Compute optimization mechanism for deep neural networks
An apparatus to facilitate compute optimization is disclosed. The apparatus includes a memory device including a first integrated circuit (IC) including a plurality of memory channels and a second IC including a plurality of processing units, each coupled to a memory channel in the plurality of memory channels.
1. An apparatus, comprising:
one or more processors including a graphics processor, the graphics processor including:
one or more processing units to provide a first set of shader operations associated with a shader stage of a graphics pipeline,
a scheduler to schedule shader threads for processing, and
a field-programmable gate array (FPGA) dynamically configured to provide a second set of shader operations associated with the shader stage of the graphics pipeline; and
wherein, upon the FPGA being dynamically configured to provide the second set of shader operations, one or more shader threads are scheduled to be processed at either the one or more processing units or the FPGA based on one or more processing characteristics targeted for each shader thread, the FPGA being asynchronously triggered to perform operations associated with the shader stage in the graphics pipeline via access to a memory location associated with the FPGA.
2. The apparatus of claim 1 , wherein the first set of shader operations have a first set of processing characteristics and the second set of shader operations have a second set of processing characteristics.
3. The apparatus of claim 2 , wherein the first set of processing characteristics include a first target speed of operation and a first target power consumption and a first shader thread is to be scheduled to be processed by the one or more processing units according to the first set of processing characteristics, and wherein the second set of processing characteristics include a second speed of operation and a second power consumption and a second shader thread is scheduled to be processed at the second set of shader operations according to the second set of processing characteristics.
4. The apparatus of claim 1 , wherein the first set of shader operations and the second set of shader operations are different versions of a same set of shader operations.
5. The apparatus of claim 1 , wherein the graphics processor further includes a buffer to store data for shader operations; and
wherein the one or more processors are to dispatch a request to the FPGA to perform a shader operation in the second set of shader operations via the buffer, the FPGA is to perform the shader operation, and the FPGA is to return a signal to the buffer upon completion of the shader operation.
6. The apparatus as in claim 1 , wherein the one or more processing units include one or more streaming multiprocessors having a single instruction multiple thread architecture.
7. A method comprising:
receiving a plurality of shader threads at a graphics processor, the graphics processor including one or more processing units to process data and a field-programmable gate array (FPGA), wherein the graphics processor provides a graphics pipeline and the one or more processing units are configured to provide a first set of shader operations having a first set of processing characteristics;
dynamically configuring the FPGA as a shader stage in the graphics pipeline, wherein dynamically configuring the FPGA as the shader stage in the graphics pipeline includes asynchronously triggering the FPGA via access to a memory location associated with the FPGA, the FPGA to provide a second set of shader operations, the second set of shader operations having a second set of processing characteristics; and
scheduling one or more shader threads for processing at the one or more processing units or the FPGA based on processing characteristics targeted for each of the one or more shader threads.
8. The method as in claim 7 , wherein the first set of processing characteristics include a first speed of operation and a first power consumption and a first shader thread is scheduled to be processed by the one or more processing units according to the first set of processing characteristics, and wherein the second set of processing characteristics include a second speed of operation and a second power consumption and a second shader thread is scheduled to be processed at the second set of shader operations according to the second set of processing characteristics.
9. The method as in claim 7 , wherein the first set of shader operations and the second set of shader operations are different versions of a same set of shader operations.
10. The method as in claim 7 , wherein the graphics processor includes a buffer to store data for shader operations and the method further comprises dispatching a request to the FPGA to perform a shader operation in the second set of shader operations via the buffer, the FPGA to perform the shader operation and return a signal to the buffer upon completion of the shader operation.
11. The method as in claim 7 , wherein the one or more processing units include one or more streaming multiprocessors having a single instruction multiple thread architecture.
12. A graphics processor, comprising:
one or more processing units to provide a first set of shader operations associated with a shader stage of a graphics pipeline,
a scheduler to schedule shader threads for processing, and
a field-programmable gate array (FPGA) dynamically configured to provide a second set of shader operations associated with the shader stage of the graphics pipeline; and
wherein, upon the FPGA being dynamically configured to provide the second set of shader operations, one or more shader threads are scheduled to be processed at either the one or more processing units or the FPGA based on one or more processing characteristics targeted for each shader thread, the FPGA being asynchronously triggered to perform operations associated with the shader stage in the graphics pipeline via access to a memory location associated with the FPGA.
13. The graphics processor of claim 12 , wherein the first set of shader operations have a first set of processing characteristics, the second set of shader operations have a second set of processing characteristics, and the first set of shader operations and the second set of shader operations are different versions of a same set of shader operations.
14. The graphics processor of claim 13 , wherein the first set of processing characteristics include a first target speed of operation and a first target power consumption a first shader thread is scheduled to be processed by the one or more processing units according to the first set of processing characteristics and the second set of processing characteristics include a second speed of operation and a second power consumption and a second shader thread is scheduled to be processed at the second set of shader operations according to the second set of processing characteristics.
15. The graphics processor of claim 12 , wherein the graphics processor further includes a buffer to store data for shader operations; and
wherein the one or more processing units are to dispatch a request to the FPGA to perform a shader operation in the second set of shader operations via the buffer, the FPGA is to perform the shader operation, and the FPGA is to return a signal to the buffer upon completion of the shader operation.
16. The graphics processor of claim 12 , wherein the one or more processing units include one or more streaming multiprocessors having a single instruction multiple thread architecture.
17. A data processing system comprising:
a memory device;
one or more processors coupled with the memory device, the one or more processors including a graphics processor, the graphics processor including:
a buffer to store data for shader operations;
one or more processing units to provide a first set of shader operations associated with a shader stage of a graphics pipeline,
a scheduler to schedule shader threads for processing, and
a field-programmable gate array (FPGA) dynamically configured to provide a second set of shader operations associated with the shader stage of the graphics pipeline;
wherein, upon the FPGA being dynamically configured to provide the second set of shader operations, one or more shader threads are scheduled to be processed at either the one or more processing units or the FPGA based on one or more processing characteristics targeted for each shader thread, the FPGA being asynchronously triggered to perform operations associated with the shader stage in the graphics pipeline via access to a memory location associated with the FPGA; and
wherein the one or more processors are to dispatch a request to the FPGA to perform a shader operation in the second set of shader operations via the buffer, the FPGA is to perform the shader operation, and the FPGA is to return a signal to the buffer upon completion of the shader operation.