High Bandwidth Memory (HBM2) Interface FPGA IP User Guide

ID 683189
Date 3/29/2024
Public
Document Table of Contents

7.2. High Bandwidth Memory (HBM2) Interface Intel® FPGA IP HBM2 IP Efficiency

The efficiency of the High Bandwidth Memory (HBM2) Interface Intel® FPGA IP estimates data bus utilization at the AXI interface. The AXI4 protocol supports independent write and read address and data channel and accepts concurrent write and read transactions. Calculated efficiency values take into consideration that the core clock frequency and the memory clock frequency are different. The HBM2 IP design example includes an efficiency monitor that reports the efficiency and the minimum latency observed in the transactions that are simulated. You can enable the efficiency monitor on the Diagnostics tab of the HBM2 IP parameter editor.

The following equation represents the HBM2 controller efficiency:

Efficiency = ((Write transactions + Read transactions accepted by HBM2 controller)/total valid transaction count) * (core clk frequency * 2/HBM2 interface frequency) * 100
  • Write transactions – Refers to user write data transactions that the HBM2 controller accepts (user-asserted AXI WVALID and corresponding controller-asserted AXI WREADY).
  • Read transactions – Refers to user read data transactions that the HBM2 controller has processed (controller-asserted AXI RVALID and corresponding user-asserted AXI RREADY).
  • Total valid transaction count – Total transaction time after first valid transaction has been issued.
  • Core frequency (MHz) – The frequency at which user logic operates. The core operates at a lower frequency than the HBM2 interface.
  • HBM2 interface frequency (MHz) - The frequency at which the HBM2 interface operates.

The HBM2 controller provides high efficiency for any given address pattern from the user interface. The controller efficiently schedules incoming commands, avoiding frequent precharge and activate commands as well as frequent bus turn-around when possible.

Factors Affecting Controller Efficiency

Several factors can affect controller efficiency. For best efficiency, you should consider these factors in your design:

  • User-interface frequency vs HBM2 interface frequency - The frequency of user logic in the FPGA fabric plays an important role in determining HBM2 memory efficiency, as shown in the example above.
  • Controller Settings
    • Disable the Reorder Buffer in the Controller Settings to achieve improved efficiency. (However, if the application requires that read data be provided in the same order as the read requests, then it is preferable to enable the Reorder Buffer.)
    • Burst length - The pseudo-BL8 mode helps to ensure shorter memory access timing between successive BL4 transactions, to improve controller efficiency.
  • Traffic Patterns - Traffic patterns play an important role in determining controller efficiency.
    • Sequential vs random DRAM addresses: Sequential addresses enable the controller to issue consecutive write requests to an open page and help to achieve high controller efficiency. Random addresses require constant PRECHARGE/ACTIVATE commands and can reduce controller efficiency.
    • Set the User Auto Precharge Policy to FORCED and set the awuser/aruser signal on the AXI interface to HIGH to enable Auto Precharge for random transactions. For sequential transactions, set the Auto Precharge Policy to HINT.
    • Sequential Read only or Write Only transactions: Sequential read-only or write-only transactions see higher efficiency as they avoid bus turnaround times of the DRAM bi-directional data bus.
  • AXI Transaction IDs – Use of different AXI transaction IDs helps the HBM2 controller schedule the transactions for high efficiency. Use of the same AXI transaction ID preserves command order and may result in lower efficiency.
  • Temperature – There are two main temperature effects that reduce the bandwidth available for data transfers in the Stratix® 10 HBM2 interface:
    • The HBM2 is exceeding a temperature threshold set in the General > Threshold temperature for AXI throttling parameter, along with a value for the AXI throttling ratio.
    • Above 85°C, HBM2 refreshes occur more frequently, so there is less bandwidth available for data transfers. At 85°C, the refresh rate is 2×, and at 95°C, 4× more frequent than the standard refresh time interval of 7.8 microseconds.
    The HBM2 TEMP register value can be read back via the AXI APB bus to help with the correlation of memory refreshes, temperature, and data bandwidth.