Introduction
In low-latency networking, SmartNIC pipelines, and custom SoC datapaths, inline packet filtering is typically one of the first operations performed on incoming traffic. However, conventional FPGA packet filters often suffer from two major bottlenecks:
- Block RAM (BRAM) Exhaustion & Read Latency: Storing match rules in on-chip SRAM macros consumes scarce memory resources needed for DMA descriptors and packet FIFOs, while introducing fixed read latencies.
-
Traffic Stalls During Rule Updates: Updating filter rules dynamically often forces the pipeline to backpressure (
TREADY = 0) or flush in-flight frames to avoid mid-packet rule tearing.
To solve both problems, I developed gen1_axis-inline-filter (DetermRTL)—a deterministic, zero-BRAM AXI4-Stream inline packet filter IP core—and just published the evaluation SDK, post-P&R PPA benchmarks, datasheet, and a zero-dependency C99/C++11 host HAL driver.
- Evaluation Repository & Datasheet: DetermRTL/gen1_axis-inline-filter-eval
Key Architectural Highlights
1. Zero Block RAM (0 DP16KD) Architecture
Instead of relying on Block RAM primitives, the core uses a pure LUT4 + Flip-Flop register-based architecture across 64-bit, 256-bit, and 512-bit AXI4-Stream datapaths.
- Eliminates BRAM read latency completely.
- Leaves 100% of on-chip BRAM/SRAM blocks free for downstream packet buffers and DMA rings.
2. Dual Pipeline Modes (PIPELINE_MODE)
Depending on whether your design prioritizes minimum wire-to-wire latency or maximum line rate, the core supports two compile-time pipeline modes:
-
PIPELINE_MODE = 0(Ultra-Low Latency): Single-cycle combinational parse and match path for minimum deterministic latency. -
PIPELINE_MODE = 1(High-Throughput): Inserts a registered pipeline stage to break critical comparator paths, achieving 146.37 MHz (9.37 Gbps @ 64-bit) and 102.48 MHz (52.47 Gbps @ 512-bit) post-P&R on a Lattice ECP5-85F (Speed Grade 8) using 100% open-source EDA tools (Yosys + nextpnr-ecp5).
3. Hitless Atomic Rule Updates
Rules are configured over an AXI4-Lite control plane into shadow registers. When firmware asserts the commit bit, the hardware waits for the exact TVALID && TREADY && TLAST frame boundary to swap the active rule set atomically:
- Zero mid-packet rule tearing.
- Zero backpressure stalls or dropped frames during live reconfiguration.
4. Zero-Dependency C99/C++11 HAL & Hardware Self-Discovery
For bare-metal, RTOS, or Linux UIO/mmap host software, the SDK includes a single-header/source HAL (determ_filter.h / example_main.c) with zero dynamic allocation. At initialization, the driver reads a read-only capability register (0x004) to dynamically discover DATA_WIDTH, PIPELINE_MODE, and NUM_RULES directly from the silicon—preventing firmware/bitstream version mismatches.
Open-Source Formal Verification & CI
Beyond directed and randomized Verilator simulations (-Wall -Wextra clean), the RTL is formally verified in CI using open-source formal tools:
-
SymbiYosys (BMC): Formally proves AXI4-Stream protocol invariants (
TVALIDand payload stability under arbitraryTREADYbackpressure stalls), zero packet drops or duplication, and exact atomic rule-commit timing on frame boundaries. - Yosys LEC: Performs Logic Equivalence Checking across synthesis configurations.
Try It Out & Feedback
The full datasheet, post-P&R resource/timing tables, and C99/C++11 HAL examples are available on GitHub:
👉 https://github.com/DetermRTL/gen1_axis-inline-filter-eval
If you’re working on AXI4-Stream datapaths, open-source FPGA flows (Yosys/nextpnr), or formal verification with SymbiYosys, I’d love to hear your thoughts and feedback in the comments!

Top comments (0)