BxBChan Design Comparison Deep Dive
BxBChan Design Comparison Deep Dive
The current alternative to purchasing a BxBChan is to build a channelizer yourself. How well is that going to work out?
To answer that question, this web page compares a BxBChan channelizer to a channelizer built by a DSP expert using Altera's DSP Builder software on top of Matlab/Simulink.
Here is the presentation of the DSP expert's solution. (This video is also publicly available on YouTube):
The slides in this video show two different channelizer solutions, with sufficient details to duplicate each solution with a BxBChan. Performance results are also contained in the slides. Extracting these design details gives the following:
| Parameter | Value |
|---|---|
| FPGA | Agilex 7 AGFB014R24A3E3V |
| Channels | 1024 |
| Type | Complex In, Complex Out |
| Parallelism | Processes 8 points per clock (PPC8) |
| Taps | 16 |
| Input Bits | 14 |
| Filter Coef Bits | 12 |
| FFT Data Bits | 16in / 23out |
| Oversampling Ratio | 1:1 (Critically Sampled) |
| Parameter | Value |
|---|---|
| FPGA | Agilex 7 AGFB014R24A3E3V |
| Channels | 1024 |
| Type | Complex In, Complex Out |
| Parallelism | Processes 16 points per clock (PPC16) |
| Taps | 16 |
| Input Bits | 10 |
| Filter Coef Bits | 12 |
| FFT Data Bits | 16in / 23out |
| Oversampling Ratio | 1:1 (Critically Sampled) |
The goal is to build matching BxBChans to compare results. One parameter that shouldn't be precisely matched is the FFT Data Bit setting of 16in/23out. The BxBFFT inside the BxBChan doesn't reduce bits at its beginning stages, since that produces excessive rounding at those stages without a significant resource benefit. So input and output bits of the BxBFFT inside the BxBChan are the same. To provide what should be superior numerical performance, the BxBChan's FFT Data Bits were set to 18in/18out. This improves performance because the 16-bit data at the first stage of the DSP_Builder FFT will have four times the rounding amplitude of an 18-bit BxBFFT input, or 16 times the noise power. There will also be increased errors in early DSP_Builder FFT twiddle multiply results. These early errors in the DSP_Builder Matlab/Simulink FFT can't be recovered by its larger bit widths in later FFT stages. A medium bit width throughout an FFT beats a low-bit-width/high-bit-width strategy.
Another change is that for BxBChan resource testing, the FPGA was changed from AGFB014R24A3E3V to AGFB014R24B3E3E, since the older FPGA is no longer supported by Quartus 25.3. This is believed to be an equivalent FPGA.
All other parameters are easily matched. In this way each of these DSP_Builder solutions can be duplicated with a BxBChan that you can purchase off-the-shelf, with less than 1 week lead time, and significantly less expertise. Below are the results.
| Parameter | DSP_Builder Channelizer | BxBChan Channelizer | Improvement from using a BxBChan |
|---|---|---|---|
| Fmax | Not in Presentation | 748.5MHz | No Data |
| Restricted Fmax | Not in Presentation | 559.91MHz | No Data |
| ALM | 42647 | 32859 | 23.0% |
| DSP | 440 | 216 | 50.9% |
| Mem Blocks | 144 | 134 | 6.9% |
| Parameter | DSP_Builder Channelizer | BxBChan Channelizer | Improvement from using a BxBChan |
|---|---|---|---|
| Fmax | 536.19MHz | 692.52MHz | 29.2% |
| Restricted Fmax | 536.19MHz | 559.91MHz | 4.4% |
| ALM | 73279 | 50271 | 31.4% |
| DSP | 824 | 416 | 49.5% |
| Mem Blocks | Not in Presentation | 171 | No Data |
These measurements obtained from Quartus reports show that the BxBChan provides improvement over a DSP_Builder Matlab/Simulink channelizer in every resource category. This allows larger designs to fit in Altera FPGAs so they can better meet customer needs. Fmax is also higher, allowing faster operation. Restricted Fmax is much higher, showing increased timing margin. The resource savings will inevitably also lead to lower power consumption.
The conclusion can only be that unless there is a specific requirement to use DSP_Builder, Matlab, or Simulink, a BxBChan channelizer should be the preferred choice. It delivers superior results with less expertise, time, and effort.
Since the FFT is a large and fundamental component of the channelizer, these results should extend as well to comparison of the BxBFFT vs the DSP_Builder Matlab/Simulink FFT.