Virtualizing external memory as local to a machine learning accelerator
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for virtualizing external memory as local to a machine learning accelerator. One ambient computing system comprises: an ambient machine learning engine; a low-power CPU; and an SRAM that is shared among at least the ambient machine learning engine and the low-power CPU; wherein the ambient machine learning engine comprises virtual address logic to translate from virtual addresses generated by the ambient machine learning engine to physical addresses within the SRAM.
1. A device comprising:
a main memory shared by multiple client devices, wherein the multiple client devices include an ambient computing device, wherein the ambient computing device comprises: multiple ambient processing devices including an ambient machine learning (ML) engine; and a shared local memory that is shared by the multiple ambient processing devices,
wherein the ambient computing device includes virtual address logic configured to translate virtual addresses used by the ambient ML engine into physical addresses of the shared local memory, the ambient computing device being configured to:
load first instructions for a first ambient processing device of the multiple ambient processing devices into the shared local memory;
execute the first instructions using the first ambient processing device;
load second instructions for the ambient ML engine into the shared local memory, the second instructions overwriting at least part of the first instructions;
execute the second instructions using the ambient ML engine, to generate a request to read or write data to a virtual address;
translate the virtual address to a physical address of the shared local memory using the virtual address logic; and
read or write the data to the physical address of the shared local memory.
2. The device of claim 1 , wherein the ambient computing device is configured to process sensor signals before other client devices sharing the main memory are activated from a low-power state.
3. The device of claim 1 , wherein the multiple client devices include a main ML engine that generates physical addresses in the main memory.
4. The device of claim 1 , wherein upon receiving an interrupt, the device is configured to stream model parameters from the main memory into the shared local memory.
5. The device of claim 4 , wherein streaming the model parameters into the shared local memory overwrites space in the shared local memory used by one or more other ambient processing devices.
6. The device of claim 4 , wherein streaming the model parameters into the shared local memory comprises streaming the model parameters into available space in the shared local memory.
7. The device of claim 4 , wherein the interrupt represents receipt of one or more sensor signals to be processed.
8. The device of claim 4 , wherein the ambient ML engine is configured to perform an inference pass of a machine learning model by using the virtual address logic to access the model parameters that were streamed into the shared local memory.
9. The device of claim 1 , wherein the shared local memory comprises multiple memory banks configured to be individually powered down when entering a low-power state.
10. The device of claim 1 , wherein the multiple ambient processing devices comprise at least one of a direct memory access controller, one or more other ML engines, or one or more processors.
11. The device of claim 1 , wherein the ambient ML engine comprises a single ML compute tile.
12. A system comprising:
multiple client devices including an ambient computing device; and
a main memory shared by the multiple client devices, wherein the ambient computing device comprises multiple ambient processing devices and a shared local memory that is shared by the multiple ambient processing devices including an ambient machine learning (ML) engine,
wherein the ambient computing device includes virtual address logic configured to translate virtual addresses used by the ambient ML engine into physical addresses of the shared local memory, the ambient computing device being configured to:
load first instructions for a first ambient processing device of the multiple ambient processing devices into the shared local memory;
execute the first instructions using the first ambient processing device;
load second instructions for the ambient ML engine into the shared local memory, the second instructions overwriting at least part of the first instructions;
execute the second instructions using the ambient ML engine, to generate a request to read or write data to a virtual address;
translate the virtual address to a physical address of the shared local memory using the virtual address logic; and
read or write the data to the physical address of the shared local memory.
13. The system of claim 12 , wherein the ambient computing device is configured to process sensor signals before other client devices sharing the main memory are activated from a low-power state.
14. The system of claim 12 , wherein the multiple client devices include a main ML engine that generates physical addresses in the main memory.
15. The system of claim 12 , wherein upon receiving an interrupt, the device is configured to stream model parameters from the main memory into the shared local memory.
16. The system of claim 15 , wherein streaming the model parameters into the shared local memory overwrites space in the shared local memory used by one or more other ambient processing devices.
17. The system of claim 15 , wherein streaming the model parameters into the shared local memory comprises streaming the model parameters into available space in the shared local memory.
18. The system of claim 15 , wherein the interrupt represents receipt of one or more sensor signals to be processed.
19. The system of claim 15 , wherein the ambient ML engine is configured to perform an inference pass of a machine learning model by using the virtual address logic to access the model parameters that were streamed into the shared local memory.
20. The system of claim 12 , wherein the shared local memory comprises multiple memory banks configured to be individually powered down when entering a low-power state.