Remove NVIDIA/CUDA-Only Constraint in Domino IQ Local Mode — Add ROCm and Vulkan Support for AMD Unified Memory Platforms

the Problem

Domino IQ Local Mode — and with it, the RAG pipeline over NSF databases — requires NVIDIA GPU hardware with CUDA drivers. This excludes an entire and rapidly growing class of AI hardware from accessing Domino IQ's most valuable capabilities.

A Shifting Hardware Landscape

AMD's Unified Memory Architecture (UMA) platform is fundamentally changing what local AI inference looks like. Current and upcoming AMD processors — combining high-performance CPU cores, RDNA-series integrated GPU, and XDNA NPU on a single die with a shared memory pool — are enabling organizations to run large language models locally, privately, and without cloud dependency. The current generation already supports up to 128 GB of unified memory addressable by the GPU. The next generation, announced for Q3 2026, scales this to 192 GB with 160 GB allocatable as effective VRAM — enabling local inference of 200B+ parameter models on a workstation-class system. Major OEMs including ASUS, HP, Lenovo, and Framework are bringing these systems to market across laptops, mini-PCs, and compact desktops at competitive price points.

This is precisely the hardware profile that Domino IQ's sovereign AI story targets: capable, affordable, air-gapped, and data-sovereign. Yet none of it can run Domino IQ in Local Mode.

Why It Matters

• RAG over NSF is lost. Domino IQ's most differentiated feature is unavailable to any customer without NVIDIA hardware, forcing them into Remote Mode where data leaves the server — directly contradicting the zero-cloud, sovereign AI value proposition.

• Business Partners are losing deals. Customers who have invested in AMD UMA platforms for local AI cannot use Domino IQ's full capabilities, and the cost of adding NVIDIA server GPU hardware is often a dealbreaker in mid-market and public sector accounts.

• The gap will widen. As AMD's next-generation unified memory platforms reach mainstream availability through OEM channels in H2 2026, the number of organizations affected by this limitation will grow significantly.

The Technical Path Is Clear

llama.cpp — the inference engine Domino IQ already uses — natively supports AMD's ROCm/HIP backend today. Building and shipping Domino IQ inference binaries compiled against ROCm is additive work, not a rewrite. The open-source ecosystem (Ollama, llama.cpp, LM Studio) already runs fully on AMD UMA hardware with ROCm. HCL would not be pioneering new ground — it would be catching up to what the ecosystem already supports.

The Ask

1. Ship Domino IQ inference binaries compiled against AMD ROCm (HIP backend).

2. Add a Vulkan compute backend as a portable fallback for non-NVIDIA, non-AMD hardware.

3. Restore CPU-only mode for development and testing environments.

4. Confirm that RAG is available across all supported backends, not just CUDA.

The CUDA-only constraint is an implementation artifact that is becoming a meaningful product limitation as the AI hardware market diversifies. Addressing it would directly strengthen Domino IQ's sovereign AI positioning and expand its addressable market at exactly the moment AMD UMA platforms are reaching enterprise mainstream.