The CUDA ecosystem received two significant boosts this week.

The CUDA ecosystem received two significant boosts this week. First, the llama.cpp project (commit b10089) completed support for all quantized GGML types on CUDA, including k-quants, i-quants, and mxfp4, enabling full device-side embedding lookups and avoiding slow host fallbacks. This improvement, detailed in the release notes, means popular GGUF models using mixed quantizations now run entirely on GPU, enhancing performance for local inference on NVIDIA hardware.

Second, Primate Labs released Geekbench 7 with CUDA support, adding CUDA alongside OpenCL, Vulkan, and Metal for GPU benchmarking. This mainstream endorsement signals that CUDA remains the go-to API for professional and AI workloads, likely increasing its visibility among developers and enterprises evaluating NVIDIA GPUs.

While minor bug reports appeared on Stack Overflow (1, 2), they reflect typical development friction rather than ecosystem weakness. The Manifold market betting on CUDA monopoly through 2027 at 54% yes suggests neutral-to-positive sentiment.

What the sources said:

• llama.cpp release notes: "cuda: add k-quant support to GET_ROWS ... This closes GET_ROWS type coverage on CUDA: every quantized GGML type now takes the direct device path."
• Tom's Hardware: "Geekbench 7 is also set to introduce support for Nvidia's CUDA API in addition to OpenCL, Vulkan, and Metal, allowing users to benchmark Nvidia GPUs using the same API that is used in a variety of professional and AI applications."