In the era of 8K ultra-high-definition video, the High Efficiency Video Coding (HEVC/H.265) standard faces significant hardware implementation challenges. This paper, published in IEEE Access, introduces a specialized VLSI architecture designed to meet 8K throughput requirements while minimizing hardware footprint and power consumption through a novel "TU-oriented" approach.
The Challenge: Memory Bottlenecks in Intra Prediction
Intra prediction in HEVC reduces spatial redundancy by predicting pixel values from previously decoded neighboring samples. This process is complex due to the support for 35 different prediction modes and variable Transform Unit (TU) sizes ranging from 4x4 to 32x32 pixels.
Traditional hardware designs often produce a fixed 4x4 output block. However, the subsequent Inverse Discrete Cosine Transform (IDCT) stage—which follows prediction—requires data in shapes matching the TU size. This mismatch necessitates large "ping-pong" SRAM buffers to store intermediate results, which consumes substantial chip area and power. Additionally, fetching the required reference samples (shown above) for each block creates a significant memory bottleneck that hinders 8K performance.
Proposed TU-Oriented Architecture
To address these inefficiencies, the researchers developed a flexible architecture centered on the Transform Unit size rather than a fixed block size. The core of this design is a set of 32 Processing Elements (PEs) that can be dynamically reconfigured.
As illustrated in the figure above, the architecture processes pixels in lines that match the width of the current TU. For a 32x32 TU, it processes a single 32-pixel line; for smaller TUs, it can process multiple lines simultaneously (e.g., four 8-pixel lines for an 8x8 TU). This alignment allows the prediction and IDCT stages to operate in a streamlined pipeline, eliminating the need for massive intermediate buffers.
Key Innovations
- SRAM-Free Reference Buffer: By analyzing the zig-zag scanning order of TUs, the authors replaced traditional SRAM with a compact 0.8K-bit register-based line buffer. This allows for single-cycle data access and significantly reduces power consumption.
- Multiplier Sharing: The mathematical formulas for "Planar" and "Angular" prediction modes were optimized to share hardware multipliers, reducing the overall logic gate count.
- High-Throughput PEs: The 32-PE array ensures a constant throughput of 32 pixels per cycle, regardless of the TU configuration, which is essential for real-time 8K decoding at high clock speeds.
Performance and Results
The proposed architecture was implemented in Verilog HDL and synthesized using a TSMC 65nm CMOS process. Key results include:
- Throughput: Supports 8K resolution at 30fps with a clock frequency of 400MHz.
- Area Efficiency: Requires only 66.2K logic gates, making it significantly smaller than previous 8K-capable designs.
- Memory Savings: The reference buffer uses only 0.8K bits, a massive reduction compared to previous works that used between 4.9K and 21K bits.
Conclusion
By prioritizing data alignment and "shape-aware" processing, this architecture successfully removes the primary memory bottlenecks in HEVC intra prediction. While the industry transitions toward the newer VVC (H.266) standard, the principles of line-based prediction and register-file-based buffers presented here provide a robust foundation for future high-resolution video decoder designs.
Summary updated