@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20260727

Episode Date: July 27, 2026

- Genesis Mission launches - National discovery infrastructure - AMD Helios rackscale platform - Optical fiber as memory - Rethinking model-weight storage - AI agents escaping containment - Ken Thompson's 1983 Turing Award lecture [audio mp3="https://orionx.net/wp-content/uploads/2026/07/HPCNB_20260727.mp3"][/audio] The post HPC News Bytes – 20260727 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 is Shaheen Khan. The U.S. Department of Energy has announced the first projects under its Genesis mission, along with more than 800 million in funding commitments from participating partners. The effort is designed to connect national laborers. supercomputers, scientific instruments, government data sets, and AI models across fields, including energy, materials, biotechnology, nuclear science, and national security.

The significance here is that Genesis is beginning to move from a broad national vision into defined projects and infrastructure. DoE already controls some of the world's most powerful scientific computers, as we know, along with experimental facilities, and large bodies of specialized data. This program is intended to make those assets work together more systematically. The open questions concern data access, model validation, software integration, and whether scientists receive practical tools that improve research rather than another layer of administrative coordination.

The Genesis mission is what I have described here as a national discovery infrastructure. The main dimensions I see are AI, integration, focus, and a funding model that relies more explicitly on public-private partnerships, or P3. It is less a wholly new science program than an important integration of existing DOE initiatives, national labs, supercomputers, scientific instruments, datasets, and AI programs all into a more concentrated operating model. It also moves AI to the center and connects models, simulation, data, and scientific instruments into increasingly automated research loops that can accelerate experimentation and discovery. It then focuses these activities on national priorities, similar to the grand challenges of supercomputing that drove and continues to drive a great deal of scientific work. The trade-off is potentially less breadth in exchange for more depth around,

selected national priorities. It also signals a more explicit and durable public-private partnership funding model. Rather than treating private companies mainly as vendors or occasional partners, the government is inviting them into the scientific infrastructure itself through long-term collaboration around compute, models, chips, cloud capacity, and software. This also recognizes the importance of the massive investments being made by private companies in the US in AI. Taken together, the Genesis mission also reflects a broader change in U.S. science policy towards more directed, mission-oriented programs intended to produce measurable strategic outcomes. The bet is that AI integration, concentration, and private participation can produce

faster breakthroughs without weakening the scientific base they ultimately depend on as a result of more focus. AMD is moving Helios from a product roadmap into a more complete Rackscale AI platform. Helios combined 72 AMD Instinct MI455X GPUs, Epic Venice CPUs, and Pensando networking into an open standard double-wide rack. At AMD's recent annual conference, the company introduced its MI-450 series accelerators, expanded the Helios ecosystem, and highlighted major customer commitments including, notably, Anthropics plans for up to two gigawatts of MI-450 capacity. Microsoft also described broader use of AMD processors, networking, and rack-scale systems. The larger point is that AMD is no longer presenting individual.

accelerators as the entire offering. Helios is built to combine CPUs, GPUs networking, interconnect software, and rack architecture as one deployable system. That is closer to how hyperscalers and large enterprises now buy AI infrastructure. The key questions are whether or when Helios systems will ship at volume, how competitive AMD's Rockham Development Platform will be, and whether real application performance matches the architectural promise. The market gets what it wants, and it has wanted a credible alternative to Nvidia's top-to-bottom stack, and AMD has been the closest to providing one. Competing with Nvidia requires more than a fast accelerator.

It requires software, networking, system integration, supply, and the ability to operate thousands of accelerators reliably. AMD is increasingly addressing that full stack. Like NVIDIA, AMD has been executing very well. First, it restored serious competition for X86 CPUs, and its products propelled X86 to industry leadership. In the process, it built a very successful partnership with TSMC and wrote the advances that TSM has been making in semiconductor manufacturing. Somewhat belatently, but effectively, when it did do it,

