Interconnect-Centric AI Computing: From Bandwidth Scaling to Physical Interface Architecture

Lei Deng; Yihong Qi; Jun Fan; James L. Drewniak; Xianbin Wang; Ming Li

doi:10.1016/j.eng.2026.09.004

Engineering ›› :202609004 DOI: 10.1016/j.eng.2026.09.004

research-article

Interconnect-Centric AI Computing: From Bandwidth Scaling to Physical Interface Architecture

Lei Deng ^a^,^b
, Yihong Qi ^b^,^c^,^d^,^f^,^*
, Jun Fan ^c^,^e
, James L. Drewniak ^c
, Xianbin Wang ^d
, Ming Li ^f

Author information +

History +

PDF (2527KB)

Abstract

Cite this article

Download citation ▾

Lei Deng, Yihong Qi, Jun Fan, James L. Drewniak, Xianbin Wang, Ming Li. Interconnect-Centric AI Computing: From Bandwidth Scaling to Physical Interface Architecture. Engineering 202609004 DOI:10.1016/j.eng.2026.09.004

登录浏览全文

4963

注册一个新账户忘记密码

References

Publishing order | Descend order by publishing year | Descend order by cited within

[1]	Keyes RW . Miniaturization of electronics and its limits. IBM J Res Develop 1988; 32(1): 84-8.

[2]	Datta S, Chakraborty W, Radosavljevic M . Toward attojoule switching energy in logic transistors. Science 2022; 378(6621): 733—40.

[3]	Landauer R . Irreversibility and heat generation in the computing process. IBM J Res Develop 1961; 5(3): 183-91.

[4]	Bérut A, Arakelyan A, Petrosyan A, Ciliberto S, Dillenschneider R, Lutz E . Experimental verification of Landauer’s principle linking information and thermodynamics. Nature 2012; 483(7388): 187—9.

[5]	Li M, Vitányi P . An introduction to Kolmogorov complexity and its applications. 4th ed. Cham: Springer; 2019.

[6]	Oikawa S, Nakayama Y, Ito S, Sagawa T, Toyabe S . Experimentally achieving minimal dissipation via thermodynamically optimal transport. Nat Commun 2025; 16: 10424.

[7]	Tirumala A, Wong R . NVIDIA Blackwell platform: advancing generative AI and accelerated computing. In: Proceedings of the IEEE Hot Chips 36 Symposium; 2024 Aug 25—27; Stanford, CA, USA. Piscataway: IEEE; 2014. p. 1-33.

[8]	Chowdhery A, Narang S, Devlin J, Roberts K, Barham P, Chung HW, et al. PaLM: scaling language modeling with Pathways. J Mach Learn Res 2023; 24(240): 11324-436.

[9]

Jiang Z, Lin H, Zhong Y, Wang Y, Wang H, Yang J, et al. MegaScale: scaling large language model training to more than 10,000 GPUs. In: Proceedings of the 21st USENIX Symposium on Networked Systems Design and Implementation; 2024 Apr 15—17; Santa Clara, CA, USA. Santa Clara: USENIX Association; 2024. p. 745—60.

[10]	Optical Internetworking Forum (OIF) . Next generation CEI—224G framework. OIF—FD—CEI—224G—01.0. White paper. Fremont: OIF; 2022.

[11]	Peripheral Component Interconnect Special Interest Group (PCI—SIG) . CopprLink internal cable specification for PCI Express 5.0 and 6.0. Version 1.0. Beaverton: PCI—SIG; 2024.

[12]	De Paulis F, Rabinovich R, Resso M, Kocsis S . Bandwidth limits of connector wipe stub for reliable 224 Gbps signaling. In: Proceedings of the IEEE 27th Workshop on Signal and Power Integrity; 2023 May 7—10; Aveiro, Portugal. Piscataway: IEEE; 2023. p. 1-4.

[13]

Narayanan D, Shoeybi M, Casper J, Puri R, Patwary M, Ali M, et al. Efficient large—scale language model training on GPU clusters using Megatron—LM. In: Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis; 2021 Nov 14—19; St. Louis, MO, USA. New York City: Association for Computing Machinery (ACM); 2021. p. 1-15.

[14]	Jiang C, Tian Y, Jia Z, Zheng S, Wu C, Wang Y . Lancet: accelerating mixture—of—experts training via whole graph computation—communication overlapping. Proc Mach Learn Syst 2024; 6: 74-86.

[15]	NVIDIA. NVIDIA Tesla P100: the most advanced datacenter accelerator ever built, featuring Pascal GP100, the world’s fastest GPU. White paper. Santa Clara: NVIDIA; 2026.

[16]	NVIDIA. NVIDIA Tesla V100 GPU architecture. White paper. Santa Clara: NVIDIA; 2017.

[17]	NVIDIA. NVIDIA A100 Tensor Core GPU architecture. White paper. Santa Clara: NVIDIA; 2020.

[18]	NVIDIA. NVIDIA H100 Tensor Core GPU datasheet [Internet]. Santa Clara: NVIDIA; undated [cited 2026 Jul 15]. Available from: https://resources.nvidia.com/en—us—gpu—resources/h100—datasheet—24306.

[19]	NVIDIA. NVIDIA DGX B200 [Internet]. Santa Clara: NVIDIA; [cited 2026 Jul 15]. Available from: https://www.nvidia.com/en—us/data—center/dgx—b200/.

[20]	Hall SH, Heck HL . Advanced signal integrity for high—speed digital designs. Hoboken: Wiley—IEEE Press; 2009.

[21]	Priyadarshi S. Advancing AI scalability and performance with optical interconnects. IEEE Commun Mag 2025; 63(11): 22-7.

