@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20240527
Episode Date: May 27, 2024- Techno-politics, China, TSMC, and ASML remote disablement - CryptoSuper500 report on Bitcoin as Specialized Supercomputing - X-ai says it will build AI supercomputer with 100,000 H100 GPUs - AI reg...ulations in EU go into effect next month - Nvidia GPUs continue strong in earnings season [audio mp3="https://orionx.net/wp-content/uploads/2024/05/HPCNB_20240527.mp3"][/audio] The post HPC News Bytes – 20240527 appeared first on OrionX.net.
Transcript
Discussion (0)
Welcome to HPC News Bites, a weekly show about important news in the world of supercomputing,
AI, and other advanced technologies.
Hi, everyone.
Welcome to HPC News Bites.
I'm Doug Black.
Hi, Shaheen.
We start once again with techno-geopolitics with a report from Bloomberg that in the event of an invasion of Taiwan by mainland
China, the ASML chip manufacturing equipment used by Taiwan Semiconductor Manufacturing Company
can be disabled remotely. We're talking about ASML's school bus-sized extreme ultraviolet
machines that TSMC uses for production of advanced HPC and AI chips. ASML has been banned
from exporting those machines to China, delivering a distinct advantage to the US-led side of our
increasingly polarized world. We had a full podcast about this, and especially after your
coverage, Doug, two years ago of a paper published by military scholars whose scenario planning included the idea that
if an invasion were to happen, then TSMC factories should literally be blown up. Six months ago,
TSMC chairman Mark Liu addressed it directly, saying any military invasion will render TSMC
factories non-operable. So it looks like they've added a remote control capability to that.
That, of course,
starts its own domino effect, since anything that can be operated remotely is a cybersecurity attack surface itself. So hopefully they have enough security attached to this remote
disablement process to prevent bad actors from disrupting production. In our recent conversation
with John Shalf of the Berkeley
Lab, we talked about specialized silicon as a way of advancing performance and energy efficiency for
supercomputing. That concept got a nice case study this past week with the release of the
Crypto Super 500 report by my OrionX colleague, Stephen Perino. The list tracks cryptocurrency mining based on the so-called
proof-of-work consensus algorithm, which is one way to validate transactions in a decentralized
way. Only Bitcoin and a handful of other coins use that consensus model, so the name CryptoSuper500
is just an allusion to Top500. Proof-of-work is based on cryptographic hashing, so it's not the same as
scientific or AI computation, but just like supercomputing, it is both very computationally
intensive and its energy use is the subject of intense focus, since it directly impacts the
economic model for miners. So in the 12th edition of the report, Stephen looks at the silicon that is used for mining and how incredibly effective they are for what they do compared to a more general purpose approach.
He's saying this decentralized, auto-balanced Bitcoin network has emerged as a great example of a specialized supercomputer demonstrating the effectiveness of the specialized silicon strategy. Very interesting. And as we expect more examples of specialized silicon in supercomputing,
I encourage our listeners to look up the AtHBC podcast episode with John Schalf two weeks ago,
who expands on this topic. There's a report from the tech news site, The Information,
that Elon Musk says he plans to build what would be
the world's biggest AI supercomputer comprised of 100,000 NVIDIA H100s.
The system will be used by Musk's startup called XAI to power the next version of its
Grok AI chatbot.
Musk said he wants to get the system running by the fall of 2025.
It would be at least four times the size of today's
biggest GPU clusters. The story brings to mind Musk's statement in 2017 that AI, quote, is the
scariest thing with the potential to destroy humanity. Comments he repeated last year, along
with calls for controlling and regulating it. I have to say, Shaheen, we can only imagine what he'd be doing with AI if he thought it was a force for good. Speaking of controlling AI,
as we covered back in March, the European Union is out ahead of the rest of the world in establishing
an AI regulatory regime. Its landmark regulations will go into force next month, having been
endorsed by the EU countries in a political
deal reached in December. This sets a potential global benchmark for a technology used in business
and everyday life, just as has been Europe's GDPR data privacy law passed in 2018.
As Reuters reported, the EU's AI Act is more comprehensive than the United States' light-touch voluntary
compliance approach or China's approach, which seems to optimize for social stability and state
control. As a summary and reminder, the EU law aims to protect people and businesses, includes
a risk taxonomy for AI practices and AI applications, and establishes clarity for AI investors for usage and adoption of AI products.
It's earnings season, and despite talk of a possible softening by the end of the year
and improving wait times to receive systems, we see continued strong demand for generative AI
and therefore hefty GPUs for learning and inference. NVIDIA said demand for H100s actually grew after they announced the B100.
That demand pulls a lot of other technologies like networking, memory, and software.
So the party continues.
Yes, anyone who's bet against NVIDIA's continued staggering rise has lost so far.
Now, it's true their quarterly revenue growth has slowed, but
that's only in comparison with the dizzying jump in revenues of a year ago. We note that in October
2022, NVIDIA stock was at $112. Today, it's above $1,000, and Jensen Wang's personal worth is more
than $90 billion. The company enabled the generative AI revolution. They're riding that
wave, and it's been amazing. Okay, that's it for this episode. Thanks so much for being with us.
HPC Newsbytes is a production of OrionX in association with InsideHPC. Shaheen Khan and
Doug Black host the show. Every episode is featured on InsideHPC.comcom and posted on orionx.net. Thank you for listening.