무슨 일이 일어났나요?
Volantis announced an $88 million Series A funding round led by Lachy Groom and Abstract Ventures, with participation from John Doerr, VXI Capital, Triatomic, Susa Ventures and several angels. The capital will fund the company’s development of vertical‑cavity surface‑emitting laser (VCSEL) technology to replace electrical wiring between dies and high‑bandwidth memory (HBM) in AI accelerators. Volantis claims the optical approach can overcome the limited reach of copper interconnects, potentially allowing a GPU to be surrounded by 220 memory chips instead of the current eight‑chip limit.
Volantis, founded in 2022 and based in San Francisco, closed an $88 million Series A round on Thursday. The round was led by Stripe veteran Lachy Groom and Abstract Ventures, with notable backers John Doerr, VXI Capital, Triatomic, Susa Ventures, and angel investors including AI podcaster Dwarkesh Patel, AI‑chip veteran Naveen Rao and Anthropic researcher Sholto Douglas.
The company’s core technology repurposes vertical‑cavity surface‑emitting lasers (VCSELs), which are already embedded in hundreds of millions of iPhones for facial‑recognition. Volantis plans to use these lasers to transmit data optically between dies and memory stacks, eliminating the electrical reach limitation that caps current GPUs at eight HBM chips.
CEO and co‑founder Tapa Ghosh said the approach could enable a GPU to be surrounded by up to 220 memory chips, dramatically expanding on‑chip memory bandwidth. He emphasized that while advanced packaging is complex, the use of existing VCSEL components could reduce both cost and supply‑chain risk.
Volantis expects to deliver a prototype chip within the next year and a production‑ready device by 2025, though exact pricing, volume commitments, and customer adoption details were not disclosed.
소스 세부정보: economictimes.indiatimes.com ↗
왜 중요한가요?
Current AI accelerators are constrained by the physical distance that electrical signals can travel between and memory, limiting memory bandwidth and scaling. By using VCSELs—laser components already mass‑produced for iPhone facial‑recognition cameras—Volantis aims to sidestep supply‑chain bottleneries tied to high‑cost HBM and copper packaging. If successful, the technology could dramatically increase the amount of model data a single GPU can access, speeding up and training for large language models and other data‑intensive AI workloads. The funding also signals continued venture interest in novel AI‑hardware solutions beyond traditional silicon scaling.
The bandwidth gap between and memory is a primary limiter for scaling large AI models. Optical interconnects could provide orders‑of‑magnitude higher data rates with lower latency, directly impacting the efficiency of training and for models that exceed current memory capacities.
By leveraging a component already mass‑produced for consumer devices, Volantis may avoid the costly, lead‑time‑intensive supply chains that have plagued HBM production, potentially lowering overall system cost for data‑center operators.
If the technology proves viable, it could reshape the competitive landscape of AI hardware, offering an alternative to the copper‑based packaging strategies pursued by Nvidia, AMD and emerging Chinese chipmakers.
대화형 메커니즘: 실제로 작동하는 방식
이 개발의 이면에 있는 기본 기술을 대화식으로 살펴보세요.
Which component of an AI application is the machine-learning model itself?
다음에 무엇을 볼 것인가
Key milestones to monitor include a prototype chip demonstration, the timeline for a first‑generation product (Volantis targets a 2025 release), and any partnership announcements with GPU manufacturers such as Nvidia or AMD. Investors will also watch for supply‑chain validation of VCSEL sourcing at scale and any performance benchmarks that compare optical interconnects to existing HBM solutions.
Prototype demonstration results and any published performance metrics compared with state‑of‑the‑art HBM‑based GPUs.
Partnerships or design‑win announcements with major GPU vendors or cloud providers, which would validate market interest.
Supply‑chain developments confirming that VCSELs can be sourced at the volumes required for data‑center‑scale production.
Regulatory or IP challenges related to repurposing consumer‑grade VCSEL technology for high‑performance computing.