@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20240219
Episode Date: February 19, 2024- AI Security, Safety, Containment, Governance - U.S. AI Safety Institute Consortium (AISIC) - Running CUDA Apps On ROCm - Exascale software linear algebra software libary - Chinese chipmaker Loongson... tapes out 16-core DragonChain-powered CPU - Photonic computing chip at University of Pennsylvania [audio mp3="https://orionx.net/wp-content/uploads/2024/02/HPCNB_20240219.mp3"][/audio] The post HPC News Bytes – 20240219 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.
Hey, everybody.
Welcome to HPC News Bites.
I'm Doug Black.
Hi, Shaheen.
Let's start today with a focus on AI security and safety.
An academic paper was released with the title of Guidelines for
Artificial Intelligence Containment, Discussing Ways to Prevent AI Systems from Going Rogue.
Then, a 78-page white paper published in the UK proposed that AI be controlled by controlling
compute. OpenAI, whose founder is currently on a mission to raise $7 trillion with a T to build dozens of AI chip
fabs around the world, is a contributor to this paper. There's some irony here that reminds us
of a year ago when Elon Musk and others proposed to regulate or pause AI development. We're also
going all in on AI R&D. We know everything with a digital heartbeat is also a target for cybersecurity
attacks and malware. IBM also announced measures to safeguard AI applications. Microsoft said its
AI was used by hackers from the usual suspect countries. Governments around the world are
hurrying up to mitigate this risk in a wider way than the usual military intelligence,
critical infrastructure,
and the like. And the Biden administration this week announced a consortium dedicated to AI safety
and responsible AI. They set it up within NIST, National Institute for Science and Technology.
This is the department that used to be called National Bureau of Standards, and where a lot
of work has been done over the years
on cybersecurity, including quantum safe cryptography.
Let's move to software.
There's a massive amount of software
that runs on NVIDIA GPUs
and depends on NVIDIA's proprietary parallel programming
and communication interface called CUDA, C-U-D-A,
that stands for Compute Unified Device Architecture, introduced in 2006 to
facilitate splitting computer tasks between CPUs and GPUs that were attached to them as
coprocessors.
So CUDA manages the back and forth and provides the ability to harness the dense compute elements
that are part of a GPU.
And it took a long time to get to its current polished
state. So what if there was software that could take your existing CUDA code as is and run it
at speed on another GPU? Well, that's what Intel is doing as part of its OneAPI and AMD is doing
as part of its Rokkim development environments, and they're closing in.
Right. These software ports have become easier for developers, but still require non-trivial work.
It's complex since each model of a chip, even those with the same architecture,
has performance variations, and it gets more complicated across architectures.
It emerged this week that AMD has funded such a, quote, drop-in piece of software to enable CUDA programs to run on AMD's Rokom platform.
Pharonix has the details of what it describes as a skunkworks project over the past two years to let applications run at the library level.
So, in principle, the same application links to a different set of native libraries and is then good to go.
We've received a post this week on InsideHPC's National Lab News portal regarding Project SLATE,
which stands for Software for Linear Algebra Targeting Exascale.
It's a performance, portable, and GPU-accelerated dense linear algebra library written in C++ for exascale systems.
There's a lot of interesting work being done on the software side by the DOE labs and the exascale computing project.
Some of it focused on taking the programming done for the three new exascale systems
and extending it to more mainstream HPC systems,
broadening the benefits and applicability of that development work.
Yeah, it really helps to actually have an exascale computer at your disposal.
And it's great to see this.
Big challenge of scaling is a super efficient foundation.
Scaling is only interesting when performance grows faster than complexity.
But because the various parts of hardware and software don't scale linearly, small inefficiencies that you could ignore because they don't matter at low scale can become prohibitive at high scale.
So you really need to start right to get to extreme scale.
And if you've done it, you can then use it at smaller systems with equal efficiency.
Now, dense linear algebra is typically well behaved, which means it's GPU-friendly, it's communications-friendly,
but getting it right is crucial to enable other apps and codes.
This is also relevant to Top500 and its associated benchmarks, HPL, HPL-MXP, the mixed precision
benchmark that used to be called HPL-AI, and the toughest of the set, HPCG.
On the hardware front, there are new developments in advanced chip design.
Chinese chipmaker Longsun has reportedly taped out the 16-core Dragon Chain-powered CPU
with 64-core coming.
According to a story in My Drivers, the server processor will rival AMD Zen 3 CPUs.
Yeah, separately, we've talked about photonics for computing,
and it came up
when Professor Karen Bergman of Columbia University, one of the leading authorities on photonics,
was a guest of the show, one of our most popular episodes, by the way, so you should go look it up.
That discussion was not so optimistic about the use of photonics for computing,
but there is a hunger for new ways of handling AI that provide speed and energy efficiency.
Along those lines, there is a report in the phys.org publication that University of Pennsylvania engineers have developed a new chip that uses light waves, photonics, rather than electricity.
Definitely intriguing, and I hope they make it work.
All right, that's it for this episode.
Thank you all 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.com
and posted on OrionX.net.
Thank you for listening.