中文 English

Pushing The Envelope With HBM2E Memory

Under the hood of an HBM2E memory interface with enough bandwidth for AI and HPC applications.

popularity

In September, Rambus announced the achievement of reaching 4 gigabits per second (Gbps) operation with our HBM2E memory interface. This milestone was demonstrated in silicon and required mastering substantial signal integrity and power integrity (SI/PI) challenges. The 4 Gbps mark represents a 20% rise from the previous maximum data rate of 3.2 Gbps for HBM2E.

To date, the industry’s fastest HBM2E DRAM comes from SK Hynix with operation of 3.6 Gbps. Rambus teamed with SK Hynix and Alchip to implement the HBM2E 2.5D system to validate in silicon the Rambus HBM2E interface using TSMC’s N7 process and CoWoS (Chip-on-Wafer-on-Substrate) advanced packaging technologies. Co-designing with the engineering team from Rambus, Alchip led the interposer and package substrate design.

What does this mean for our customers? Paired with the 3.6 Gbps SK Hynix HBM2E DRAM, the Rambus solution can deliver 460 gigabytes per second (GB/s) of bandwidth. In an architecture with 4 to 6 HBM2E DRAM devices, that translates to two to three terabytes of bandwidth. That enormous bandwidth meets the needs of accelerators targeting the most demanding AI/ML training and high-performance computing (HPC) applications.

If you open the hood on the Rambus HBM2E memory interface, you’ll find a fully integrated and verified PHY and memory controller. The PHY design benefits from Rambus’ 30 years of high-speed signal expertise. It’s thanks to that deep SI/PI design experience that we’re able to achieve 4G operation over a 1024-bit wide interface. With command and address, there are over 1700 traces between the PHY on the SoC and the HBM2E DRAM.

The pedigree of the Rambus HBM2E memory controller is also extremely impressive. It is a second-generation design with a track record of implementation in over 50 customer designs and nine test chips. In addition, in every case it’s been implemented with 100% first-time silicon success.

With AI/ML training models rising by 10X every year, capacity requirements are also a critical requirement for memory. The Rambus memory controller supports HBM2E DRAM with up to 12-high memory stacks. Further, the controller supports densities from 4 to 24 Gbits. With a 12-high stack and 24 Gbit density, that works out to 36 GByte per channel.

The controller is delivered configured to customer requirements to minimize size, power and latency. The Rambus HBM2E solution provides high bus efficiency across a wide variety of configurations (AXI, native interface) and traffic scenarios (random and sequential accesses; short and long bursts, etc.). It has features for reliability, availability, serviceability (RAS) support including ECC scrubbing and data path parity protection.

To ease in-lab bring-up or debug, the solution offers full featured memory test support is provided to isolate the HBM2 Memory/PHY and Controller data path prior to full system bring-up. In addition to hardware-based validation, we also have an extensive factory-based verification environment using universal verification model (UVM). We use the Samsung and SK Hynix memory models, as well as Avery Design Systems memory models and monitors. You can be assured that you’re not going to spend time debugging the PHY and controller interaction. We provide a customer focused version of this UVM testbench as part of every HBM2E controller delivery.

And we support customers with industry-leading technical expertise during pre-sales and in all project phases including design, tapeout and bring up. Our offered Lab Station development environment further accelerates the bring up process.

The rate of advancement in AI/ML is breathtaking with demands for greater bandwidth and capacity that are unrelenting. HBM2E delivers the highest bandwidth of any external memory solution by a wide margin, and Rambus offers the highest speed HBM2E interface solution. With our HBM2E memory interface, you can harness the highest level of performance using a proven, flexible and efficient memory solution backed up by world-class support.

Additional Resources:
White Paper: GDDR6 and HBM2E: Memory Solutions for AI
Web: Rambus HBM2E PHY and Rambus HBM2E Controller
Product Briefs: HBM2E PHY and HBM2E Controller
Solution Brief: HBM2E Interface Solution



Leave a Reply


(Note: This name will be displayed publicly)