The first system partitioning step identifies the system parameters to which the SAR BP engine will be designed. To this end, the following table shows a proposed set of system parameters. These are influenced strongly by the GOTCHA data set [1] proposed for evaluating the system performance.
| Parameter | Value | Units | Notes |
|---|---|---|---|
| Image Width/Height | 512 | pixels | Assume square image |
| # of pulses | 586 | pulses | For 5 azimuth angles of GOTCHA data set |
| Target throughput | 1 | GOPs/sec | Rate of per-pixel back-proj OPs |
| IFFT Transform Size | 2048 | points | Based on system model |
The parameters above drive the overall system performance and cost of the solution. The computational complexity of the algorithm is \(O(N^3)\) for $N\times N$ pixel images. The workload scales linearly with the number of radar pulses to be combined coherently. The IFFT cost varies as $(N\cdot\log(N))$ and can require a large memory footprint.
The “1 GOPs/sec” figure of merit above represents the rate of BP operations on a pixel-by-pixel basis. It’s value represents a placeholder for now. A fundamental open question at this point is what throughput can an AI Engine implementation sustain for the SAR BP algorithm? Some early prototyping is required to answer this question. For now however, it is useful to consider what system performance can be achieved based on this number. We resort to spreadsheets to answer this question.
A key focus of the system partitioning activity is to identify both system performance measures and block-level requirements for the design based on the parameters in the preceding table. The following table computes various system parameters of interest and block requirements based on the parameters above. The values below assume a 1 GOPs/sec throughput to start. This is a somewhat useful and optimistic value as it aligns well to the 1 GHz clock rate of the AI Engine array. To further refine this value, we must engage in some early prototyping to understand what sampling rates are feasible (and with what resource profile) before we can architect a solution.
| Parameter | Value | Units | Notes |
|---|---|---|---|
| Final frame rate | 6.5 | fps | All pulses accumulated |
| Per-Pulse frame rate | 3820 | fps | For a single pulse |
| Image Storage in DDR | 2 | MB | Assume 8B per pixel |
| IFFT Transform Rate | 3820 | Hz | One transform per radar pulse |
| IFFT Sampling Rate | 8 | Msps | Assume streaming solution |
| Total # of AI Engine tiles | TBD | tiles | Require prototyping |