Altering performance of computational units heterogeneously according to performance sensitivity

• US 8,447,994 B2
• Filed: 07/24/2009
• Issued: 05/21/2013
• Est. Priority Date: 07/24/2009
• Status: Active Grant

• Current Assignee: Advanced Micro Devices, Inc.
• Original Assignee: Advanced Micro Devices, Inc.
• Inventors: Nussbaum, Sebastien, Branover, Alexander, Kalamatianos, John
• Primary Examiner(s): Elamin, M
• Application Number:
  US12/508,929
• Publication Number:
  US 20110022833A1
• Time in Patent Office: 1,397 Days
• Field of Search: 713/300, 713/310, 713320-324, 713/330, 713/340
• US Class Current: 713/300
• CPC Class Codes: G06F 1/324 by lowering clock frequency
• CPC Class Codes: Y02D 10/00 Energy efficient computing, e.g. low power processors, power management or thermal management

Abstract

One or more computational units of a computer system are selectively altered in terms of performance according to which of the one or more computational units has a higher performance sensitivity than others of the computational units.

Claims

1. A method for operating a computer system including a plurality of computational units comprising:

removing one or more of the computational units with low performance sensitivity from a group of the computational units until a predicted power margin of computational units remaining in the group is greater than zero; and

boosting performance of the computational units remaining in the group.

2. The method as recited in claim 1 further comprising comparing the respective performance sensitivities of each of the computational units to a threshold value to determine the one or more units with low performance sensitivity.

3. The method as recited in claim 1 wherein the computational units whose performance is boosted are in a same power state prior to being boosted.

4. The method as recited in claim 3 wherein the same power state is a nominal maximum power state.

5. The method as recited in claim 1 wherein the computational units include a plurality of processing cores and the method further comprises:

removing one or more of the processing cores with low performance sensitivity from a group of the processing cores until a predicted power margin of processing cores remaining in the group is greater than zero; and

boosting performance of the processing cores remaining in the group.

6. A method for operating a computer system including a plurality of processing cores comprising:

if a predicted power margin resulting from boosting performance of a group of the cores is less than zero,

removing a core with low performance sensitivity from the group to form a smaller group; and

calculating a new predicted power margin and determining if the new predicted power margin is greater than zero if performance of the cores in the smaller group is boosted;

if the new predicted power margin is greater than zero for the cores in the smaller group, boosting the performance of the cores in the smaller group; and

if the new predicted power margin for the current group is still less than zero, removing another core with low boost sensitivity from the smaller group to form another smaller group.

7. The method as recited in claim 6 wherein determining the new predicted power margin according to, a current actual power margin-Σ

(boosted power-current power), where boosted is power of the cores in the smaller group operated at a boosted power level and current power is power of the cores in the smaller group operating at the current power levels and the current actual power margin is a power margin corresponding to a current power consumption of the computational units.

8. The method as recited in claim 6 further comprising calculating the new predicted power margin according to a difference between power consumption of the cores in the smaller group when boosted from current power consumption of the cores in the smaller group.

9. A method for operating a computer system including a plurality of processing cores comprising:

if a predicted power margin resulting from boosting performance of a group of the processing cores is less than zero, eliminating one or more processing cores from the group, according to performance sensitivity of the one or more processing cores being lower than performance sensitivity of other of the processing cores, until the predicted power margin is greater than zero; and

boosting performance of remaining processing cores in the group by increasing at least a frequency of clock signals being supplied to the remaining cores.

10. The method as recited in claim 9 further comprising:

accessing a storage to determine performance sensitivity of the processing cores, the storage storing a performance sensitivity corresponding to respective process contexts executing on each of the processing cores.

11. An apparatus comprising:

a plurality of computational units;

a storage to store performance sensitivity information associated with the computational units; and

a power allocation function configured to boost performance of one or more of the computational units according to which one or more of the computational units has a higher performance sensitivity, based on the performance sensitivity information, than others of the computational units.

12. The apparatus as recited in claim 11 wherein the power allocation function is configured to compare the performance sensitivity of each of the computational units to a threshold value and to boost computational units according to the comparison.

13. The apparatus as recited in claim 11 wherein the boosted one or more cores are in a same power state immediately prior to being boosted.

14. The apparatus as recited in claim 13 wherein the same power state is a nominal maximum power state.

15. The apparatus as recited in claim 11 wherein the power allocation function is further responsive to a predicted power margin not being sufficient to boost all of a group of cores to a boosted performance state, to remove one or more computational units from the group of computational units, and recalculate a new predicted power margin, the removal being determined according to the one or more computational units from the group having respective boost sensitivities lower than boost sensitivities of other ones of the computational units of the group, the removal and recalculation being repeated until the new predicted power margin is greater than zero to thereby accommodate boosting performance to the boosted performance state.

16. The apparatus as recited in claim 11 wherein the apparatus comprises at least one integrated circuit and the computational units comprise at least one of processing cores, a memory controller, and a graphics processing unit.

17. The apparatus as recited in claim 11 wherein the power allocation function is implemented in one or more of hardware, firmware, and software stored on a computer readable medium.

18. A non-transitory computer readable medium encoding a computer readable description of circuits that include, a plurality of computational units;

a storage to store performance sensitivity information associated with the computational units; and

a power allocation function configured to alter performance of one or more of the computational units according to which one or more of the computational units has a higher performance sensitivity, according to the performance sensitivity information, than others of the computational units.