Intel Phi 7120A SC7120A Fiche De Données
Codes de produits
SC7120A
Intel
®
Xeon Phi™ Coprocessor Datasheet
Document ID Number: 328209 003EN
66
Note that the PL1, PL0 default thresholds are intended to be percentages of the TDP,
and the SMC will dynamically determine actual values for the thresholds during
coprocessor boot-up.
and the SMC will dynamically determine actual values for the thresholds during
coprocessor boot-up.
No user intervention is necessary to enable power threshold
throttling.
System administrators may program PL1, PL0 thresholds and their
respective time durations. The Software Development Kit (SDK) packaged in the
coprocessor software stack, or MPSS (
coprocessor software stack, or MPSS (
), contains
documentation on programming the power and time registers. The Out-Of-Band
mechanism is explained in sections 6.6.3.6.1 and 6.6.3.6.2.
mechanism is explained in sections 6.6.3.6.1 and 6.6.3.6.2.
6.6
Out of Band / PCI Express* SMBus / IPMB
Management Capabilities
The Intel
®
Xeon Phi™ coprocessor PCI Express* card exists as part of a system-level
ecosystem. In order for this system to manage its cooling and power demands, the
Intel
Intel
®
Xeon Phi™ coprocessor telemetry must be exposed to ensure that the system is
adequately cooled and that proper power is maintained. Manageability code running
elsewhere in the chassis, through the SMC, can retrieve SMC sensor logs, sensor data,
and vital information required for robust server management. Note that logging, in this
context, is completely separate from and has nothing to do with the MCA error log.
elsewhere in the chassis, through the SMC, can retrieve SMC sensor logs, sensor data,
and vital information required for robust server management. Note that logging, in this
context, is completely separate from and has nothing to do with the MCA error log.
The SMC public interface (SMBus) is a compliant IPMB interface. It supports a minimal
IPMB command set in order to interact with manageability devices such as BMCs and
the Manageability Engine (ME).
IPMB command set in order to interact with manageability devices such as BMCs and
the Manageability Engine (ME).
The IPMB implementation on the SMC can receive additional incoming requests while
responses are being processed. This enables the interleaving of requests and responses
from multiple sources using the SMC’s IPMB, thus minimizing latency.
responses are being processed. This enables the interleaving of requests and responses
from multiple sources using the SMC’s IPMB, thus minimizing latency.
Upon initial power-on or restart, the SMC selects an IPMB slave address from the range
0x30 - 0x4e in increments of 2 (e.g., 0x30, 0x32, 0x34, etc.). The IPMB slave address
self-select starting address is nonvolatile, starting at the last selected slave address.
This ensures that the card doesn’t move nondeterministically in a static system. To
determine the address of the Intel
0x30 - 0x4e in increments of 2 (e.g., 0x30, 0x32, 0x34, etc.). The IPMB slave address
self-select starting address is nonvolatile, starting at the last selected slave address.
This ensures that the card doesn’t move nondeterministically in a static system. To
determine the address of the Intel
®
Xeon Phi™ coprocessor card scan the range of
addresses issuing the Get Device ID command for each address. A valid response
indicates the address used is a valid address.
indicates the address used is a valid address.
For the Intel
®
Xeon Phi™ coprocessor cards, the IPMB slave address will be found at
0x30 if only a single card is installed. If the motherboard has an exclusive connection to
the SMBus on each PCI Express* connection, then the Intel
the SMBus on each PCI Express* connection, then the Intel
®
Xeon Phi™ coprocessor
will assign itself a default address (0x30). If the SMBus connections are shared, each
Intel
Intel
®
Xeon Phi™ coprocessor in a chassis will negotiate with each other and select
addresses in the range from 0x30 to 0x4e. If a mux is incorporated into the design to
isolate devices on a shared link the address negotiation process should result in each
card having address 0x30. However, if the mux in use allows for the channels to be
merged, i.e., creating a shared bus scenario, the address negotiation may result in
each card having a unique address behind the mux.
isolate devices on a shared link the address negotiation process should result in each
card having address 0x30. However, if the mux in use allows for the channels to be
merged, i.e., creating a shared bus scenario, the address negotiation may result in
each card having a unique address behind the mux.
Power management and power control are performed through the host driver interface
(in-band). An SDK is provided as part of the Intel
(in-band). An SDK is provided as part of the Intel
®
Xeon Phi™ coprocessor software
stack and can be found in the standard MPSS release.
The SMC’s PCI Express*/SMBus interface operates as an industry standard IPMB with a
reduced IPMI command implementation. The SMC supports a system event log (SEL)
via the IPMI interface.
reduced IPMI command implementation. The SMC supports a system event log (SEL)
via the IPMI interface.