Abstract
This paper presents a floating point fused dot-product (FDP) unit with latency reduced. The proposed FDP unit performs the dot-product operation of four floating point numbers: $ab \pm cd$ and is implemented with dual-path algorithm. The proposed FDP is modeled in Verilog-HDL and synthesized using TSMC 65 nm technology library. Synthesis results show that our proposed FDP unit is 24–30% faster and 36.4% less area than the fastest FDP in previous work. We also use the proposed FDP unit and our previously designed FAS (fused add-subtract) unit to implement a FFT Radix-2 Butterfly (R2BF) unit. The latency of our proposed R2BF unit is improved roughly by 34% and the area is reduced by 41.6%, compared to the fastest 2’s-complement butterfly unit.
Introduction
In high-performance computing and digital signal processing (DSP), floating-point arithmetic is a fundamental but computationally expensive operation. Traditionally, calculating a dot product such as $(a \times b) + (c \times d)$ requires discrete multiplication steps followed by an addition. Each stage introduces potential rounding errors and increases overall latency.
"Fused" units address these issues by combining operations into a single hardware block, thereby reducing both error and delay. However, existing FDP units often suffer from high latency due to the complexities of exponent alignment and handling "near-zero" results where significant cancellation occurs. This research aims to eliminate these bottlenecks by redesigning the internal pipeline and optimizing the critical timing path.
Methodology
The proposed architecture utilizes a dual-path algorithm, which bifurcates the calculation into a "FAR" path (for large exponent differences) and a "CLOSE" path (for small differences where cancellation is likely). To maximize performance, five specific hardware optimizations were implemented:
- Logic Rebalancing: The mantissa swap logic was moved from the second pipeline stage to the first, balancing the delay across stages.
- Hardware Sharing: A specialized
Shift2Bitblock is shared between the FAR and CLOSE paths to minimize area. - Parallel Sticky-Bit Computation: The sticky bit used for rounding is computed in parallel with the main shifter.
- Optimized Injection-Based Rounding: The authors demonstrated that the injection signal for rounding can be reduced from 24 bits to just 3–5 bits, significantly decreasing adder width and delay.
- Dual-Way Rounding: A dual-way structure in the final addition stage handles rounding cases more efficiently.
Evaluation Metrics
To ensure a fair comparison across different manufacturing processes (e.g., 45nm vs 65nm), the authors employed technology-independent metrics: Logic Levels ($L$) and Fan-Out-of-4 (FO4) delays.
As shown in the logic diagrams above, basic gate operations are assigned fractional logic levels (e.g., $0.5L$ for simple AND/XOR gates and $1L$ for more complex combinations). This allows for an architectural comparison that transcends specific CMOS technology nodes.
Results and Discussion
The proposed FDP unit was modeled in Verilog-HDL and synthesized using a TSMC 65 nm CMOS technology library. It demonstrated significant improvements over the previous state-of-the-art:
- Speed: The design achieved a 23.7% speedup in terms of logic levels (30.5L vs 40L) and approximately 29.8% in FO4 delay.
- Area: A 36.4% reduction in silicon area was achieved, primarily through the reduction of internal shifter and comparator widths.
- System Integration: When integrated into an FFT Radix-2 Butterfly (R2BF) unit, the new design was 34% faster and 41.6% smaller than the fastest existing 2’s-complement butterfly unit.
While the design excels in latency and area, it recorded a power consumption of 30.34 mW, which is higher than some slower, non-optimized designs. The primary focus of this work was the optimization of the critical path and physical footprint.
Conclusion
By optimizing pipeline balance and streamlining rounding logic, this work provides a highly efficient FDP unit suitable for high-speed DSP applications. Future work may involve applying these dual-way rounding and parallel sticky-bit techniques to more complex arithmetic units, such as three-term dot products or fused multiply-accumulate (FMA) units, while exploring further power reduction strategies.
Summary updated