# I need an AI sanity check (9700 Pro x2) > Source: > Published: 2026-09-03 19:28:37+00:00 Don’t install AMD’s amdgpu folks unless there is a reason. AMD suggests installing the proprietary driver because its the one they commercially support but in reality you just need the runtime and compilers because mainline kernel has all that you need to run ROCm since 5.11. You are most likely running an “unsupported” GPU to begin with. Hi People, I am pretty sure you are going to me mad with me, but I am exploring the same setup, on Windows, yes, Microsoft windows server… Using QWEN 3.6 27B Q4_K_M, and with one card I can barely hit 22 tokens per sec… I am using OpenCode CLI and ROCM GGUF –> llmala.cpp (Windows) 2.24.0… Unfortunately, I have zero knowledge on Linux… My second card will be here within 10 days, but I am quite sure I will not get pass 30 token per sec. Any suggestions? Brutal honesty is appreciated… Let me know how it goes. I could not get two cards to work in windows at all without crashing. Hi Strawberry. I got my second card today. Installed it, and windows recognized it using my customized driver. The card is also recognized on LMStudio, and the VRAM has doubled to 65GiB. So far everything seems to be good, even though I do not have enough Knowledge to evaluate the behavior; Model is loaded half on each card, and token per second, both read and write are at the same level as before 16 to 18 on generation. power draw on both cards and utilization are at around 50% on BOTH cards. BUT when I press with more complicated prompts, I see both read and write tokens per sec to double. Not NEAR double, literally double. at some point I got 37,8 tokens per sec, and utilization and power draw reached max for both cards at 304 Watt. So my first impression is that this is working; it does feel snappier and more interactive, but not for all workloads. If anyone has any benchmark prompt or process I Can use to test I am happy to do so. This would be very good because I could do a direct comparison with linux distros performance. I will post more info soon. Please ping me if you need hardware config You need to activate MTP, will basically double your token output. Use tensor or layer split depending on task, read up on best ways to split between two cards. Not always - there’s a cost to MTP: loss of significant prefill performance and concurrency. It very much depends on the use case. [thro](https://forum.level1techs.com/u/thro) 28 My AI adventures on linux were a shit show with ubuntu (fighting ROCm support, etc.), and everything “just worked” with vulkan support on CachyOS. Seriously, broke my Ubuntu 24.04 LTS video output trying to make it work spent a couple of days messing with it. CachyOS just worked. My baseline observation is that if you’re on AMD you want current driver stack and CachyOS gives you that. Ubuntu is a clusterfuck for this in my experience. My advice is don’t try fight the tide, go with a modern rolling release like CachyOS for good AI support on AMD. Also consider llamacpp (llama-cli / llama-server) instead of ollama - its faster. I can get over 200 tokens/sec on unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF with 7900XTX on CachyOS with llama-cli. Ubuntu 24.04 is trash when it comes to Rocm, installed 7.2.4 and I never could get anything to work because every build included the 5.2 system Rocm headers aswell as the new 7.2.4 ones. 26.04 seems better here with consistent newer preshipped Rocm? Anyways, DeepSeek V4 Flash wrote a 10 line build script that only includes 7.2.4 1 Like [thro](https://forum.level1techs.com/u/thro) 30 I had similar issues with Ubuntu and rocm. Cachyos just works [Janos](https://forum.level1techs.com/u/Janos) 31 I have no issues with Ubuntu 24.04-26.04 and ROCm with 2x7900XTX. I started with Ollama, then Lemonade/llama.cpp and because of Qwen3.8-Flash-Next since last week Unsloth-Studio/llama.cpp. I’m wondering right now if three 9700s run well, or do odd numbers of GPUs cause problems? Three 9700 Pros work just fine, with llama.cpp Vulkan under Linux. I have yet to try four.