3. C600 register map

3.1. API version history

When commands are added to or removed from the C600 firmware, the API version will be incremented. The API version at which each command or field has been added is indicated in the following tables to ensure that the user can identify which information is expected. The API version is itself available via command 0x03, see API version for details.

API Version	ICU Version
1	2.6.3
2	2.6.7

3.2. Command list

A full listing of all SMBus commands is given in Table 3.1. More details about each of the commands can be found in Section 3.3, Command descriptions.

Table 3.1 SMBus commands

ID	Name	Bytes (r/w)	Protocol	API
0x01	Vendor ID	2	Read Word	1
0x02	Product ID	2	Read Word	1
0x03	API version	2	Read Word	1
0x04	ICU major version	2	Read Word	1
0x05	ICU minor version	2	Read Word	1
0x06	ICU patch version	2	Read Word	1
0x07	ICU version string	31 / 1	Block Process Call	1
0x08	Board public ID	24	Block Read	1
0x09	Board revision	22	Block Read	1
0x0A	Board PCB information	2	Read Word	2
0x10	System uptime	8	Block Read	1
0x12	POST status	4	Block Read	1
0x13	IPU errors	24 / 1	Block Read or Write	1
0x15	dmesg update	2	Write Word	2
0x16	dmesg read	30 / 2	Block Process Call	2
0x17	IPU status	2	Read Word	2
0x20	Clock frequency	2	Read Word	1
0x22	Frequency limiter	2 / 2	Block Read or Write	1
0x23	Board TDP limit	2 / 2	Block Read or Write	2
0x31	Board power consumption	2	Read Word	2
0x40	Temperatures - max board	2	Read Word	1
0x41	Temperatures - group A	24	Block Read	1
0x42	Max temperatures - group A	24	Block Read	1
0x43	Temperature thresholds	24	Block Read	1
0x45	Thermal status	6	Block Read	1
0x80	Driver error state	4	Block Read	1

3.3. Command descriptions

3.3.1. System information

Commands 0x00 - 0x0F are reserved for system information. These values should be constant once the ICU has booted.

ID	Name	Bytes (r/w)	Protocol	API
0x01	Vendor ID	2	Read Word	1
0x02	Product ID	2	Read Word	1
0x03	API version	2	Read Word	1
0x04	ICU major version	2	Read Word	1
0x05	ICU minor version	2	Read Word	1
0x06	ICU patch version	2	Read Word	1
0x07	ICU version string	31 / 1	Block Process Call	1
0x08	Board public ID	24	Block Read	1
0x09	Board revision	22	Block Read	1
0x0A	Board PCB information	2	Read Word	2

3.3.1.1. Vendor ID (0x01)

Read word command 0x01 returns a data word fixed to contain the Graphcore PCIe vendor ID, and used to verify that the I2C device is a Graphcore ICU.

Response:

Byte	Name	Format	Value	API
1-2	Vendor ID	uint16_t	0x1d95	1

3.3.1.2. Product ID (0x02)

Read word command 0x02 returns a data word fixed to contain the PCIe device ID of the C600 card, and used to verify that the I2C device is a Graphcore C600 PCIe card.

Response:

Byte	Name	Format	Value	API
1-2	Product ID	uint16_t	0x0600	1

3.3.1.3. API version (0x03)

Read word command 0x03 returns a data word containing the currently supported version of the SMBus command API - see API version history.

Response:

Byte	Name	Format	Unit	API
1-2	API version	uint16_t		1

3.3.1.4. ICU major version (0x04)

Read Word command 0x04 returns a data word containing the running firmware’s major version number.

Response:

Byte	Name	Format	Unit	API
1-2	ICU major version	uint16_t		1

3.3.1.5. ICU minor version (0x05)

Read Word command 0x05 returns a data word containing the running firmware’s minor version number.

Response:

Byte	Name	Format	Unit	API
1-2	ICU minor version	uint16_t		1

3.3.1.6. ICU patch version (0x06)

Read word command 0x06 returns a data word containing the running firmware’s patch version number.

Response:

Byte	Name	Format	Unit	API
1-2	ICU patch version	uint16_t		1

3.3.1.7. ICU version string (0x07)

Block process call command 0x07 returns up to 31 data words containing the running firmware’s null-terminated full version string, starting at the index supplied. If the returned string is not null-terminated, more characters can be read by repeating the command after increasing the argument index given in the block write command, and concatenated to build up the full string.

