Hardware architecture for accelerating artificial intelligent processor
View Patent ↗A hardware architecture that may include: a host, a frontal engine, a parietal engine, a renderer engine, an occipital engine, a temporal engine, and a memory. The frontal engine may obtain a 5D tensor from the host and divide it into several groups of tensors. These groups of tensors may be sent or transmitted to the parietal engine, and the parietal engine may take the groups of tensors to further divide them into several tensors. The parietal engine may send these tensors to the renderer engine for execution and may send a partial amount of tensors to the occipital engine. The occipital engine may accumulate the partial amount of tensors and may execute them. The occipital engine may send the output feature as the final tensor to the temporal engine. The temporal engine may compress the final tensor before storing or saving it to the memory.
1. A hardware architecture for accelerating an artificial intelligence processor, comprising:
artificial intelligence task is regarded as a 5D tensor in an application-specific integrated circuit (ASIC), and in each dimension, the task is divided into groups, and each group being further divided into a plurality of waves;
a host,
a frontal engine,
a parietal engine,
a renderer engine,
an occipital engine,
a temporal engine, and
a memory, wherein the memory comprises random access memory (RAM);
wherein the frontal engine obtains a 5D tensor from the host and divides the 5D tensor into a plurality of tensors, wherein the parietal engine sends the plurality of tensors to the renderer engine, wherein the renderer engine is configured to execute an input feature renderer and a partial tensor output to the occipital engine, wherein the occipital engine accumulates a partial tensor and executes an output feature renderer to obtain a final tensor sent to the temporal engine, wherein the temporal engine performs data compression and writes the final tensor into the memory.
2. The hardware architecture according to claim 1 , wherein the parietal engine processes the plurality of tensors according to a user-defined input feature renderer and outputs the partial tensor output to the occipital engine.
3. The hardware architecture according to claim 1 , wherein in a unified rendering architecture, the execution of the output feature renderer is sent back to the parietal engine, and in response to the parietal engine finishes rendering, the result is sent back to the occipital engine.
4. The hardware architecture according to claim 1 , wherein n is the split rendering architecture, wherein the output feature renderer is specifically, wherein the output feature renderer is processed in the occipital engine, the occipital engine sends the output tensor to the temporal engine, the temporal engine performs post processing, and sends it to DRAM or keep it in the cache for further processing.
5. The hardware architecture according to claim 1 , wherein the number of parietal engines is configurable, wherein the frontal engine sends the group tensors to the parietal engine in a polling schedule, and each parietal engine consists of a specific stream-aware processor, all stream-aware processor shares an L2 cache and an export block.
6. The hardware architecture according to claim 5 , wherein The stream perceptron processor comprises an operator core and a neuron kernel as a computational core, wherein the operator core is configured for a general computation, wherein the neuron kernel is configured for an artificial intelligence computation, and the neuron kernel is composed of an L1 buffer and a multiply accumulator.
7. The hardware architecture according to claim 6 , wherein the tensor wave is sent to the neuron kernel, wherein the neuron kernel runs the input feature renderer on the tensor wave, wherein an output from the neuron kernel is exported to the occipital engine, wherein the occipital engine performs accumulation and sends the output back to the parietal engine, wherein the parietal engine is executed on a calculator output feature renderer and output the result to the temporal engine.
8. The hardware architecture according to claim 6 , wherein the number of parietal engines is four, wherein each parietal engine has two stream perceptron processors, wherein each stream perceptron processor has four neurons, and wherein each neuron has eight multiply accumulator sets, wherein each of the eight multiply accumulator sets has 4 multiply accumulators.