A 0.65-pJ/bit 3.6-TB/s/mm I/O Interface With XTalk Minimizing Affine Signaling for Next-Generation HBM With High Interconnect Density

Hyunjun Park, Jiwon Shin, Hanseok Kim, Jihee Kim, Haengbeom Shin, Taehoon Kim, Hyeri Roh, Jung-Hun Park, Woo-Seok ChoiIEEE Trans. Circuits Syst. I Regul. Pap 2025 · DOI

TL;DR

This paper introduces XMAS (XTalk Minimizing Affine Signaling), a novel I/O interface designed for high-density interconnects like those found in High Bandwidth Memory (HBM). XMAS significantly reduces crosstalk-induced jitter and improves robustness against noise, achieving a remarkable edge density of 3.6 TB/s/mm with an energy efficiency of 0.65 pJ/bit.

What Problem Does This Paper Address?

As computing demands surge, particularly in areas like AI and high-performance computing, the need for faster and more efficient data transfer between processors and memory is paramount. High Bandwidth Memory (HBM) is a key technology for this, utilizing dense interconnects over silicon interposers. However, this density leads to a significant challenge: crosstalk. Crosstalk occurs when signals on adjacent wires interfere with each other, degrading signal integrity and limiting data rates. Furthermore, the sheer number of I/O connections in HBM exacerbates issues like Simultaneous Switching Noise (SSN), which can also corrupt signals. Existing solutions often compromise pin efficiency (the number of data bits transmitted per unit of physical interface width) or introduce significant power and area overhead. This paper tackles the problem of achieving high bandwidth and edge density in these dense interconnects while maintaining signal integrity and energy efficiency, specifically addressing the challenges of crosstalk and SSN.

Methods

The core innovation is the XTalk Minimizing Affine Signaling (XMAS). XMAS employs an affine transformation at the transmitter (TX) and a linear transformation at the receiver (RX) to encode and decode parallel data. This approach is fundamentally different from traditional methods that rely on complex equalization or redundant coding schemes.

Key aspects of the XMAS method include:

  • Affine/Linear Transformation: Parallel input data is transformed by an $n \times m$ integer matrix ($T$) at the TX, and the received signals are transformed by an $m \times n$ integer matrix ($R$) at the RX. This allows for a sophisticated manipulation of signal correlations.
  • Co-optimization of Signaling and Channel: The paper emphasizes the importance of jointly optimizing the XMAS matrices ($T$ and $R$) with the physical channel characteristics (width, spacing, capacitance). An analytical model is developed to facilitate this co-design process, enabling the exploration of a vast parameter space to find optimal configurations.
  • Orthogonal Matrix Design: For binary decision at the receiver, the XMAS matrices $T$ and $R$ are designed such that their product is a diagonal matrix ($RT = \Lambda$). This orthogonality is crucial for minimizing sensitivity to intersymbol interference and ensuring robust binary decisions.
  • Crosstalk Cancellation (XTC) through Matrix Design: The affine and linear transformations are carefully designed to actively cancel out crosstalk. The paper illustrates how specific matrix elements can introduce opposite-signed signals on interfering paths, effectively nullifying the crosstalk at the receiver. This is achieved without adding extra circuitry for equalization or redundant bits for coding.
  • SSN Mitigation: The encoding matrix is constructed such that the set of voltage levels transmitted by the TX drivers remains constant across unit intervals, regardless of the specific input data. This helps to stabilize the current drawn from the power supply, significantly reducing SSN.

Datasets

Not discussed in the paper. The evaluation relies on simulations and measurements from a fabricated prototype.

Key Insights and Ideas

  • Crosstalk as a Design Opportunity: Instead of solely treating crosstalk as a problem to be mitigated with complex circuits, XMAS leverages it as a controllable factor through affine/linear transformations. By carefully designing the encoding/decoding matrices, crosstalk can be actively canceled.
  • The Power of Co-design: The paper strongly advocates for the joint optimization of signaling schemes and physical interconnects. This holistic approach is shown to be critical for maximizing edge density and performance in dense environments.
  • Simplicity with High Performance: XMAS achieves impressive crosstalk cancellation and noise robustness without resorting to complex equalization circuits or bandwidth-inefficient coding schemes. This leads to a more power-efficient and area-efficient solution.
  • Affine Transformation for Signal Integrity: The use of affine transformations at the TX and linear transformations at the RX provides a flexible and powerful mechanism to manipulate signal characteristics and achieve desired outcomes like crosstalk cancellation and noise immunity.

Key Results

  • Edge Density: Achieved an edge density of 3.6 TB/s/mm.
  • Energy Efficiency: Demonstrated an energy efficiency of 0.65 pJ/bit.
  • Crosstalk-Induced Jitter (CIJ) Reduction: Reduced peak-to-peak CIJ by 75% compared to single-ended signaling at 10 GS/s/pin.
  • Eye Opening: Achieved a horizontal eye opening extending to 0.2 UI at a bit error rate (BER) less than $10^{-12}$.
  • SSN Robustness: The XMAS eye diagram remained largely unaffected by SSN, whereas the single-ended signaling eye completely closed under similar conditions.
  • Pin Efficiency: Achieved a high pin efficiency of 87.5%.
  • Fabrication: The prototype transceiver was fabricated in a 28-nm CMOS process.

Limitations and Future Work

  • Orthogonality Constraint: While the orthogonality condition ($RT = \Lambda$) is beneficial for binary decisions, it might limit the flexibility in achieving the absolute best XTC performance in all scenarios. The paper notes that for $n$ greater than 7 or 8, this orthogonality condition becomes unattainable, impacting the maximum symbol rate.
  • Channel Model Dependence: The effectiveness of XMAS is tied to the accuracy of the channel model and the ability to co-optimize the signaling with the channel. Complex or highly variable channel conditions might require further refinement.
  • Future Work: The paper suggests that further exploration of XMAS with different channel characteristics and in more complex multi-drop or multi-cast scenarios could be beneficial. Investigating the trade-offs between pin efficiency, XTC performance, and complexity for even higher data rates would also be a valuable direction.

Summary updated

BibTeX
@article{park2025pj,
  author = {Hyunjun Park and Jiwon Shin and Hanseok Kim and Jihee Kim and Haengbeom Shin and Taehoon Kim and Hyeri Roh and Jung-Hun Park and Woo-Seok Choi},
  title = {{A 0.65-pJ/bit 3.6-TB/s/mm I/O Interface With XTalk Minimizing Affine Signaling for Next-Generation HBM With High Interconnect Density}},
  journal = {IEEE Trans. Circuits Syst. I Regul. Pap.},
  year = {2025},
  doi = {10.1109/tcsi.2025.3573564},
}
Figure