Design & Reuse

ComputeRAM in AI accelerators: An LLM case study

synthara.ai, May. 13, 2025 – 

In the rapidly evolving landscape of AI data centers, system designers face two critical challenges: maximizing compute efficiency (TOP/s/W) and minimizing silicon cost (TOP/mm²). Synthara’s latest study showcases how ComputeRAM (CxR), our revolutionary in-memory computing IP, dramatically addresses both challenges. We demonstrate that integrating ComputeRAM into a state-of-the-art AI accelerator can achieve a 3.7x improvement in compute efficiency (TOP/s/W), significantly reducing operational costs related to energy consumption, and a 20% increase in compute density (TOP/mm²), substantially increasing compute capability in each server rack.
The analysis utilizes a reference architecture modeled after industry-leading accelerators such as Nvidia Tensor Cores, AMD Matrix Core Units, Tenstorrent Tensix Cores, and Graphcore IPUs. ComputeRAM’s integration into this architecture proved seamless and non-disruptive, ensuring straightforward adoption without altering existing software or hardware frameworks. This performance leap holds profound implications for datacenters investing billions in AI accelerator technologies, potentially translating into savings of hundreds of millions of dollars in capital expenditure (CapEx) and operational expenses (OpEx).

Discover how ComputeRAM can elevate your AI accelerator designs and significantly reduce your datacenter costs. Read the full report here or contact us directly at business@synthara.ai for further information.