According to technology media outlet TechRadar, startup Majestic Labs recently announced the launch of the Prometheus AI server. The new product combines AIUs with LPDDR6 memory and aims to challenge NVIDIA’s GPU+HBM combination.
According to reports, the Prometheus server does not use GPUs, but instead uses Ignite AI processing units (AIUs), which combine Arm cores with RISC-V vector and tensor engines. Each Prometheus can be configured with up to 12 AIUs and 8TB–128TB of LPDDR6 memory.
At the same time, a standard 40U rack can accommodate up to 4 Prometheus servers, with total power consumption of 120kW. It uses a liquid-cooling system instead of traditional air cooling. The company claims that a single Majestic rack offers the equivalent memory capacity of 25 NVIDIA NVL72 Vera Rubin racks, while consuming less power than NVIDIA’s solution.
In addition, the server is expected to launch next year. It complies with OCP (Open Compute Project) standards and is compatible with the PyTorch, vLLM, and OpenAI Triton frameworks.