What LLM computing Need

1. Since LLM inference is memory BW bounded, cluster supercomputing for AI inference with large scale concurrency is much more efficient than server, due to much less frequent model loading from HBM; .
2. distributed parallel computing with tight coupling(large BW>8Tb/s) is efficient(80%); but inefficient(40%) in loose cluster due to low internode BW.

concurrent inference(deepseek R1) efficiency (tokens per GPU), with >30token/s per user	16 H100	320 H100	320 H100 with NVL
5000token prompt+1000token generation	800	3000	6000 (BW1.5Tb/s)
1000token prompt+1000token generation		5000	15000(BW 4Tb/s)

Training is more BW hungry!!!

Current Solutions

Moore’s law is the bottleneck of the performance of a single chip, thus GPU cluster is the unique future of AI training, concurrent inference and AI4S; Interconnection bandwidth BW is the NASC for the scability

Current Cluster Parallelism

Intranode: NVLINK for 8 GPU 4Tb/s
Internode: infiniband, 400Gb/s
low GPU efficiency

DGX GH200

links 256 GPUs via 4x800G optics for full-speed NVLink, achieving 76% cluster efficiency.
speeds $millions for optical modules；$6000 for each GPU

GB200 nvl72

Custom copper backplane for 8Tb/s full-bandwidth interconnect.
Max 72 GPUs per system. $2k/GPU.
120kW power challenges reliability; cooling (CDU) >$300k.

What We Do

optical IO chip
shorten serdes signal chain to optics in one die, for signal integrity and density

cake	eat	fit
Large BW>4Tb/s	range>100m (optic fiber chain，and WDM for less fiber thus higher reliability	cheap，cost $50, price $300

Technology Innovation

Cutting-Edge Photonic Integration for AI Cluster Interconnection

Material Innovation

disruptive optical material LN/SiN/Si+hybrid bonding with serdes

Architecture Innovation

850nm DWDM，for low cost and high yield

Process Innovation

12inch customerization.

CMOS foundry compatible.

Competitive Advantage Analysis

Technology Feature	TROE Solution	Ayarlabs	TROE merits	current tech
eletrical device node optical material integration of both	22nm electronics LN-SiN-SI photonics wafer bonding	45nm SOI electronics Silocon Photonic monolithic	much faster(both electrical and optical) and lower power	1. optical modules 4*800G, more than $2000 2. copper cables more than $500 and much thermal and power supply cost
wavelength	850nm	1310nm	much cheaper and smaller
WDM	DWDM 3-fiber coupling	LWDM array 30-fiber coupling	much cheaper and higher yield and reliability compatible with optical switch	optical IO is only $500