@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20250519
Episode Date: May 19, 2025- Computex 2025 - NVLink Fusion, DGX Cloud Lepton - Intel 18A Fab Panther Lake, Gaudi 3 PCIe, Intel Arc Pro - Jack Dongarra, US Leadership in HPC - Matrix Algebra, 64 bit precision - Chiplet Alliance..., UCIe consortium - "Motherchip" vs. Motherboard [audio mp3="https://orionx.net/wp-content/uploads/2025/05/HPCNB_20250519.mp3"][/audio] The post HPC News Bytes – 20250519 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 of InsideHPC and with me is Shaheen Khan of OrionX.net.
Jensen Wong spoke last night, US time,
at the big Computex conference in Taiwan,
and with his keynote came the expected raft
of NVIDIA announcements.
And as we've noted before, NVIDIA's ambitions
are an engine that knows no rest.
Either they're announcing new technology
that pushes the AI compute envelope,
or they're forming new business alliances
that extend their reach across the AI compute envelope, or they're forming new business alliances that extend their reach across the AI compute complex.
Among the news from Computex is what NVIDIA calls
DGX Cloud Leptum, which connects NVIDIA GPUs
among various cloud partners.
That means a customer can draw on NVIDIA Blackwells
that reside on CoreWeave, Lambda, or Crusoe,
and other cloud platforms.
NVIDIA also announced NVLink Fusion, designed to speed chip-to-chip communications,
that they'll sell to other chip designers for building custom AI systems with multiple chips integrated together.
High-end networking gear for NVIDIA chips have been in short supply,
so it's not clear whether this is a way to expand
production or really an effort to make NVLink more of an industry standard or maybe both.
Computex has emerged as a major global show and an opportunity to catch up with tech trends
in Asia in general and Taiwan in particular.
So vendors also make a lot of announcements focused on that part of the world.
Nvidia announced big supercomputers to be built in Taiwan, including one at their National
Center for High Performance Computing, NCHC, and also talked up HPC quantum integration
and the use of GPUs to facilitate quantum computing research and also to emulate quantum
systems.
Intel had demos of its first chip coming out
of their new 18A factory, that'd be Panther Lake,
targeting high-performance notebooks
and other mobile processing.
They also announced new workstation GPUs called Arc Pro
for AI and graphics, and their server AI chip, Gaudi 3,
that is now also available on a PCIe card.
The conference is just starting,
so we'll catch up with other announcements next week.
Jack Dungara is one of the leading lights of our industry, and he was a special guest,
by the way, of this podcast in 2022. His recent op-ed in the publication called The Conversation
discussed the state of HPC. He called for more attention to traditional HPC, not just AI, and he warned that the nation
risks falling behind without a long-term national strategy and federal investment in several
areas including R&D, research at universities and national labs, procurement and deployment
of high-end systems, workforce development, new algorithms, and AI-HPC integration.
He also called for sustained funding models that avoid boom and bust cycles tied to one-off
milestones or geopolitical urgency. And he calls for more public-private collaboration
to bridge gaps between academic research, industry innovation, and national security needs.
By the way, Shaheen, Jack's article has created something of a stir, including a feature story
on Politico.
We've talked about public-private partnership, P3 here, and the important role it plays in
advancing society.
On the technical side, he mentioned the gap in memory speeds compared to processing speed,
the urgent need for more energy-efficient performance, the industry's growing focus
on low-precision arithmetic driven by AI, while many HPC algorithms continue to need 64-bits,
and then assembling all of the above into exascale systems and beyond. Of course, as you all know,
Jack's a pioneer of matrix algebra for
computational science, and his work includes the development of the Linpack library and its
associated benchmark that is used to rank the top 500 list of the world's most powerful supercomputers.
It turns out, computational modeling of much of nature can take advantage of matrix algebra, and that includes AI.
In his conversation with us, he mentioned ongoing work to look for ways to leverage
low-precision hardware to achieve high-precision results.
The relatively recent HPL-MXP mixed-precision benchmark is an example of that,
but many scientific algorithms require actual 64-bit hardware.
AI today seems effective with very low precision arithmetic, but as AI becomes more physics-based
to mimic reality, it is possible that it too will need 64-bits in a hurry.
At its recent Foundry conference, Intel made news in the chiplet space announcing the chiplet alliance that includes Intel Ansys ARM, Cadence, Siemens, Synopsys, and other
companies. The effort dovetails nicely with the UCIE consortium that aims to
facilitate connectivity among chiplets. The basic idea with chiplets, which are
small modular dyes, is to connect and combine them in a package
formulated into a new and more complex chip.
It's like using a Lego set for chip assembly.
The chiplet concept isn't new,
but it's caught on in recent years
in the advanced computing area
because the conventional way of developing high-end chips
has become complex and enormously expensive.
We've talked about a time when a chip designer could select the capabilities they want from a
menu of chiplets or tiles, add them to, say, a shopping cart, and click a button and it is all
incorporated and produced in a custom chip. That sort of mass customization could be the way to
address the needs of large and small markets. You could go another step and include those basic capabilities
into your cluster configurator, let's say.
So you're not selecting a brand of chip,
but multiple brands of chiplets
possibly produced in different technologies
according to your specific need,
and out comes a customized cluster
and the software layers that make your apps take advantage of
what the hardware provides. That sure looks like a worthy stop on the roadmap to technology in
Nirvana. These ideas have been around for a number of years and a lot of people are working on them.
Chiplets were a solution to the ever-growing size of chips. They enable super large chips and also
reduce complexity by compartmentalizing things.
Chips reside on a motherboard that traditionally had a lot of other chips and other electronics on
it. But as chips get larger and larger, they leave less and less for the motherboard. The Apple M1
chip already pointed in that direction in 2020. If the whole system is really just one big chip,
you kind of have a mother chip instead
of a mother board. And that really changes system design and packaging. So the key now is to have a
large enough menu of IP and chiplets and tiles and a cross industry standard helps with that.
And then to use it for small and large chips so everyone can participate and the ecosystem is
larger.
Very interesting and exciting to watch.
All right.
That's it for this episode.
Thank you all for being with us.
HPC News Bytes is a production of OrionX in association with Inside HPC.
Shaheen Khan and Doug Black host the show.
Every episode is featured on insidehpc.com and posted on orionx.net.
Thank you for listening.