CMP170Hx “Spark” Machine

I got the CMP170 cards and unlocked them. I wanted to share my set up for CUDA since maybe it would be useful to others. First off, I hate e-waste and we are in a special time for RAM. I wanted to have a DIY CUDA box, and I had started by adding additional cards to an old asus predator prebuilt I had around, which also had 64gb DDR5. To add the CMPs I needed more CPU lanes and newegg had some really good deals on CPU/MB/etc combos. Didn’t need a combo with RAM, otherwise I would have gotten it in newegg microcenter. Anyway, I got a cheap case, some noctua fans for the cards, and transferred the memory/ssds. Placed previously owned cards on oculink slots, and used the main x16 for the GPU switch that houses the two CMP170s, so their effective speed is 2x16 across and with the other cards (which are 4x4, and therefore same speed). Qwen Flash Next, turns out, fits very nicely in these cards. There is also a repository for deepseek, but you’d need at least 3 64GB cards to run it, and with prices rising, it will be hard to justify the gamble of buying ex mining cards for LLMs. However…so far, these cards are great. Concurrency is good, prompt processing averages 4000 tps on Flash Next, decode is 80+ on a single stream. No MTP added. Third picture shows the 3 models I am now running in this CUDA box (flash next, qwen 27b, gemma 26b). Anyone else trying out Flash Next on these cards?