@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20260323

Episode Date: March 23, 2026

- GTC postview, list of announcements - GPUs and China - DoE $293m funding RFA for technology challenges - Volume on HPC-AI-Quantum co-design, call for contributions - Turing Award goes to QKD inventors, quantum cryptography pioneers [audio mp3="https://orionx.net/wp-content/uploads/2026/03/HPCNB_20260323.mp3"][/audio] The post HPC News Bytes – 20260323 appeared first on OrionX.net.

Welcome to HPC Newsbytes, a weekly show about important news in the world of supercomputing, AI, quantum computing, and other advanced technologies. Hi, everyone. Welcome to HBC Newsbytes. I'm Doug Black. And with me, of course, is Shaheen Khan. Invidia's GTC conference was last week. Shaheen, you were there in person. But by all accounts, it seemed to have been as expected. There were about 30,000 attendees in person from around the world, 20% more than last year, which indicates obviously a continued fascination, fixation, focus on AI, but also continued domination of the space by Nvidia. In his keynote, CEO Jensen Wong projected a $1 trillion demand for AI data center infrastructure by 2027,

reinforcing the scale and urgency of the AI buildout. Agentic AI and physical AI were front and center as they represented, the next phase of growth and what NVIDIA clearly sees as drivers of GPU sales. Strong financial momentum continues with revenue surpassing $215 billion, which is up 66% year-over-year. Data Center revenue from 2025 reached $194 billion, and the company projects that computing demand will continue to soar. It mentioned a whopping $1 trillion in demand through 2027, which analysts interpreted as the total value of cumulative purchase orders for Nvidia's high-end data center platforms from 2025 through 2027.

This figure doubles a similar number shared last year, which was $500 billion for 2025 through 26. Reportedly, roughly, 60% of Nvidia's revenue currently comes from hyperscalers with the remaining 40% diversified across industries, massive growth driven by AI, and Adventists. Vidia is the primary beneficiary of it. Jensen has repeatedly said, the more you buy, the more you save, referring to it as CEO math,

and using the phrase to indicate that purchasing more new GPUs, reduce data center energy costs, and create lower total cost of ownership, TCO, and long-term savings despite high upfront costs. What do you hear this time, Shaheen? Well, teaming is the word. It was packed,

and the conference was a new peak. A lot of technology announcements, and we'll get to that in a second, but it seemed like Envideo was up-leveling the message towards business metrics. So as some witty person put it, it went from the more you buy, the more you save, to the much stronger, the more you buy, the more you make.

This also meant the audience expanded beyond customers to include vendor partners, most of whom had a banner year thanks to continued strong demand for AI. This shift in messaging is consistent with the evolution of both the technologies and the economics of AI, towards inference and agentic AI and AI factories and sovereign AI that function as foundational economic infrastructure for societies. It is also consistent with so-called physical AI and robotics that represent a very wide spectrum and perform many tasks today even without AI. So let's go through those technology announcements.

ISEC Groot N2 is a new foundation model for humanoid robots, enabling them to complete tasks in unfamiliar environments, twice as fast as previous models. There were over 100 robots on the show floor and around the conference, and there is huge progress there. But demos being demos and liable to fail, you could see at least a few of them were being controlled by remote control by folks standing by in a corner with their hands behind their backs, fiddling with the control units.

Drive Hyperion is a level four robot taxi software for NVIDIA and partners like Uber, BYD, and Nissan. Metropolis VSS3 Blueprint enables AI agents to search and reason over video footage at the edge, for example, in a retail store and to do it 100 times faster than manual review. The Vera Rubin platform, of course, was introduced again and formally launched as the successor to Blackwell. That's the new Vera CPU that replaces the Grace CPU, the Ruben GPU, and the special Ruben CPX GPU, which is optimized for million token massive context AI inference. They are three to five times higher performance and ten times lower in inference costs compared to Blackwell. And just like Blackwell, you get them in racks and rows of racks to build your AI factory.

For 64-bit applications, however, the choice is more complicated and depends on workloads and whether simulating 64-bit operations using lower precision hardware gets you the desired accuracy and speed. And of course, assuming you can even get your hands on the new systems. Following the acquisition of Grok's IP, they also announced a new NVIDIA-GROC 3 LPU, LPU standing for language processing unit. Nvidia paid some lip service to data centers in space by mentioning the concept and talking up what they called the Space One Verarubin module, a specialized version of the architecture designed for data centers in orbit and autonomous space

operations. We shall see, but as we have discussed here, it's one thing to have computers in space to be used there, and entirely another thing to have a data center there to be used by users on Earth. The latter is too far out, and yet another technology that needs many breakthroughs before it might become a possibility. But they made me talk about it, so well done there. Under the label of operating system for agentic computers, they announced NemoClaw, an enterprise-grade version of OpenClaw, the open-source orchestration frame, framework for AI agents. By enterprise grade, they mean the usual. Enhanced security, sandboxing, user management, permission management, etc. We talked about Nematron 3 super

last week, a new reasoning model that is available now. DLSS-S-5 is for neural rendering for gaming that uses 3D-guided AI to generate photorealistic 4K videos. Initial feedback was mixed, however, described as making everything look the same, and it shows how hard it is to automate creative tasks. Their right-brain, left-brain integration remains a challenge. Kuda-X is a suite of libraries, microservices, and tools for data science and HBC applications. And then they had several announcements focused on traditional enterprise and integration with legacy applications. This is something that was started last year or before that event. and will be one of those final frontiers for AI to get baked into enterprise IT.

