Overview
This datasheet describes GUC HBM (High Bandwidth Memory) PHY IP, which could be integrated with HBM memory controller to provide HBM functionality. The HBM PHY is compliant to the JEDEC HBM2E & HBM2 standard. In 5nm, it supports data rates up to 3600 Mbps per data pin at DFI 1:2 mode. A DFI interface is provided by this HBM PHY IP to connect to HBM memory controller IP. Power information is described at subsection 3.2.1 in detail. Complete signal and power integrity analysis are verified on the GUC design flow to make sure all signal and power requirements are met.
HBM PHY IP includes one hard macro and one RTL module. The hard macro called as IGAHBMY02A is HBM Hard PHY includes command address module, data module, IO pads, PLL and DLL components required for HBM PHY function. The RTL module called as MLB (miscellaneous logic block) is provided to work with HBM Hard PHY. MLB RTL module includes functions such as training logic, JTAG controller interface for PHY DFT registers and BIST(Built-In Self Test) logic. HBM PHY evaluation mode is provided through MLB RTL module without the help of controller to complete the training sequences such as read data EYE training and write data EYE training which are required to guarantee HBM PHY IP functions. MLB training procedure is default enable in initial sequence to set HBM PHY by hardware as described in subsection 3.4.3. This MLB RTL module shall be integrated in SOC by users to work with HBM Hard PHY correctly.
Learn more about HBM IP core
As AI and high-performance SoCs increasingly rely on HBM, memory bandwidth alone is no longer enough to maximize performance. This article discusses why the intelligent data movement and cache efficiency are critical to unlocking the full benefits of HBM-based architectures.
LLM inference is increasingly memory-bound, and HBM cost per GB now dominates system cost. Today’s HBM stacks include short on-die ECC, which tightens binning, raises price, and locks reliability policy inside the device. This paper asks a simple question: can we tolerate a much higher raw HBM bit error rate (BER) and still keep end-to-end correctness and throughput, without changing the HBM PHY or the fixed 32B transaction size?
HBM4 is the latest generation of the High Bandwidth Memory (HBM) that has become analogous to the Artificial Intelligence (AI) boom that is everywhere in today’s world. HBM is also increasingly being used in other applications like Data centers, autonomous driving systems, servers, cloud computing just to mention few domains where bandwidth and performance in a key requirement.
High-Bandwidth Memory (HBM) delivers exceptional bandwidth and energy efficiency for AI workloads, but its high cost per bit, driven in part by stringent on-die reliability requirements, poses a growing barrier to scalable deployment. This work explores a systemlevel approach to cost reduction by eliminating on-die ECC and shifting all fault management to the memory controller.
As artificial intelligence (AI) reshapes industries and advances technological frontiers, its success hinges on advanced memory capabilities. Leading this transformation is High Bandwidth Memory (HBM), which offers unparalleled speeds and efficiencies.
In wide chip interfaces like DDR, HBM and ONFI, it can be challenging to synthesize and connect high-frequency controllers to the PHY hard macros. Clock trees can be expansive, pushing tools to their limits, and often multiple clock domains are needed.