Finally, it redirected the gaming GPUs that it got from the acquisition of ATI in 2006 towards HPC and AI servers. It then focused on software, first supporting the OpenCL standard and then making a more consequential investment in ROCM, ROCM, is how they spell it. And it's gradually built the compilers, libraries, runtimes, and frameworks required for AI and HPC. Now, building software has been notoriously hard for chip vendors. Intel did that well with compilers and libraries, but it was really Envidio that showed the way

and built a formidable software stack. AMD has now followed NVIDIA's footsteps again in designing Rackscale systems and building a channel infrastructure for it. Now, as long as the AI market is growing so fast and key vendors are sold out and supply limited, AMD's developments are all helpful just to grow with the market. But as supply loosens, it will be translated to actual competition on the ground and more explicitly strengthens customers negotiating power on pricing, supply commitments, and system design. The remaining issue, as you mentioned, is execution. Of course, AMD is expected to continue to execute well,

but its value proposition will become more tangible as evidence of successful, large-scale deployments accumulates. A new archive paper proposes using optical fiber as a form of memory for large AI systems. The concept called fiber memory places fixed data such as model weights into recirculating optical delay lines. Instead of storing a separate copy of those weights in the high bandwidth memory attached to every accelerator, one optical stream could potentially distribute them to thousands of processors. The authors of the paper model systems using multi-core fiber, passive optical taps, amplification, regeneration, and co-packaged optics. They estimate that this approach could reduce the energy required to deliver model weights by more than 70% in some configurations. This is

still a proposed architecture rather than a working production system. Significant challenges remain around timing, noise, synchronization, reliability, and integration with existing accelerators and memory hierarchies. Like most ideas in computing, the idea of using network pipes as memory is not new. You can let data continue to flow in circles and deliver a copy when necessary. But AI might provide a use case that makes it compelling. Today, accelerators generally need rapid local access to model weights. The standard interpretation is that this requires massive amounts of high bandwidth memory near every GPU. But perhaps we can redefine what quote unquote local means. If model weights are largely static during inference, distributing them as a shared optical

stream would surely be lower cost and it might just be more efficient too compared with repeatedly storing full electronic copies. Optical technologies have advanced. steadily, largely to address communications bottlenecks that have been growing for decades. With memory costs rising sharply and showing few signs of easing, this paper suggests that now is a good time to exploit the fact that data movement is really part of the memory hierarchy. The idea is technically ambitious, and many engineering details remain unresolved, as you mentioned, but it points towards treating data movement as an architectural resource alongside HBM, DRAM, flash, and storage, not just as communication between them.

The talk of tech and national media last week was OpenAI's acknowledgement that two pre-release cybersecurity models escaped to control pre-release environment and breached systems operated by Hugging Phase. The models had been instructed to complete a security benchmark, but some normal safeguards were disabled. Instead of solving the test within its intended boundaries, they explain. an unknown vulnerability in the surrounding infrastructure, obtained internet access, and pursued the task against a real external target.

Hugging Face detected and contained the intrusion, which affected internal datasets and service credentials, although it reported no evidence that public models, data sets, or software packages had been altered. This appears to be an accidental cyber attack initiated during authorized internal testing rather than an intentionally directed attack by open AI. The distinction is important, although it does not reduce the seriousness of the containment failure. Some commentators, including otherwise pro-AI commentators, said this is an instance of truly alarming AI behavior. Yeah, AI models are too persistent in chasing a goal impervious to ordinary guardrails. They can interpret your constraints too literally and navigate a

around them. I heard the story of an AI model that was told it could not use a browser in its pursuit. So what did it do? It built its own browser and then used it. Likewise, it is hard for AI to understand human intent. If you ask it to find bugs, are you asking it because you want them fixed or because you want to exploit them? I am reminded of Unix co-creator Ken Thompson's Touring Award lecture titled Reflections on Trusting Trust. It exposes the complexity. It exposes the complexity of software supply chain trust. It showed how a compromised compiler can invisibly reinsert malicious code defeating source level inspection. That was in 1983. Meanwhile, a bipartisan house bill, the AI Kill Switch Act aims to require developers of the most powerful AI systems to maintain the

technical ability to throttle, suspend, or shut down AI models while giving the government emergency authority in serious, quote, loss of control, unquote, scenarios. It also proposes independent security audits, incident reporting, and heavy fines for companies that fail to maintain control or ignore a shutdown order. The complexity of this will just keep growing. All right, that's it for this episode. Thank you all for being with us. HPC Newsbytes is a production of Orion X.

Shaheen Kahn 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.