Abstract
Ubiquitous multicore processors nowadays rely on an integrated packet-switched network for cores to exchange and share data. The performance of these intra-chip networks is a key determinant of the processor speed and, at high core counts, becomes an important bottleneck due to scalability issues. To address this, several works propose the use of mm-wave wireless interconnects for intra-chip communication. However, these works often assume idealized channel conditions. This paper demonstrates that such assumptions do not hold in commercial chips and proposes a co-design methodology to engineer the package and adapt the transceiver. Our methods reduce path loss and delay spread by 47 dB and 7.3x, respectively, enabling intra-chip wireless communications over 10 Gb/s.
The Scalability Wall in Manycore Processors
As we move into the "manycore" era with hundreds of cores on a single die, traditional wired Networks-on-Chip (NoCs) are hitting a scalability wall. Multi-hop wired communication creates significant latency and power bottlenecks for chip-wide data exchange. Wireless NoC (WNoC) aims to solve this by overlaying a layer of millimeter-wave (mm-wave) links for low-latency broadcasts and long-range transfers.
However, most WNoC research relies on "ideal" channel models that ignore the physical reality of the chip package. In reality, the metallic and semiconducting layers of a commercial chip introduce massive signal loss and multipath dispersion (echoes). This paper addresses the gap between theoretical WNoC potential and the harsh physical constraints of actual hardware.
The Package as a Designable Component
A central insight of this work is that the chip package is not just a protective shell; it acts like a reverberation chamber or waveguide. By treating the package as a design variable rather than a fixed constraint, we can optimize the environment for wireless propagation.
The authors propose a three-pillar co-design methodology:
- Channel Characterization: Using full-wave electromagnetic simulations, they modeled a realistic flip-chip package to extract path loss and delay spread trends.
- Channel Engineering: Using Simulated Annealing, they optimized silicon thickness ($T_s$), heat spreader thickness ($T_h$), and carrier frequency ($f_c$) to minimize attenuation and dispersion.
- Static Transceiver Optimization: Exploiting the deterministic nature of the on-chip environment, they designed a transceiver using threshold adaptation and Return-to-Zero (RZ) modulation to combat Inter-Symbol Interference (ISI) without complex dynamic equalization.
Key Results and Performance
The results demonstrate that "thinning is winning." Reducing silicon thickness from 0.7 mm to 0.1 mm transitions the environment from a lossy medium to a guided medium, drastically improving signal reach.
Through package engineering, the authors achieved:
- Massive Signal Improvements: Path loss was reduced by up to 47 dB and worst-case delay spread by 7.3x compared to a standard package.
- Breaking the 10 Gb/s Barrier: The optimized system achieved a data rate of 10 Gb/s with a Bit Error Rate (BER) of $10^{-15}$.
- Efficiency: The proposed transceiver design reached these speeds while staying only 3.1 dB away from the ideal, dispersion-free case.
- Path Loss Exponent: Engineering the channel reduced the path loss exponent ($n$) from 4.32 to 1.32, which is better than free-space propagation.
Conclusion
This work provides a reality check for WNoCs, showing that while standard packages are hostile to wireless signals, they can be engineered into efficient communication channels. By leveraging the static, deterministic nature of the intra-chip environment, 10+ Gb/s wireless communication becomes a practical reality for future manycore architectures. Future work will explore higher frequencies in the Terahertz band and the integration of Forward Error Correction to further push efficiency limits.
Summary updated