HBM Bit ECC Fault

Symptom

The event logs on the device (slog/dev-os-id/run/event/event_*.log) contain the keyword event_id=0x80E01809 or event_id=0x80E01801.

The device black box logs (in the hisi_logs /device-id /*/bbox directory) contain the keyword "Hardware Error."

/hisi_logs/device-2/20240420162417-271713000/bbox/kbox.txt:2014[2454.830692] {6}[Hardware Error] Hardware error from APEI Generic Hardware Error Source: 0
/hisi_logs/device-2/20240420162417-271713000/bbox/kbox.txt:2015[2454.830693] {6}[Hardware Error]event severity: recoverable
/hisi_logs/device-2/20240420162417-271713000/bbox/kbox.txt:2016[2454.830694] {6}[Hardware Error] Error 0, type: recoverable
/hisi_logs/device-2/20240420162417-271713000/bbox/kbox.txt:2017[2454.830696] {6}[Hardware Error]  section_type: memory error
/hisi_logs/device-2/20240420162417-271713000/bbox/kbox.txt:2018[2454.830697] {6}[Hardware Error]  physical_address: 0x0000101efe36d3c0
/hisi_logs/device-2/20240420162417-271713000/bbox/kbox.txt:2019[2454.830699] {6}[Hardware Error]  node: 2 card: 259 module: 51 rank: 1 bank: 10 row: 30691  column: 56
/hisi_logs/device-2/20240420162417-271713000/bbox/kbox.txt:2020[2454.830701] {6}[Hardware Error]  error_type: 3, multi-bit ECC
/hisi_logs/device-2/20240420162417-271713000/bbox/kbox.txt:2024[2454.830702] {6}[Hardware Error]  DIMM location: not present. DMI handle: 0x0000

Fault Root Causes

Based on the error information in the preceding logs, it can be determined that the AI Core error is caused by the HBM bit ECC fault.

Solution

Click Health Management Fault Definition to obtain the manual of the corresponding version. The HBM bit ECC fault description is as follows (some key fields are listed):

Event ID

0x80E01809

Event Name

Multi-bit ECC errors during the HBM patrol

Fault Description/Possible Cause

A multi-bit ECC error is triggered during HBMC patrol scrubbing and demand scrubbing. The possible cause is that some HBM chips fail or the HBM cannot store data properly. This is usually caused by hardware faults.

Impact

1. If an incorrect address is accessed during the startup, the startup may fail.

2. If a service accesses an incorrect address, error data is returned, which may cause service failures.

3. The service does not access an incorrect address, and the current service is not affected.

System Action

1. Reports a notification event to the device management module, records the incorrect address, attempts to isolate the device online, and performs offline isolation after the device is restarted.

2. Records error logs.

System Handling Suggestion

No operation is required.

Event ID

0x80E01801

Event Name

Multi-bit ECC errors when accessing the HBM space

Fault Description/Possible Cause

A multi-bit ECC error is triggered and address hot isolation fails when the HBM space is accessed. The possible cause is that some HBM chips fail and the failed addresses are still occupied. This is usually caused by hardware faults.

Impact

Incorrect data is returned when accessing the HBM, which may cause NPU startup failures or service failures.

System Action

1. Reports fault events to the device management module.

2. Records error logs.

System Handling Suggestion

If the reported host PID is 0, you are advised to reset the SoC. If the reported host PID is not 0, kill the process and wait for a restoration period. If the fault persists, you are advised to reset the SoC.