AI cloud provider Zyphra has conducted a benchmark test using AMD's MI355X chip, comparing its performance against NVIDIA's B200. The tests, which ran models like DeepSeek V3.2 and GLM-5.1, highlighted AMD's advantage in VRAM capacity, with 288GB compared to B200's 180GB, allowing for more efficient long-text processing without multi-GPU setups. However, AMD's MI355X faces challenges in inter-chip connectivity, lagging behind NVIDIA's NVLink bandwidth capabilities. To address these limitations, Zyphra developed Tensor Sequence Parallelism and Tree Attention algorithms, enhancing data transfer efficiency. While NVIDIA's B200 maintains a lead in single-request speed, Zyphra's optimized software stack enables the MI355X to close the throughput gap in long-context scenarios. Zyphra plans to further enhance the MI355X's capabilities by supporting larger models and developing new quantization schemes.