QDF, CUDF, is a GPU-accelerated data frame library of APIs to speed up data processing and ETL tasks. ETL stands for extract, transform, load, and is an important step in loading structured data into databases for data warehousing and corporate decision-support applications. NV-comp is a library for lossless data compression and decompression for AI and data-intensive applications. QVS, CUVS, is a library designed to accelerate vector search and similarity search algorithms for AI applications, including high-performance support for approximate nearest neighbor ANN search.

These are all focused on SQL applications, SQQ, standing for structured query language, which is the standard way to interact with databases that are the foundation of enterprise applications. They've been working with Oracle AI database, Microsoft SQL Server, IBM PrestoSQL, Google BigQuery, Amazon, etc. Finally, as they let AI agents access and query SQL databases and file systems, the idea of AI native storage comes up again, and they pointed to a future product called Bluefield 4 STX that would provide intelligence for dumb storage.

So, like we said last week, expansive view of the turf and lots of cooperation and competition as they pushed industry forward. The allure of selling into the world's second biggest economy was in the spotlight in two ways last week, as Nvidia has revived manufacturing of its H-200 AI processors for sale in China. The issue of chip exports to China has taken many turns over the past year, with shifting policies and rules from the Trump administration as the U.S. tries to balance the broad goal of winning the AI race with China over against the financial rewards of selling AI technology into China's enormous AI market. Meanwhile, server maker Supermicro has gotten embroiled in the China AI business with some of its employees, including a co-founder, being named in an indictment for alleged, participation in a scheme to smuggle NVIDIA chips into China. For its part, China has also shifted and changed its own balancing act between a

determination to compete with the U.S. on AI by building up its domestic chip capability versus its desire to access the most powerful AI chips that the U.S. is willing to export. And according to NVIDIA last week, demand signals out of China have strengthened. The U.S. Department of Energy announced funding for the Genesis Missions efforts to take on complex science and technology challenges that includes a 293 million request for application for, quote, transforming science and energy with AI. DOE invites interdisciplinary teams to leverage AI models and frameworks for more than 20 national challenges that include advanced manufacturing, biotechnology, critical materials, nuclear energy,

and quantum information science. Yeah, very exciting. The requirement. Quest for Application, RFA, is open to teams from the National Laboratories, U.S. industry, and academia. Phase 1 awards will range from $500,000 to $750,000, pretty significant, and will support nine-month projects. Phase 1 applications and Phase 2 letters of intent are due by April 28th, and Phase 2 applications are due May 19th. The OE plans to hold an informational webinar about this RFA on March 26th. We came across an open invitation from one of the Nature journals, this one called Quantum Information,

that is collecting peer-reviewed papers on a major issue in the HPC world, and that is development of practical quantum technologies based on integration with classical HPC and AI-based control layers. As the publication said, this kind of co-designed strategy is crucial from moving beyond proof-of-concept experiments and toward scalable, robust, and distributed quantum systems. The project is called Foundations and Advances in HPCAI Quantum Co-Design, and the publication will accept papers until this coming December 20th. Yes, the framing is great, less of a snapshot

of the state of these technologies, and more of a strategic nudge and blueprint, that the next generation of computing will need cross-layer integration and end-to-end optimization, as you said, and not just component-level excellence. I also liked how they structured their topics of interest, which include unified programming models, network-centric scaling, energy-aware design, hybridization of AI simulation and quantum, and ecosystem orchestration, and several other familiar themes. It is good to see the community mobilizing around these emerging needs. Various articles covered the 2025 Touring Award winners Charles Bennett and Gilles Broussard for establishing the field of quantum information science and inventing quantum cryptography.

They developed the foundational protocol called BB-84 for quantum key distribution, QKD, which uses quantum mechanics to enable theoretically unbreakable communication. You pointed me to the great article by the Quantum Magazine, which tells the story in a very engaging way. Nice job there. Just a couple of weeks ago in our February 23rd episode, we talked about quantum communications and how it works in broad brush terms. Please check that out since it's at the heart of this award. Cryptography depends on two parties being able to decipher a message that is liable to be seen by eavesdroppers. They need a shared key and they need to. to encrypt a message in a way that is easy to encrypt, but really hard to decrypt. The usual methods are mathematical, looking for math functions that act like a trapdoor, easy to go in in one direction,

and impossible to go from the other direction. Bennett and Brassar grounded cryptography in quantum physics rather than mathematical hardness. Their BB-84 protocol introduced the idea that he can secure information physically in such a way that the action of eavesdropping destroys it and alerts the sender and the receiver. Their work connects computer science with quantum mechanics and is a foundational milestone in creating the field of quantum

information theory. So the award is not as much about new cryptography, although that's pretty cool, as it is about a new way to secure information as a physical entity, not a logical entity. All right, that's it for this episode. Thank you all for being with us. HPC Newsbytes is a production of OrionX. Shaheen Khan and Doug Black host the show. Every episode is posted on Orionx.net. If you like the show, please rate and review it. Thank you for listening.