OPEN LINK ↗
// 45d agoBENCHMARK RESULT
RX 6900 XT Benchmarks Split ROCm, Vulkan
A Reddit user shared quick llama.cpp benchmarks on an AMD Radeon RX 6900 XT after building with ROCm 6.4.2 and the latest Vulkan backend. The results show a split picture: Vulkan often leads token generation throughput, while ROCm can be faster on prompt processing for some workloads, especially Qwen 3.5 4B Q8_0. For Gemma 4 E2B Q4_K, the winner shifts with ubatch size, which reinforces that backend performance here is workload-dependent rather than universally favoring one stack.
// ANALYSIS
ROCm is not a blanket win or loss here; it looks stronger for prompt processing on the Qwen run, while Vulkan is more consistently better at token generation.
- –Gemma 4 E2B Q4_K is mixed: Vulkan wins pp512 at larger ubatches, but ROCm holds the edge on tg128.
- –Qwen 3.5 4B Q8_0 is clearer: ROCm is substantially faster on pp512 across all tested ubatches, while Vulkan is faster on tg128.
- –The post is a quick local benchmark, not a controlled long-context study, so it is best read as a practical snapshot for AMD RDNA2 users.
// TAGS
amdradeon rx 6900 xtrocmvulkanllama.cppgemma 4qwen 3.5local llmgpu benchmark
DISCOVERED
45d ago
2026-04-28
PUBLISHED
45d ago
2026-04-28
RELEVANCE
8/ 10
AUTHOR
grumd