Systems and methods for updating memory side caches in a multi-GPU configuration
Systems and methods for updating remote memory side caches in a multi-GPU configuration are disclosed herein. In one embodiment, a graphics processor for a multi-tile architecture includes a first graphics processing unit (GPU) having a first memory, a first memory side cache memory, a first communication fabric, and a first memory management unit (MMU). The graphics processor includes a second graphics processing unit (GPU) having a second memory, a second memory side cache memory, a second memory management unit (MMU), and a second communication fabric that is communicatively coupled to the first communication fabric. The first MMU is configured to control memory requests for the first memory, to update content in the first memory, to update content in the first memory side cache memory, and to determine whether to update the content in the second memory side cache memory.
1 . A heterogenous processing system, comprising:
a first integrated circuit comprising a general-purpose processor including a plurality of processor cores and a first cache, wherein the general-purpose processor is associated with a first memory;
a coherent chip-to-chip interconnect coupled with the first cache;
a second integrated circuit coupled with the first integrated circuit via the coherent chip-to-chip interconnect, the second integrated circuit including a first graphics processor having a second memory, wherein the coherent chip-to-chip interconnect is configured to enable coherent access to the first memory by the first graphics processor; and
memory management circuitry coupled with the coherent chip-to-chip interconnect, the memory management circuitry to enable the general-purpose processor to cache data from the second memory in the first cache at cache line granularity, wherein the general-purpose processor and the first graphics processor have unified virtual memory.
2 . The heterogenous processing system of claim 1 , wherein a first graphics processor is associated with a second cache.
3 . The heterogenous processing system of claim 2 , wherein the coherent chip-to-chip interconnect is operable to enable coherency between the first cache and the second cache.
4 . The heterogenous processing system of claim 1 , wherein the unified virtual memory includes an entire virtual address space of the first graphics processor and is accessible to the general-purpose processor.
5 . The heterogenous processing system of claim 4 , wherein the unified virtual memory includes the entire virtual address space of the general-purpose processor and is accessible to the first graphics processor.
6 . The heterogenous processing system of claim 1 , comprising a second graphics processor having a third memory, the second graphics processor coupled with the first graphics processor via a point-to-point interconnect.
7 . The heterogenous processing system of claim 6 , further comprising a third integrated circuit including the second graphics processor.
8 . The heterogenous processing system of claim 6 , wherein the first graphics processor is configured to perform a first atomic operation to the third memory via the point-to-point interconnect.
9 . The heterogenous processing system of claim 8 , wherein the general-purpose processor is configured to perform a second atomic operation to the third memory via the coherent chip-to-chip interconnect and the point-to-point interconnect.
10 . An apparatus comprising:
an interface to a coherent chip-to-chip interconnect;
a first graphics processor having a second memory and associated with a second cache, wherein the coherent chip-to-chip interconnect is configured to enable coherent access to a first memory by the first graphics processor, the first memory is associated with a general-purpose processor having a plurality of processor cores, the general-purpose processor is coupled with the first graphics processor via the coherent chip-to-chip interconnect, and the general-purpose processor has a first cache; and
memory management circuitry coupled with first graphics processor and the interface to the coherent chip-to-chip interconnect, the memory management circuitry operable to maintain a unified virtual memory for the first graphics processor and the general-purpose processor, wherein the coherent chip-to-chip interconnect is operable to enable coherency between the first cache and the second cache for the unified virtual memory and the memory management circuitry is configured to enable the general-purpose processor to cache data from the second memory in the first cache at cache line granularity.
11 . The apparatus of claim 10 , comprising a second graphics processor having a third memory, the second graphics processor coupled with the first graphics processor via a point-to-point interconnect.
12 . The apparatus of claim 11 , wherein the first graphics processor is configured to perform an atomic operation to the third memory via the point-to-point interconnect.
13 . The apparatus of claim 11 , wherein the memory management circuitry is configured to maintain a unified virtual memory for the general-purpose processor, first graphics processor, and the second graphics processor.
