge gains that are coming from more efficient models and software techniques. Where today a power user might be running 4-5 model instances on and off during the day, in a few short years we’ll each be able to run 2,000-3,000, the vast majority of which will be autonomous working tirelessly to advance every facet of our physical and virtual lives.
Some interesting points:
- We assume that other firms will start to compete more effectively with Nvidia, but outside of some radical breakthrough, that can’t happen fast enough to shift their compute domination and production deployment head start. Nvidia is likely to become the most valuable company in the world.
- This is of course assuming there are no disruptions to the chip output from the single TSMC facility in Taiwan where ALL of these chips come from today (and that the data centres can be built or upgraded, but we are not estimating an impossible rate in this regard).
- No matter how fast competitors ramp up or how many H100s are shipped, from 2025 and beyond the vast majority if AI computations in the cycle (and potentially the emergence of AGI) will happen on a Blackwell, by shear dint of its relative power.
- Its a cliché in the AI world to see a “you are here” on an exponential curve. But as of this week’s news, this is a real curve. You are here..
