Rambus wants your AI superchip to stop choking on memory bandwidth. The outfit has rolled out its fastest HBM controller yet, built to the HBM4E standard and rated at up to 16Gbps per pin.
Rambus says the controller is aimed at next-generation AI data centre “superchips”, with a claimed 60 per cent uplift on its HBM4 controller. It is talking 16Gbps per pin versus 10Gbps for HBM4, and up to 4.1TB/s total bandwidth per module versus 2.56GB/s quoted for HBM4.
The HBM4E spec is expected to show up in Nvidia Rubin Ultra GPUs and AMD’s MI500 series accelerators.
In its press release, Rambus billed the part as the “industry’s leading HBM4E Memory Controller IP”, and said it stretches its lead in HBM IP with performance and reliability upgrades for demanding accelerators and GPUs.
It relies on a claimed track record of more than 100 HBM design wins to sell “first-time silicon success” and says the IP delivers up to 16 gigabits per second per pin with low latency.
Rambus said the controller expands its portfolio of high-performance memory IP, which is corporate speak for “pay us and ship it”.
Rambus reckons the controller enables a new wave of HBM deployments for AI accelerators, graphics and HPC, pushing up to 4.1 terabytes per second to each memory device.
Pair an AI accelerator with eight HBM4E devices, and Rambus says that means more than 32TB/s of memory bandwidth for next-generation workloads, assuming your design does not trip over itself first. The controller can be paired with third-party standard or TSV PHY options to build a full HBM4E subsystem in a 2.5D or 3D package, as part of an AI SoC or a custom base die.
Rambus says the HBM4E Controller IP is now available for licensing, and early-access design customers can get stuck in today, with the bandwidth arms race set to get noisier.
