Supercomputers for AI - H+ Weekly - Issue #431
This week - Chinese LLM chatbots are available to the public; lab-grown fruits; BBC talks about longevity; AI-powered drones race better than human champions; and more!
AI is once again the hottest trend in the tech world, and everyone, from giant corporations to startups, wants to get a piece of the AI pie. But to enter this game, you need to have access to the computing power required to run AI models. One option to acquire that computing power is to use cloud AI services, a quick and cost-effective way to get started. The second option is to build your own computing power. It is a way more expensive option, but for some companies, building a dedicated AI supercomputer makes sense.
The newest AI supercomputer is at Tesla. The new cluster, worth over $300 million, boasts 10,000 Nvidia H100 compute GPUs, offering a peak performance of 340 FP64 PFLOPS for technical computing and 39.58 INT8 ExaFLOPS for AI applications. Tesla already has a smaller cluster consisting of 5,760 Nvidia A100 GPUs that came online in June 2012. The primary purpose of the new cluster is to accelerate the training for AI used in Tesla’s Full Self-Driving technology, but I wouldn't be surprised if researchers at xAI (Musk’s own AI research lab) also get access to it.
Tesla is also in the process of building Dojo, an AI supercomputer powered by in-house-designed chips, with the purpose of training AI models for self-driving cars. The company is investing $1 billion into this project, projected to be completed in 2024. According to Elon Musk, Tesla will spend over $2 billion in 2023 alone on training compute and will do so again in 2024.
Tesla’s newest AI supercomputer is just one of many planned or already operational supercomputers designed specifically for AI computations.
In May, Meta completed the second phase of building the Research SuperCluster (RSC). When finished, RSC will offer almost 5 exaflops of computing power through 16,000 NVIDIA A100 GPUs.
Microsoft, whose Azure cloud services power OpenAI’s ChatGPT, is also building a new supercomputer solely for AI calculations, powered by “tens of thousands of NVIDIA A100 and H100 GPUs.”
Earlier this year, Inflection announced they are building the most powerful AI supercomputer, consisting of 22,000 NVIDIA H100 GPUs capable of 22 exaFLOPS in the 16-bit precision mode. This would give Inflection’s supercomputer the second-place spot on the Top500 list and put it very close to the top spot, even though it's optimized for AI calculations.
You might have noticed that almost all of these supercomputers use Nvidia’s GPUs. This is good news for Nvidia, which is currently having its best year in its 30-year history, but it's not good news for everyone else trying to get any of the highly sought-after Nvidia GPUs. One solution to this issue is to not use Nvidia’s GPUs and design custom chips instead. That is the path Tesla is taking with Dojo. Another company that has opted for custom chips is Google. Since 2015, Google has been using custom Tensor Processing Units (TPUs) in their data centres for AI computations. Today, TPUs account for more than 90% of the computing power used to train AI models at Google.
There are benefits to creating custom chips to power AI supercomputers. One benefit is independence from suppliers such as Nvidia, which is especially advantageous now when everyone is fighting to get the newest H100 GPUs, often paying well above the original price tag just to secure these cards. Another benefit is the ability to customize the chip to the company’s specific needs. Nvidia designs their chips to be general-purpose, suitable for AI calculations as well as engineering and scientific applications. Going with custom chips means the chips can be optimized to perform a single task exceptionally well. According to Google, an AI system using the fourth generation of TPUs is up to 1.7 times faster and 1.9 times more power-efficient than a similar system based on Nvidia's A100 chip.
However, building a supercomputer with custom AI chips is not a viable option for everyone. The required scale and significantly higher upfront investment are hurdles that only tech giants can afford to overcome.
Whichever path the AI companies take to build their own supercomputers, they believe it's an investment that will massively pay off. With potentially billions, if not trillions, of dollars available from the generative AI revolution, investing hundreds of millions in infrastructure appears to be a small price to pay.
From H+ Weekly
Becoming a paid subscriber is the best way to support the newsletter.
If you enjoy and find value in my writing, please hit the like button and share your thoughts in the comments. Share the newsletter with someone who will enjoy it, too.
You can also buy me a coffee if you enjoy my work.
🦾 More than a human
The Transformative, Alarming Power of Gene Editing
In 2018, He Jiankui shocked the world by revealing his experiment that resulted in the birth of the first genetically modified humans. Soon after that, he was sentenced to three years in prison for “illegal medical practices”. In 2022, He Jiankui was released from prison and quickly resumed his efforts at gene editing. This article from The New Yorker details He Jiankui’s life and the environment both in China and the US that led to his historical revelation. He Jiankui also shared that he is back working on altering human genes and that he is building a new lab in China to cure as many as five genetic diseases in two to three years, with therapy for Duchenne muscular dystrophy being the first on the list.
‘Complete’ models of human embryos created from stem cells in lab
Researchers have developed "complete" models of human embryos using stem cells in a lab and grown them outside the womb. While not identical to real human embryos, these model embryos could help scientists better understand early human development, uncover causes of miscarriages, create organs for transplantation from patients' skin cells, and assess the effects of medicines on real human embryos.
The Immortals
The Immortals is a new, 10-episode podcast series from BBC focusing on longevity. The first couple of episodes are about how blood and plasma transfusion could reverse the effects of ageing and follow the experiments Bryan Johnson does on himself, his son and his father. After that, the remaining episodes introduce the concept of “longevity escape velocity” alongside other billionaires who believe that a future in which death is optional is closer than people think.
