REACH Aims to Reduce HBM ECC Controller Overhead for AI Inference
Researchers at Rensselaer Polytechnic Institute and IBM T.J. Watson Research Center have developed a new technical paper titled 'REACH: Controller-Managed Long-Span ECC for HBM AI Inference.'
The paper, published in September 2026, addresses the high cost of High-Bandwidth Memory (HBM) by proposing a stronger controller protection mechanism that can support a wider range of device error rates.
Current long-span error-correcting codes provide strong protection but require costly decoding at HBM bandwidth. The researchers propose using established inner codes to correct common errors and identify unresolved chunks, reserving a long outer code for known-erasure repair.
The proposed REACH microarchitecture takes advantage of the sequential read-dominated nature of Large Language Model (LLM) decode operations, aggregating span-wide state while limiting parity-update traffic.