Critical Issue
If the Hybrid Memory Cube (HMC) Controller IP core receives a packet in the four FLITs that follow an incoming error recovery sequence (20 consecutive IRTRY packets), the forward retry pointer (FRP) of the packet might not appear in the reverse retry pointer field (RRP) of an outgoing packet.
If the HMC detects a second error before it receives the retry pointer, it resends the packet and the HMC Controller IP core processes the packet twice. The impact depends on the type of the packet that is processed twice. For example, if the twice-processed packet is a TRET packet, input buffer overflow could occur later in the HMC device. If the twice-processed packet is a read response packet, the IP core passes it to the data path response application interface, where it appears as an unexpected response.
This issue is a special case of another related knowledge article: HMC Controller IP Core Does Not Properly Process HMC Packets Received Immediately Following a Short Retry Sequence.
This issue affects all full-width variations of the Hybrid Memory Cube Controller IP core.
To avoid this issue, configure the Hybrid Memory Cube to send at least 24 IRTRY packets in a retry sequence. To configure this property, set the Init retry packet transmit number field of the HMC device Link Retry register to the value of 7 or higher. However, to avoid the general issue stated in this related knowledge article: HMC Controller IP Core Does Not Properly Process HMC Packets Received Immediately Following a Short Retry Sequence, you must implement the workaround for that erratum: force at least 50 IRTRY packets by setting Init retry packet transmit number to the value of 13 or higher.
This issue will be fixed in a future version of the HMC Controller IP core.