Block write command:

Byte	Name	Format	Unit	API
1	Index	uint8_t		1

Block read response:

Byte	Name	Format	Unit	API
1-31	Version string	ASCII		1

3.3.1.8. Board public ID (0x08)

Block read command 0x08 returns a 24 byte null-terminated string containing the C600 card’s human readable board description.

Response:

Byte	Name	Format	Unit	API
1-24	Board public ID	ASCII		1

3.3.1.9. Board revision (0x09)

Block read command 0x09 returns a 22 byte null-terminated string containing the C600 card’s board identification string, also known as the card serial number.

Response:

Byte	Name	Format	Unit	API
1-22	Board revision	ASCII		1

3.3.1.10. Board PCB information (0x0A)

Read Word command 0x0A returns a data word containing the PCB and BOM identification information for the C600 card.

Response:

Byte	Name	Format	Unit	API
1	PCB identifier	uint8_t		2
2	BOM identifier	uint8_t		2

3.3.2. Active system state

Commands 0x10 - 0x1F are reserved for the active system state. These values may change as the system is running.

ID	Name	Bytes (r/w)	Protocol	API
0x10	System uptime	8	Block Read	1
0x12	POST status	4	Block Read	1
0x13	IPU errors	24 / 1	Block Read or Write	1
0x15	dmesg update	2	Write Word	2
0x16	dmesg read	30 / 2	Block Process Call	2
0x17	IPU status	2	Read Word	2

3.3.2.1. System uptime (0x10)

Block Read command 0x10 returns a 64-bit integer containing the ICU’s uptime in milliseconds.

Response:

Byte	Name	Format	Unit	API
1-8	System uptime	int64_t	millisec	1

3.3.2.2. POST status (0x12)

Block read command 0x12 returns the ICU’s Power On Self Test (POST) result. Outline details for reading this result are below, however any result other than 0x0 indicates an issue was encountered during the boot sequence.

Response:

Byte	Name	Format	Unit	API
1-4	POST status	int32_t	bitfield	1

POST status bitfield:

Bit	Description
31	Failure bit. 1 if failed else 0
30	Warning bit. 1 if multiple warnings else 0
29:24	Reserved
23:16	First encountered Zephyr error code from failed/warning code unit (truncated to 8-bit)
15:8	First encountered failure/warning number from failed/warning code unit (unit specific)
7:0	gc_init_id number associated with the failed/warning code unit

3.3.2.3. IPU errors (0x13)

3.3.2.3.1. Update errors

Block write command 0x13 triggers the ICU to check for logged errors, according to the provided update request type. Up to 1000 error records can be maintained within the ICU

Block write command:

Byte	Name	Format	Unit	API
1	Update request	uint8_t	enum	1

Update request enumeration:

Value	Name	Description	API
0	Next	Ready the next error log to be read, and increments read index	1
1	Most recent	Ready the most recent log to be read. Does not update	1
2	Mark all read	Updates the read index to point to the most recent error log	1

3.3.2.3.2. Read errors

Block read command 0x13 retrieves 24 bytes of data describing the requested error log. If no more errors are available, this command will return all zeros excluding the State field. The ICU needs time to perform the action of a preceeding Block Write command - if the read is performed too soon, State may return Updating and the Block Read command should be repeated after a short delay. The error information consists of three error registers and a timestamp indicating when the error was detected.

Block Read command:

Byte	Name	Format	Unit	API
1-4	Timestamp	uint32_t	seconds	1
5-8	CMGMTEVVR	uint32_t	bitfield	1
9-12	CICERRVR	uint32_t	bitfield	1
13-16	CIUERRVR	uint32_t	bitfield	1
17-18	Index	uint16_t		1
19-20	Remaining	uint16_t		1
21-22	Bootcount	uint16_t		1
23	Source	uint8_t	enum	1
24	State	uint8_t	enum	1

Source enumeration:

Value	Name	Description	API
0	No records	No (more) errors to report	1
1	Newman reset	Logged during a Newman reset operation	1
2	Shutdown	Logged during an ICU shutdown operation	1
3	Monitor	Logged during a regular monitoring operation	1
4	Shellcheck	Logged via a development interface	1

State enumeration:

Value	Name	Description	API
1	Updating	Previous request is still being handled	1
2	Updated	Error log is ready and available	1
3	No records	No logs have been recorded	1
4	Error	An error occurred while reading the error logs	1

3.3.2.4. dmesg update (0x15)

Write Word command 0x15 triggers the ICU to cache its dmesg log. Up to 100 log messages can be cached. The most recent messages will be cached if there are more than 100 logs. The update request input can be any numeric value.