🧠 Artificial Intelligence
Chinese ChatGPT alternatives just got approved for the general public
Chinese authorities have granted approval for the initial batch of chatbots powered by large language models to be released to the public. Amongst the selected chatbots was Baidu’s Ernie Bot, the first Chinese ChatGPT rival, which was revealed in March to a mix of excitement and disappointment. Other chatbots approved for public release were Baidu’s Doubao, Zidong Taichu 2.0 from the Institute of Automation at the Chinese Academy of Sciences and chatbots from smaller startups - SenseTime, Baichuan Intelligent Technology, Zhipu AI and MiniMax.
‘I hope I’m wrong’: the co-founder of DeepMind on how AI threatens to reshape life as we know it
In this interview, Mustafa Suleyman, co-founder of DeepMind and now founder of Inflection, shares his hopes and fears of AI. He acknowledges the potential AI has to rapidly reduce the price of achieving any goal. But “any goal” also includes those with negative outcomes. The example used in this conversation was biotech. On the plus side, biotech holds the promise to cure diseases and to create “radical abundance”, as Suleyman puts it. But the same technology can also produce a deadly virus using a DNA synthesiser.
The 100 Most Influential People in AI 2023
Time has published a list of the 100 most influential people in AI in 2023. The list includes names that you might expect to see, like Sam Altman, Demis Hassabis or Elon Musk, but also includes many founders of the new wave of AI startups, investors, academics, researchers, policymakers and artists using AI in their work.
▶️ Andrew Ng: Opportunities in AI (36:54)
Andrew Ng sees the opportunities in more niche, tailored solutions that previously would require a large team of highly trained engineers to create. These projects, which he calls “5 million dollar projects”, will be built with small teams on top of generative AI and data-centric AI systems (systems where you insert data with a minimum amount of code). Ng argues that the coming decade will be like the SaaS explosion in the 2010s but powered by AI. He also makes a note that people inside tech are focusing too much on tools (an example, in this case, could be all the projects trying to replicate ChatGPT) instead of focusing on real-life applications that impact more people outside the tech.
Rise of AI newsbots shakes up India's media landscape
Odisha TV, an Indian broadcasting network, has introduced an AI news presenter called Lisa. Lisa's role is to present news updates on digital platforms, offer horoscope readings, and provide weather and sports updates. India has joined other countries, including China (which pioneered the use of an AI news anchor in 2018), Indonesia, Malaysia, Taiwan, and Kuwait in deploying AI avatars to deliver news on television.
🤖 Robotics
The USAF Pairs Piloted Jets With AI Drones
Many countries are researching the concept of a “loyal wingman” - a swarm of drones working together with a human fighter pilot. The US is the closest to moving beyond the research phase. That means having an operational concept, a chosen contractor, a sizable production line and a line of funding. The USAF is asking Congress for $5.8 billion over five years. If the program is successful, US fighter pilots will be flying combat missions together with AI-controlled drones to aid them in a dogfight or to rush ahead into enemy airspace without risking the pilot’s life or the jet.
Doctors couldn't operate on my tumour, but this robot did — and it may have saved my life
Glenn Deir shares how a surgical robot Da Vinci saved his life. Deir only hope was surgery to remove tumours from his mouth. The problem was that no surgeon was able to perform the delicate operation. But Da Vinci was, and it successfully removed the tumours.
AI-powered drone beats human champion pilots
Developed by researchers at the University of Zurich, the Swift AI won 15 out of 25 drone races against world champions and clocked the fastest lap on a course where drones reach speeds of 50mph (80km/h) and endure accelerations up to 5g, enough to make many people black out. “Our result marks the first time that a robot powered by AI has beaten a human champion in a real physical sport designed for and by humans,” said Elia Kaufmann, a researcher who helped to develop Swift.
A Japanese city is installing robots in classrooms so kids who are worried about attending school can tune in remotely
Kumamoto, a city in southwest Japan, plans to install robots in some of its classrooms to help home-based pupils feel more included and less anxious. Japan has been experiencing a rise in the number of truant schoolchildren, a trend potentially accelerated by the COVID-19 pandemic. These two-meter-tall robots are equipped with cameras and microphones for two-way communication and can move freely around the school.
🧬 Biotechnology
In the face of climate change and food insecurity, New Zealand considers lab-grown fruit
Researchers from New Zealand are applying a similar technique used to grow meat in a lab to grow fruits. The plan is to grow fruit tissue from plant cells that will one day taste, smell and feel like real fruit. The program, which started 18 months ago, focuses on cells from blueberries, apples, cherries, feijoas, peaches, nectarines and grapes, but the researchers warn that the end goal of harvesting something that is nutritional and enjoyable to eat is some years away and might not be attainable at all. But if successful, lab-grown fruits could help secure the food supply that might be disrupted due to climate change.
Synthego and bit.bio join forces on genetic engineering platform for cell therapy
Synthego, a company specializing in genome engineering, has joined forces with bit.bio, a synthetic biology company, to create a platform implementing synthetic circuitry in cells for therapeutic purposes. The goal of this collaboration is to advance the genetic engineering sector for cell-based therapies, improving their safety, efficiency, and global accessibility for countless patients.
H+ Weekly sheds light on the bleeding edge of technology and how advancements in AI, robotics, and biotech can usher in abundance, expand humanity's horizons, and redefine what it means to be human.
A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!
You can follow H+ Weekly on Twitter and on LinkedIn.
Thank you for reading and see you next Friday!
Despite the fact that Nvidia chips are highly sought after, isn't everything still actually made by TSMC?