[22]	Netherton A, Dumont M, Nelson Z, Koo J, Jhonsa J, Mo A, et al. High capacity, low power, short reach integrated silicon photonic interconnects. Photon Res 2024; 12(11): A69-86.

[23]	Razdan S, Traverso M, Torza A . Co—packaged optics integration for hyperscale networking. Report. San Jose: Cisco Systems, Inc.; 2026.

[24]	Peripheral Component Interconnect Special Interest Group (PCI—SIG) . PCI Express base specification. Revision 7.0. Beaverton: PCI—SIG; 2025.

[25]	Institute of Electrical and Electronics Engineers (IEEE) . IEEE standard for Ethernet. IEEE Std 802.3—2022. IEEE standard. New York City: IEEE; 2022.

[26]

Institute of Electrical and Electronics Engineers (IEEE) . IEEE standard for Ethernet—Amendment 4: physical layer specifications and management parameters for 100 Gb/s, 200 Gb/s, and 400 Gb/s electrical interfaces based on 100 Gb/s signaling. IEEE Std 802.3ck—2022. IEEE standard. New York City: IEEE; 2022.

[27]

Institute of Electrical and Electronics Engineers (IEEE) . IEEE draft standard for Ethernet—Amendment 13: media access control parameters for 1.6 Tb/s and physical layers and management parameters for 200 Gb/s, 400 Gb/s, 800 Gb/s, and 1.6 Tb/s operation. IEEE P802.3dj/D3.1. IEEE standard. New York City: IEEE; 2026.

[28]	SFF Committee. Specification for GBIC (Gigabit Interface Converter). SFF—8053, Rev. 5.5. Oklahoma City: SFF Committee ; 2000 Sep 27.

[29]	SFF Committee. Specification for SFP (Small Formfactor Pluggable) transceiver. INF—8074i, Rev. 1.0. Oklahoma City: SFF Committee ; 2001 May 12.

[30]	SFF TA TWG . Specification for 0.6 mm 4/8X unshielded I/O connector. SFF—8654, Rev. 1.2. Oklahoma City: SFF Committee; 2018 Apr 27.

[31]	Amphenol Communications Solutions. Mini Cool Edge 0.60 mm connector series [Internet]. Wallingford: Amphenol Communications Solutions ; [cited 2026 Jul 15]. Available from: https://www.amphenol—cs.com/product—series/mini—cool—edge—0—60mm.html.

[32]	rf M. Mister rf. NORTEL AA1419001 [Internet]. San Francisco: Wikimedia Commons; 2022 Oct 31 [cited 2026 Jul 15]. Available from: https://commons.wikimedia.org/wiki/File:NORTEL_AA1419001.jpg.

[33]	Nosachev D. SlimSAS x8 male connector [Internet]. San Francisco: Wikimedia Commons; 2021 May 12 [cited 2026 Jul 15]. Available from: https://commons.wikimedia.org/wiki/File:Slimsas—x8—male—connector.jpg.

[34]	Nosachev D. LR—Link 10GBASE—SR SFP+ transceiver [Internet]. San Francisco: Wikimedia Commons; 2016 Apr 27 [cited 2026 Jul 15]. Available from: https://commons.wikimedia.org/wiki/File:LR—Link_10GBASE—SR_SFP%2B_transceiver.jpg.

[35]	Intel Corporation. Desktop 3rd generation Intel core processor family, desktop Intel pentium processor family, desktop Intel Celeron processor family, and LGA1155 socket: thermal mechanical specifications and design guidelines. Report. Santa Clara: Intel Corporation ; 2013.

[36]	JEDEC Solid State Technology Association . LPDDR5 CAMM2 connector performance standard. JEDEC Standard PS—007A. Arlington: JEDEC; 2024.

[37]	Spekking R. LGA 1155 socket—6510 [Internet]. San Francisco: Wikimedia Commons; 2020 Jan 23 [cited 2026 Jul 15]. Available from: https://commons.wikimedia.org/wiki/File:LGA_1155_socket—6510.jpg.

[38]	streetcar. Amphenol CAMM2 RAM display [Internet]. San Francisco: Wikimedia Commons; 2025 May 23 [cited 2026 Jul 15]. Available from: https://commons.wikimedia.org/wiki/File:Amphenol_CAMM2_RAM_display.jpg.

[39]	Deng L, Tang Y, Lin Q . A novel connector design for PCIe Gen6. In: Proceedings of the IEEE International Workshop on Advanced Interconnects; 2024 Nov 6—8; Ningbo, China. Piscataway: IEEE; 2024.

[40]	Storage Networking Industry Association (SNIA) . Specification for internal unshielded high speed connector system. SFF—TA—1016, Rev. 1.3. Santa Clara: SNIA; 2024 Nov 15.

[41]	Song HI, Lee JY, Won H, Kim CA, Jin H, Eu J, et al. E—TUBE: dielectric waveguide cable for high—speed communication. Sci Rep 2020; 10(1): 18263.

[42]	Marvell Technology . Marvell CEO Matt Murphy keynote at COMPUTEX 2026 [Internet]. San Bruno: Youtube; 2026 Jun 10 [cited 2026 Jul 15]. Available from: https://www.youtube.com/watch?v=DSnugwGCBrQ.

[43]	Gore B, Josephson A, Mellitz R, De Paulis F, Boluna L, Calvin J, et al. Beyond 200G: brick walls of 400G links per lane. In: DesignCon 2025 Compiled Paper Proceedings; 2025 Jan 28—30; Santa Clara, CA, USA. Hong Kong: Wonder Idea Technology Limited; 2025.