Write Word command:

Byte	Name	Format	Unit	API
1-2	Update request	uint16_t		2

3.3.2.5. dmesg read (0x16)

Block Process Call command 0x16 returns up to 29 ASCII bytes containing the ICU’s dmesg, starting at the log and message indexes supplied, along with the update state. If the returned string is not null-terminated, more characters can be read by repeating the command after increasing the argument Message index given in the Block Write command, and concatenated to build up the full string. An array of strings can be built up in a similar fashion by increasing the argument Log index. If the first character of the returned string is null-terminated, the end of the log cache has been reached.

Block Write command:

Byte	Name	Format	Unit	API
1	Message index	uint8_t		2
2	Log index	uint8_t		2

Block Read response:

Byte	Name	Format	Unit	API
1	State	uint8_t	enum	2
2-30	dmesg string	ASCII		2

State enumeration:

Value	Name	Description	API
0	Updating	Previous request is still being handled	2
1	Updated	dmesg cache is ready and available	2
2	Index error	Log index out of range	2
3	Error	An error occurred while reading the dmesg	2

3.3.2.6. IPU status (0x17)

Read Word command 0x17 returns a data word containing a bitfield representation of the IPU’s current status.

Response:

Byte	Name	Format	Unit	API
1-2	IPU status	bitfield		2

IPU status bitfield:

Bit	Description
15:3	Reserved
2	PCIe Bus Master Enable detected
1	PCIe Power Brake activated
0	IPU in use

3.3.3. Clock control

Commands 0x20 - 0x2F are reserved for clock control information and operations.

ID	Name	Bytes (r/w)	Protocol	API
0x20	Clock frequency	2	Read Word	1
0x22	Frequency limiter	2 / 2	Block Read or Write	1
0x23	Board TDP limit	2 / 2	Block Read or Write	2

3.3.3.1. Clock frequency (0x20)

Read word command 0x20 returns the status of the IPU’s current upper clock speed.

Response:

Byte	Name	Format	Unit	API
1-2	Fast PLL IPU	uint16_t	MHz	1

3.3.3.2. Frequency limiter (0x22)

Block Read or Block Write commands 0x22 are used to read or set a limit to the maximum frequency that the IPU can run at, with an upper bound of the system default maximum frequency. If set to 0xFFFF then the system default will be used.

Block read:

Byte	Name	Format	Unit	API
1-2	Frequency limit	uint16_t	MHz	1

Block write:

Byte	Name	Format	Unit	API
1-2	Frequency limit	uint16_t	MHz	1

3.3.3.3. Board TDP limit (0x23)

Reads or sets the limit for the maximum thermal design power (TDP) that the IPU on the board can consume. This value is rounded down to the nearest multiple of 12. For example a setting of 203 watts is rounded to 192 watts.

Block read:

Byte	Name	Format	Unit	API
1-2	TDP limit	uint16_t	Watts	2

Block write:

Byte	Name	Format	Unit	API
1-2	TDP limit	uint16_t	Watts	2

3.3.4. Telemetry - power usage

Commands 0x30 - 0x3F are reserved for IPU and C600 power usage information.

ID	Name	Bytes (r/w)	Protocol	API
0x31	Board power consumption	2	Read Word	2

3.3.4.1. Board power consumption (0x31)

Read Word command 0x31 returns the average board power consumption over the last second. This is a digital average of readings from the on-board sensors, which provide an analogue average of the power usage over 63ms periods.

Response:

Byte	Name	Format	Unit	API
1-2	Board power consumption	Linear11	Watts	2

3.3.5. Telemetry - temperatures

Commands 0x40 - 0x4F are reserved for IPU and C600 temperature information.

ID	Name	Bytes (r/w)	Protocol	API
0x40	Temperatures - max board	2	Read Word	1
0x41	Temperatures - group A	24	Block Read	1
0x42	Max temperatures - group A	24	Block Read	1
0x43	Temperature thresholds	24	Block Read	1
0x45	Thermal status	6	Block Read	1

3.3.5.1. Temperatures - max board (0x40)

Read Word command 0x40 returns the current maximum temperature reported by any of the temperature sensors on the C600, in whole degrees Celsius. This value should be used to control system cooling.

