There’s a growing challenge in how we build high-performance computing infrastructure. As specialized chips, accelerators, fabrics, and storage solutions flood the market, one question has become central:
“How do we navigate this maze of options – and make the right choices to architect the best solution for our workloads?”
Reshaping Every Industry: Post-Moore’s Law Meets AI Everywhere
The infrastructure behind HPC/AI is evolving just as quickly. AI-infused applications are emerging everywhere, edge and datacenter workflows are increasingly connected, and customization is the way to get performance. As a result, hardware decisions that once changed every few years now shift by the quarter. The old model—simple to choose, simple to configure—of buying systems is no longer viable. Nowadays, you have to architect a distributed platform that’s optimal for your workflow for today and tomorrow, pick from among a dozen or more options for each minor or major component, make sure they are all tested together—only to find out, just as you’re about to place the order, that another equivalent component could ship five weeks earlier and would be easier to service.
That’s where co-design comes in.
The Age of Complexity Demands Collaboration and Co-Creation
The explosion of specialized technologies has made designing modern IT infrastructure more complex than ever. With so many choices across compute, networking, storage, and software—as well as supply chain considerations—it’s nearly impossible to make optimal decisions without deep expertise in each domain.
The sheer volume of options is both a blessing and a curse:
- X86, Arm, and RISC-V CPUs for workstations and servers, and even more choices for embedded devices
- Accelerators such as GPUs, FPGAs, and custom AI chips are available from many vendors or via public cloud services, each tuned to different stages of a workload
- Interconnects, fabrics, DPUs, storage stacks, AI frameworks, simulation software
- Edge devices, cloud platforms, and everything in between
On one hand, you can finetune performance like never before. On the other hand, making the wrong call—on precision formats, topology, cooling, or future upgradability—can mean wasted resources and missed goals.
Co-design Flips the Process
Instead of starting with pre-built configurations, you start with your mission. Then, a team of domain experts—from hardware to software to environmentals, to testing—architects the platform and builds the systems around that mission.
From Device to Datacenter: Why the Whole Fabric Matters
The rise of AI, digital twins, and edge analytics has changed the game. Today, data is generated everywhere, and needs to be processed, moved, and acted upon across a distributed computing fabric.
Co-design ensures that the system isn’t just fast in isolation, it’s fast in context. Whether you’re running fluid dynamics simulations in the cloud, analyzing drone footage on the edge with intermittent connectivity, or training foundation models in your datacenter, the system needs to perform seamlessly for each task across all layers.
AI Changed Everything (Again)
AI has brought more than just buzz. It has brought new methods and algorithms, new hardware needs, and an insatiable appetite for compute. But not always in the way you’d expect.
For example, many AI workloads thrive on low-precision arithmetic and dense accelerator clusters. This continues to be a good example of how the industry, motivated by AI, has created powerful new tools for traditional HPC simulations. At the same time, many applications continue to demand double-precision, large memory bandwidth, and fine-grained control. Trying to force one architecture to do both is inefficient, but if you can anticipate how and when your application requirements may change, that could make a big positive difference.
Co-design allows you to optimize. You can pair high-throughput AI clusters with simulation nodes, link them through smart fabrics, and scale with purpose—ensuring that each workload gets the resources it truly needs. Or you can phase-in the right upgrades or set up a system that can be partitioned with finer granularity.
Real-World Highlight: Gryf — First Suitcase-Sized AI Supercomputer
Gryf is a prime example of co-design in action. Co-designed by GigaIO and SourceCode, this edge AI system delivers petaflop-class performance in a TSA-friendly carry-on form factor. Gryf integrates server-class GPUs, compute, storage, and networking into modular mix-and-match sleds.
Built in SourceCode’s U.S.-based ITAR-compliant facility, Gryf shows the creativity that the co-design model unleashes, translating mission requirements into real-world solutions, and bringing datacenter-class performance and reliability to a portable, suitcase-sized system.
Why Co-Design Works
Co-design is a practical approach that delivers clear, measurable benefits:
- Accelerates Time-to-Market
Alignment and streamlined decisions shorten design and deployment cycles. - Reduces Technical Risk
Validation, architectural trade-off analysis, and interoperability checks mean fewer surprises. - Optimizes Design Constraints
System-level co-optimization balances power, cooling, density, and performance. - Enhances Adaptability to Emerging Tech
Modular, fabric-agnostic designs support next-gen components. - Ensures Compliance and Certification Readiness
Security, ruggedization, and regulatory needs are built in from the start.
Why SourceCode?
Trusted for over 30 years, SourceCode has refined the co-design process into an effective business model. In a world where everything is an AI-enabled computer, your enterprise is code, your infrastructure is intelligent, and co-design is the best way to build it!
Explore what’s possible. Visit www.sourcecode.com or reach out to contact@sourcecode.com to start a conversation.
Note: This article was first published in HPCwire on September 29, 2025.
