Huge Memory AI Server Aims to Shatter the Memory Wall

https://spectrum.ieee.org/huge-memory-ai-server

Publish Date: 2026-06-01 11:00:01

Summary:
Modern AI large language models (LLMs) suffer from a significant memory constraint limiting their performance, a phenomenon known as the “memory wall.” To tackle this challenge, AI hardware startup Majestic Labs has unveiled its Prometheus server, equipped with an unprecedented 128 terabytes of memory—far surpassing current competition from companies like Nvidia. Prometheus adopts a DRAM-centric architecture using a proprietary high-speed interface and aggregation chips to address this memory overcapacity. The server is engineered around Majestion’s custom AI processing unit called Ignite, which blends ARM cores with RISC-V vector and tensor cores for enhanced LLM processing. Designed to be Open Compute Project-compliant, Prometheus promises to reduce both capital expenditure and power consumption significantly. Expected to ship in 2027, Majestic Labs aims to make this advanced memory solution both affordable and easy to integrate for existing models with popular frameworks.

Key Points:

Large language models are hampered by significant memory constraints, referring to the memory wall phenomenon.
Majestic Labs’ Prometheus server aims to solve this issue through an unprecedented 128 TB of DRAM.
Prometheus uses a unique DRAM-centric architecture with an innovative proprietary memory interface.
The server features Majestion’s custom Ignite processor, which unifies ARM, RISC-V cores for efficient AI model inference.
Prometheus promises substantial reductions in both power consumption and capital expenditure, positioned to be both powerful and affordable.