14 . A multi-chip module comprising:
a first packaged integrated circuit comprising a general-purpose processor including a plurality of processor cores and a first cache, wherein the general-purpose processor is associated with a first memory;
a coherent chip-to-chip interconnect coupled with the general-purpose processor;
a second packaged integrated circuit coupled with the first packaged integrated circuit via the coherent chip-to-chip interconnect, the second packaged integrated circuit including a first graphics processor having a second memory, wherein the coherent chip-to-chip interconnect is configured to enable coherent access to the first memory by the first graphics processor; and
memory management circuitry coupled with the coherent chip-to-chip interconnect, the memory management circuitry to enable the general-purpose processor to cache data from the second memory in the first cache at cache line granularity, wherein the general-purpose processor and the first graphics processor have a unified virtual address space.
15 . The multi-chip module of claim 14 , wherein a first graphics processor is associated with a second cache.
16 . The multi-chip module of claim 15 , wherein the coherent chip-to-chip interconnect is operable to enable coherency between the first cache and the second cache.
17 . The multi-chip module of claim 14 , wherein the general-purpose processor and the first graphics processor have a unified virtual address space.
18 . The multi-chip module of claim 17 , wherein the unified virtual address space includes all virtual memory of the first graphics processor and is accessible to the general-purpose processor.
19 . The multi-chip module of claim 17 , wherein the unified virtual address space includes all virtual memory of the general-purpose processor and is accessible to the first graphics processor.
20 . The multi-chip module of claim 14 , wherein the first graphics processor includes a point-to-point interconnect that is configurable to couple the first graphics processor with a second graphics processor, the second graphics processor having a third memory.
21 . The multi-chip module of claim 20 , wherein the second graphics processor is included in a third packaged integrated circuit that is external to the multi-chip module.
22 . The multi-chip module of claim 20 , wherein the first graphics processor is configured to perform a first atomic operation to the third memory via the point-to-point interconnect.
23 . The multi-chip module of claim 22 , wherein the general-purpose processor is configured to perform a second atomic operation to the third memory via the coherent chip-to-chip interconnect and the point-to-point interconnect.
24 . A method for accessing memory in a heterogeneous processing system, the method comprising:
providing a first integrated circuit comprising a general-purpose processor including a plurality of processor cores and a first cache, wherein the general-purpose processor is associated with a first memory;
coupling a coherent chip-to-chip interconnect with the first cache;
coupling a second integrated circuit with the first integrated circuit via the coherent chip-to-chip interconnect, the second integrated circuit including a first graphics processor having a second memory;
configuring the coherent chip-to-chip interconnect to enable coherent access to the first memory by the first graphics processor; and
enabling the general-purpose processor to cache data from the second memory in the first cache at cache line granularity using memory management circuitry coupled with the coherent chip-to-chip interconnect, wherein the general-purpose processor and the first graphics processor have unified virtual memory.
25 . The method of claim 24 , comprising caching, by the first graphics processor, data from the first memory to a second cache that is associated with the first graphics processor.
26 . The method of claim 25 , comprising caching, by the first graphics processor, the data from the first memory to the second cache at cache line granularity.
27 . The method of claim 26 , comprising enabling coherency between the first cache and the second cache at least in part via the coherent chip-to-chip interconnect.
28 . The method of claim 27 , comprising enabling coherency between the first cache and the second cache additionally via the memory management circuitry.
29 . The method of claim 28 , comprising synchronizing, via the memory management circuitry, a virtual address space of the first graphics processor with the virtual address space of the general-purpose processor.
30 . The method of claim 28 , comprising providing a third integrated circuit including a second graphics processor and coupling the second graphics processor with the first graphics processor via a point-to-point interconnect, the second graphics processor having a third memory.
31 . The method of claim 30 , comprising performing, by the first graphics processor via the point-to-point interconnect, a first atomic operation to the third memory.
32 . The method of claim 31 , comprising performing, by the general-purpose processor via the coherent chip-to-chip interconnect and the point-to-point interconnect, a second atomic operation to the third memory.