Response:

Byte	Name	Format	Unit	API
1-2	Max board temp (current)	int16_t	Celsius	1

3.3.5.2. Temperatures - group A (0x41)

Block Read command 0x41 reports an aggregated list of all available temperature sensors on the C600 card. All values are reported in a Linear11 representation of Celcius allowing for decimal degrees. IPU PVT values refer to sensors within the IPU chip itself, while other values are located on the C600 card at their given location.

Response:

Byte	Name	Format	Unit	API
1-2	IPU PVT east	Linear11	Celsius	1
3-4	IPU PVT west0	Linear11	Celsius	1
5-6	ADC inlet	Linear11	Celsius	1
7-8	ADC exhaust	Linear11	Celsius	1
9-10	ADC phase0 bottomside	Linear11	Celsius	1
11-12	ADC phase1 bottomside	Linear11	Celsius	1
13-14	ADC IPU chip bottomside	Linear11	Celsius	1
15-16	ADC mid	Linear11	Celsius	1
17-18	I2C inlet	Linear11	Celsius	1
19-20	I2C IPU chip	Linear11	Celsius	1
21-22	I2C exhaust	Linear11	Celsius	1
23-24	I2C mid	Linear11	Celsius	1

3.3.5.3. Max temperatures - group A (0x42)

Block read command 0x42 reports an aggregated list of the maximum temperatures reported by all available temperature sensors on the C600 card since the last card powercycle.

Response:

Byte	Name	Format	Unit	API
1-2	IPU PVT east	Linear11	Celsius	1
3-4	IPU PVT west0	Linear11	Celsius	1
5-6	ADC inlet	Linear11	Celsius	1
7-8	ADC exhaust	Linear11	Celsius	1
9-10	ADC phase0 bottomside	Linear11	Celsius	1
11-12	ADC phase1 bottomside	Linear11	Celsius	1
13-14	ADC IPU chip bottomside	Linear11	Celsius	1
15-16	ADC mid	Linear11	Celsius	1
17-18	I2C inlet	Linear11	Celsius	1
19-20	I2C IPU chip	Linear11	Celsius	1
21-22	I2C exhaust	Linear11	Celsius	1
23-24	I2C mid	Linear11	Celsius	1

3.3.5.4. Temperature thresholds (0x43)

Block Read command 0x43 reports an aggregated list of the temperature thresholds used by the different temperature sensors present on the C600 card.

Response:

Byte	Name	Format	Unit	API
1-2	PVT emergency threshold	Linear11	Celsius	1
3-4	PVT warning threshold	Linear11	Celsius	1
5-6	Thermal control maximum	Linear11	Celsius	1
7-8	Thermal control minimum	Linear11	Celsius	1
9-10	I2C inlet emergency	Linear11	Celsius	1
11-12	I2C inlet warning	Linear11	Celsius	1
13-14	I2C IPU chip emergency	Linear11	Celsius	1
15-16	I2C IPU chip warning	Linear11	Celsius	1
17-18	I2C exhaust emergency	Linear11	Celsius	1
19-20	I2C exhaust warning	Linear11	Celsius	1
21-22	I2C mid emergency	Linear11	Celsius	1
23-24	I2C mid warning	Linear11	Celsius	1

3.3.5.5. Thermal status (0x45)

Block Read command 0x45 reports information about the thermal history of the card, including active thermal events, and incrementing counters for these events since the last card powercycle.

Response:

Byte	Name	Format	Unit	API
1-2	Thermal status register	uint16_t	<bit-field>	1
3-4	Temperature warning count	uint16_t	x1	1
5-6	Temperature excess count	uint16_t	x1	1

Thermal status register

Bit	Name	API
0	Thermal shutdown active	1
1	Temperature warning active	1
2	Temperature excess active	1

3.3.6. Host messaging

Commands 0x80 - 0x8F are reserved for the PCIe host system to report information

ID	Name	Bytes (r/w)	Protocol	API
0x80	Driver error state	4	Block Read	1

3.3.6.1. Driver error state (0x80)

Block Read command 0x80 reports the error status for the card as determined by the host driver. This status is non-volatile, and once set requires the card to be RMA’d.

Response:

Byte	Name	Format	Unit	API
1-4	Driver error state	uint32_t		1