Case Study: RadixArk achieves 3.5K+ tokens/sec/GPU for DeepSeek V4 with DigitalOcean AMD GPU Droplets

RadixArk serves DeepSeek V4 at 3.5K+ tokens/sec/GPU with DigitalOcean

RadixArk, an AI infrastructure company, needed dedicated GPU capacity to develop and validate its SGLang inference engine for running frontier models like DeepSeek V4 in production. They required reliable, on-demand access to AMD hardware to iterate quickly without being locked into a single vendor. They turned to DigitalOcean for its AMD Instinct GPU Droplets.

DigitalOcean provided RadixArk with immediate access to AMD Instinct MI350X GPU Droplets, which were ready in minutes. A joint engineering effort with AMD enabled HIP graph support, which resulted in a 10x throughput improvement. This allowed RadixArk to achieve a performance of over 3.5K tokens/second/GPU running DeepSeek V4, proving their open-source engine could serve models efficiently on multiple hardware platforms in production.

RadixArk