Gemma 4 26B posts strong R9700

Local Gemma 4 inference on AMD looks backend-sensitive: on this card, Vulkan materially outpaced ROCm in prefill, while decode stayed strong but closer together. That makes the result useful less as a universal Gemma score and more as a signal that the runtime stack can dominate real-world throughput.

• –Vulkan beat ROCm on this setup by a wide margin in prompt processing: 2,949 vs 1,422 tok/s at `pp1000`, and 1,450 vs 681 tok/s at `pp1000 @ d50000`.
• –Generation speed was also higher under Vulkan, but by a smaller gap: 92.9 vs 70.9 tok/s at `tg2500`, narrowing to 78.2 vs 61.5 tok/s at the longest context.
• –The test was run on a 210W power cap with ROCm 7.2, so the result reflects both software maturity and power-policy constraints, not just raw GPU capability.
• –For people trying to run 26B-class open models locally, this is a reminder to benchmark the whole stack: driver, backend, quantization format, and batch settings all matter.