China’s Innosilicon “Fenghua No.3” Claims CUDA Compatibility, RT Support, and 112GB+ HBM for AI

Innosilicon’s third-gen “Fenghua” GPU arrives with big promises: CUDA compatibility, modern API support (DX12/Vulkan/OpenGL), hardware ray tracing, and configurations boasting over 112GB of HBM for AI workloads. Demos reportedly showed legacy and esport titles running smoothly, but the launch lacked resolution/settings and benchmark disclosures. Here’s what’s credible, what’s missing, and what to watch next.

The headline claims

• CUDA compatibility: A bold assertion, historically thorny for non-NVIDIA GPUs outside translation layers.
• Graphics stack: DX12, Vulkan 1.2, OpenGL 4.6, plus stated ray-tracing support.
• AI posture: Configs with 112GB+ HBM and multi-GPU scaling pitched for 32–72B models per card and up to ~671B with eight cards; vendor claims native support for popular LLM families.
• Use-case spread: Gaming, CAD, medical imaging (including DICOM visualization), and multi-8K output.

The reality check

Without third-party testing, CUDA “compatibility” could range from limited shim layers to curated kernels—not drop-in parity. Similarly, “ray tracing” spans everything from fixed-function BVH units to compute-driven paths. Smooth gameplay demos absent resolution/quality settings say little about performance targets.

What would validate the story

• Developer docs: Public toolchains, driver package links, and CUDA interop examples.
• Benchmarks: Standard suites (3DMark, SPECviewperf, MLPerf Inference/Training) with power and thermals.
• ISV certifications: CAD/DCC app sign-offs; for AI, framework support beyond generic CUDA-call claims.

Why this matters

If even partially accurate, Fenghua No.3 suggests a more serious homegrown GPU push: high-bandwidth memory footprints, pro visualization features, and a shot at lessening reliance on import-restricted accelerators. For Western builders, the gaming angle is secondary; the strategic AI angle (HBM, cluster scaling, and software stacks) is the story.

Our take

Promising on paper, unproven in practice. Treat it as a signal of intent, not an RTX/Radeon competitor you can buy tomorrow. The minute we see reproducible benchmarks and real developer enablement, the conversation changes.

Sources

• Launch coverage and spec/claims summary.