Two-dimensional processing array with a vertically stacked memory tile array
The present disclosure discloses a processor. The processor is used to perform parallel computation and includes a logic die and a memory die. The logic die includes a plurality of processor cores and a plurality of networks on chip, wherein each processor core is programmable. The plurality of networks on chip are correspondingly connected to the plurality of processor cores, so that the plurality of processor cores form a two-dimensional mesh network. The memory die and the processor core are stacked vertically, wherein the memory die includes a plurality of memory tiles, and when the processor performs the parallel computation, the plurality of memory tiles do not have cache coherency; wherein, the plurality of memory tiles correspond to the plurality of processor cores in a one-to-one or one-to-many manner.
1. A multi-core processor, used in performing parallel computation, and the multi-core processor comprising:
a logic die, comprising:
a plurality of processor cores, wherein each processor core is programmable; and
a plurality of networks on chip, wherein the plurality of networks on chip correspondingly connected to the plurality of processor cores, so that the plurality of processor cores form a two-dimensional mesh network; and
a memory die, vertically stacked with the processor core, the memory die comprising:
a plurality of memory tiles, wherein when the multi-core processor performs the parallel computation, the plurality of memory tiles do not have cache coherency,
wherein, the plurality of memory tiles correspond to the plurality of processor cores in a one-to-one or one-to-many manner,
wherein, each processor core comprises:
a computation logic module, coupled to the corresponding network on chip; and
a memory interface module electrically coupled to corresponding the computation module and used in accessing the corresponding memory tile of the plurality of memory tiles without passing through the networks on chip, so that the computation logic module accesses the corresponding memory tile through the corresponding memory interface module without passing through the networks on chip, and wherein each of the computation logic module is allowed to access the corresponding memory tile and is unallowed to access the memory tiles other than the corresponding memory tile.
2. The multi-core processor of claim 1 , wherein the plurality of memory tiles are respectively connected to one processor core of the plurality of processor cores, so that the plurality of memory tiles correspond to the plurality of processor cores in a one-to-one manner.
3. The multi-core processor of claim 1 , wherein the plurality of memory tiles are respectively connected to a portion of processor cores in the plurality of processor cores, so that the plurality of memory tiles correspond to the plurality of processor cores in a one-to-many manner.
4. The multi-core processor of claim 1 , wherein the memory interface module comprises a physical layer for coupling to the corresponding memory tile and a memory controller, so that each processor core can access the corresponding memory tile.
5. The multi-core processor of claim 4 , wherein the computation logic module comprises:
a central processing unit cluster, arranged to perform various types of parallel control and perform computation; and
a matrix operator, coupled between the central processing unit cluster and the memory interface module, and configured to perform matrix computation on a large multidimensional matrix entering the processor core.
6. The multi-core processor of claim 5 , wherein the computation logic module further comprises:
a vector processor, coupled between the central processing unit cluster and the memory interface module, and configured to perform vector computation on vector inputted in the processor core.
7. The multi-core processor of claim 6 , wherein the computation logic module further comprises:
a data manager, coupled between the corresponding network on chip, the central processing unit cluster and the memory interface module, and arranged to selectively distribute work items or deliver data to the central processing unit cluster, the matrix operator, or the vector processor, depending on the data format.
8. The multi-core processor of claim 7 , wherein each processor core further comprises:
a memory on logic die, for use as a cache or a scratch memory of the processor core, and when the multi-core processor performs the parallel computation, the plurality of memories on logic die of the plurality of processor cores do not have cache coherency.
9. The multi-core processor of claim 8 , wherein the memory on logic die is implemented by an SRAM.
10. The multi-core processor of claim 9 , wherein the plurality of memories on logic die are coupled between the data manager, the vector processor, the matrix operator and the memory interface module.
11. The multi-core processor of claim 10 , wherein each processor core further comprises an application-driven accelerator, coupled between the central processing unit cluster, the data manager and the memory on logic die.
12. The multi-core processor of claim 1 , wherein the plurality of memory tiles are implemented by a DRAM.
13. The multi-core processor of claim 1 , wherein the plurality of memory tiles are implemented by an SRAM.
14. The multi-core processor of claim 1 , wherein the logic die is connected to the memory die in a hybrid bonding manner.