HACI 2026: Notes from an HPC User

Last weekend¹¹ 22–25 May 2026, I attended HACI 2026, the International Forum for HPC & AI Co-Driven Innovation.

I build scalable scientific simulation codes, and I also use AI tools in daily development. However, my simulation stack is still mostly FP64/FP32 brute force. Given that constraint, I listened with one question: What can a small lab with two consumer GPUs actually do next?

Four talks stayed with me.

• Jack Dongarra: The plausible path to “effective zettascale” is not brute-force FP64, but certified mixed-precision algorithms, communication-avoiding methods, AI-augmented reduced-order models, and hybrid AI+simulation workflows with rigorous error control and uncertainty quantification. His weather prediction example was concrete: AI trained on ~40 years of data can predict 10-day weather in seconds, but should run alongside numerical models.
• Taisuke Boku: Fugaku-NEXT and a practical migration path: from Fortran CPU code with OpenMP pragmas to C/C++ CPU-GPU code using OpenACC/OpenMP target offload. His HAIRDESC public GPU course reinforces this point: migration needs method and training, not only hardware.
• Bingsheng He: In LLM systems, small teams (2–3 people) plus agents can build production systems quickly.
• Zeyi Wen (parallel session 2): Efficient fine-tuning and inference of frontier LLMs on a single RTX 4090, with >95% peak performance and lossless results.

For a small lab, the conclusion is straightforward: optimize for attainable performance first. My possible next steps would be:

1. Measure data movement bottlenecks before adding compute.
2. Expand mixed precision only where error behavior is controlled.
3. Build one small AI+simulation prototype with validation.
4. Reproduce one single-GPU optimization result on our own RTX hardware.

HACI 2026 was broad. My output is narrow: one clearer direction and one smaller next step.