HBM+ Logic Die Offloads Host Processing

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current High Bandwidth Memory (HBM) systems face challenges in processing complex logic operations due to their asynchronous communication nature, which leads to uncertainty in operation completion time and reduced determinism, making it difficult to achieve energy efficiency and low latency in applications like deep neural networks and high-performance computing.

Innovation Solution

The introduction of an HBM+ system with a logic die positioned beneath the HBM modules, equipped with a host manager, memory controller, prefetch engine, cache controller, and offload processing logic, which offloads processing operations from the host, enabling deterministic and efficient processing through specialized logic functions and signaling protocols.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If asynchronous communication interface is used in HBM, then bandwidth and power efficiency are improved, but determinism and processing reliability for complex logic operations deteriorate

Engineering Contradiction:
ImprovebandwidthVSAvoiddeterminism
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The system is segmented into distinct functional components: the HBM stack handles high-bandwidth data storage and transfer, while the separate logic die handles complex logic operations. This segmentation allows each component to be optimized for its specific function, with the logic die providing deterministic processing for complex operations while the HBM stack maintains high bandwidth through asynchronous communication.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The logic die acts as an intermediary between the host and the HBM stack, bridging the gap between asynchronous memory operations and synchronous processing requirements. It receives data from the HBM stack, performs deterministic logic operations, and returns results to the host, thus mediating between the conflicting requirements of high bandwidth and determinism.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Adaptability or versatility

If complex logic operations are performed by the host, then processing flexibility is maintained, but latency and energy consumption increase

Engineering Contradiction:
Improveprocessing flexibilityVSAvoidlatency
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

Complex logic operations are extracted from the host processor and offloaded to the dedicated logic die. This extraction removes the latency burden from the host, allowing it to focus on high-level control and data preparation, while the logic die handles time-intensive logic operations in parallel, thereby reducing overall system latency without sacrificing processing flexibility.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The system transitions from a single-processor architecture to a multi-dimensional architecture where the logic die operates in parallel with the host. This dimensional change allows complex logic operations to be performed simultaneously with host operations, reducing latency through parallel processing while maintaining the host's processing flexibility through independent control planes.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Productivity

If more processing operations are offloaded to the logic die, then host workload is reduced, but device complexity increases

Engineering Contradiction:
Improveprocessing throughputVSAvoidarchitecture complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The logic die is designed with multi-functional capabilities, incorporating both storage functions and logic processing functions in a single integrated component. This universality allows it to handle diverse processing operations without requiring additional specialized hardware, thereby increasing processing throughput while limiting the growth of overall device complexity through functional consolidation.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentUS10545860B2Intelligent high bandwidth memory appliance
Publication Date: 2020.01.28 SAMSUNG ELECTRONICS CO LTD
  • US10545860B2 patent drawing
  • US10545860B2 patent drawing
  • US10545860B2 patent drawing

AI summary

Inventive aspects include An HBM+ system, comprising a host including at least one of a CPU, a GPU, an ASIC, or an FPGA; and an HBM+ stack including a plurality of HBM modules arranged one atop another, and a logic die disposed beneath the plurality of HBM modules. The logic die is configured to offload processing operations from the host. A system architecture is disclosed that provides specific compute capabilities in the logic die of high bandwidth memory along with the supporting hardware and software architectures, logic die microarchitecture, and memory interface signaling options. Various new methods are provided for using in-memory processing abilities of the logic die beneath an HBM memory stack. In addition, various new signaling protocols are disclosed to use an HBM interface. The logic die microarchitecture and supporting system framework are also described.