Performance Optimizations for CMP 170HX in llama.cpp
Recent technical reports from the developer community have highlighted significant performance variations in the AMD CMP 170HX GPU when running large language model (LLM) inference via the llama.cpp framework. According to unverified reports shared on Reddit, specific software configurations can lead to substantial increases in prompt processing (PP) speeds, though results appear highly dependent on the specific model architecture being utilized.
Impact of CUBLAS on Inference Speeds
A user testing the CMP 170HX reported that rebuilding llama.cpp with the GGML_CUDA_FORCE_CUBLAS=ON flag resulted in a measurable performance increase for the Qwen 3.6 27B model. In this specific test case, prompt processing speeds rose from approximately 1,000 tokens per second (t/s) to 1,500 t/s when the entire model and context were held in VRAM.
However, the user noted that this optimization did not yield consistent results across different model versions. While the Qwen 3.6 27B model saw a 50% improvement, the Qwen 3.5 122B model reportedly performed twice better when CUBLAS was disabled, yielding roughly 1,000 PP t/s compared to 500–600 PP t/s with the optimization enabled. These findings, while detailed with instruction benchmarks, lack independent verification.
Hardware Limitations and Benchmarking
Despite the observed speed increases, the CMP 170HX continues to trail high-end consumer hardware in certain metrics. Benchmarks provided by the tester indicated that the card’s f16x2_fma and VRAM speeds were measured at approximately 500–600 GB/s, which remains slower than the NVIDIA RTX 3090.
The hardware’s performance profile suggests it may be particularly efficient for INT8 compute tasks, a sentiment echoed by community members discussing the card’s potential value in specialized AI workloads.
Comparative Analysis: Apple Silicon vs. llama.cpp
In a separate technical comparison involving the M4 Max chip, benchmarks conducted on August 15, 2026, provided insight into the trade-offs between llama.cpp and MLX-based frameworks. Using the Qwen 3.6 27B MTP GGUF model, llama.cpp achieved an average prediction speed of 24.24 t/s and an average prompt processing speed of 159.25 t/s.
While llama.cpp demonstrated stability in speed as context windows increased, it was noted to place a significantly higher thermal load on the hardware compared to the mlx-vlm framework. In the tested M4 Max environment, llama.cpp reportedly caused the system to reach temperatures of 100 degrees Celsius, triggering thermal throttling. In contrast, the MLX-based inference stack, which utilized a draft model for optimization, achieved higher prediction speeds of 37.59 t/s without reaching the same thermal thresholds.