Hybrid Memory Cube Controller IP Core User Guide - Intel Stratix 10 Beta Version

ID 683854
Date 8/08/2016
Public
Document Table of Contents

3.6. Flow Control

The HMC specification describes two possible flow control schemes for traffic between the host and the HMC memory device, token based flow control and Response Open Loop Mode.

In token-passing mode, the host sends information about its buffering capacity to the HMC device during transaction layer initialization. In Response Open Loop Mode, the host does not send information about its buffering capacity to the HMC device. Instead, the host only sends a request packet when it has room to receive the response at any time.

The HMC Controller IP core operates in Response Open Loop Mode. The IP core is designed to have the capacity to accept all response packets from the HMC device.

When user requests come in faster than the HMC Controller IP core can send them out on the HMC link, the HMC Controller IP core backpressures the application by deasserting the dp_req_ready signal.

In addition, the IP core supports the Limit Outstanding FLITs feature. You can turn on this feature to ensure that the IP core limits the number of outstanding FLITs in expected read responses. The IP core delays sending a request to the HMC if sending it would increase the number of pending response FLITs from the HMC above a user-specified threshold.

The Limit Outstanding FLITs feature limit the worst-case round-trip latency of packets in the system by avoiding the congestion that can occur when response packets have large payloads. A sequence of RD128 requests, for example, causes congestion in the response path. Designers building read-intensive, latency-sensitive applications can use the Limit Outstanding FLITs feature to improve latency and throughput.

Figure 15. Improving Latency and Throughput Flow Chart