India’s Aum HPC Processor: Architecture, Specifications, and Market Position

India’s C‑DAC has unveiled the Aum HPC processor, a 48‑core Arm‑based chip built on the Ze​us Neoverse V1 core, aiming to challenge Fujitsu’s 48‑core A64FX used in the Fugaku supercomputer and Nvidia’s upcoming 72‑core Grace CPU.

The Aum design, supported by the Indian government’s national supercomputing initiative, targets both high‑performance computing (HPC) and AI workloads, with an expected launch in 2024 for pilot systems exceeding 1 petaflop theoretical peak performance.

Each Aum package (A48Z) integrates two 48‑core small dies connected via a fully coherent D2D interconnect, providing 96 MB of L2 cache and an additional 96 MB of shared system‑level cache. The memory subsystem includes eight DDR5 channels (up to 5.2 GHz) and a pair of HBM3 controllers (initially 5.6 GHz), delivering up to 332.8 GB/s DDR5 bandwidth and 2.87 TB/s HBM3 bandwidth.

The chip operates at a base clock of 3 GHz with turbo speeds above 3.5 GHz, achieving a thermal design power of roughly 300 W while delivering about 10 TFLOPS per node; a dual‑socket server can support four GPU accelerators.

Beyond DDR5 and HBM3, the design incorporates 64 PCI‑Express 5.0 lanes per small die (128 total) and supports CCIX‑based chip‑to‑chip (C2C) interconnects for NUMA scaling across multiple sockets, similar to AMD’s Infinity Fabric approach.

By leveraging Arm’s V8.4‑A Neoverse V1 cores rather than the newer V9, the Aum processor emphasizes a balance of memory bandwidth, compute density, and supply‑chain independence, aligning with India’s strategic goal of HPC autonomy without foreign backdoors.

Overall, the Aum processor represents India’s push for a domestically sourced, high‑performance, and AI‑ready CPU platform, positioned to compete with global offerings such as Fujitsu’s A64FX, Nvidia’s Grace, AWS’s Graviton 3, and Ampere’s Altra series.