NVFP4-Model

A hardware implementation of the NVFP4 (E2M1) floating point format in SystemVerilog, synthesized in Intel Quartus, with a Python reference model for verification.

Overview

This project implements:

1. Quantization of BF16 vectors to NVFP4 format with a shared group exponent
2. Dequantization of NVFP4 back to BF16
3. Dot product of two NVFP4 vectors
4. Quantization error analysis datapath
5. A Python reference model that mirrors the hardware behavior

Format

• NVFP4 (E2M1): 1 sign bit, 2 exponent bits, 1 mantissa bit
• Group size: 16 elements per shared exponent
• Bias: 2, Max exponent: 3
• Input: BF16 (16-bit)
• Output: 4-bit compressed elements + 8-bit shared exponent

Files

Synthesis

• conv_bf16tonvfp4.sv — Converts BF16 vector to NVFP4
• conv_nvfp4tobf16.sv — Dequantizes NVFP4 vector back to BF16
• quantization_error_datapath.sv — Chains quantization and dequantization to measure error
• dot_fp.sv — Computes dot product of two NVFP4 vectors
• nvfp4_datapath_top.sv — Top level datapath
• unsigned_max.sv, fp_rnd_rne.sv, clz_int.sv, vec_mul_fp.sv, vec_sum_int.sv, mul_fp.sv, mul_int.sv — Submodules

Testbenches

• tb_conv_bf16tonvfp4.sv — Verifies BF16 to NVFP4 conversion
• tb_conv_nvfp4tobf16.sv — Verifies NVFP4 to BF16 dequantization
• tb_quantization_error_datapath.sv — Verifies quantization error datapath
• tb_dot_fp.sv — Verifies dot product computation
• tb_nvfp4_datapath_top.sv — Top level simulation

Python

• BF16_to_NVFP4.py — Converts BF16 vector to NVFP4
• NVFP4_to_BF16.py — Dequantizes NVFP4 back to BF16
• quantization_error.py — Computes average quantization error

Simulation Results

• All 1.0 inputs → NVFP4 = 0110, shared_exp = 125 ✓
• All 1.5 inputs → NVFP4 = 0111, shared_exp = 125 ✓
• All 2.0 inputs → NVFP4 = 0110, shared_exp = 126 ✓
• Dot product of 16×0110 · 16×0110 = 1024 ✓
• Dot product of 16×0110 · 16×0111 = 1536 ✓
• Round-trip 1.0 → NVFP4 → BF16 = 1.0, error = 0 ✓
• Round-trip 0.125 in mixed group → dequantizes to 0.15625, nonzero error ✓

References

Based on the MX (Microscaling) open source repo, parameterized for NVFP4.