LightMatter, a US company specializing in photonical-based technologies, revealed this week its new ambitious project for the artificial intelligence market: Passage M1000a high performance optical interconnection platform that promises support bandwidths up to 114 tbps.
Innovation aims to eliminate communication bottlenecks between modular processors chipples, one of today’s greatest technical challenges.
The proposal is simple and bold: replace traditional interconnections with Extremely fast optical connectionsallowing the construction of multi-chip processors with unprecedented performance and scalability.
A new architecture for gigantic and fast chips
O Passage M1000 It is described as a interposition 3D Multi-Reticle with eight-tiles connected – A structure that covers a total area of 4.000 mm²far beyond what is common in current multi-chip processors. Technology enables the use of 1,024 serial data channelseach operating at 56 Gbps with simple modulation.
In addition, the M1000 has 256 FIRE FIBERSeach supporting 8 wavelengths and delivering Up to 448 Gbps per linetotaling the impressive 114 tbps.
The physical package of the chip comes to 7.735 mm² and can provide up to 1,500W of power for the internal components. Something aligned with the energy demands of upcoming generations of artificial intelligence chips.
The Passage M1000 represents a milestone for the semiconductor industry, breaking physical boundaries of traditional interconnections with photonic -based solutions
Nick Harris, Lightmatter Founder and CEO
Passage L200: Connectivity without borders within chips
In addition to the M1000, Lightmatter also announced the Passage L200a 3D optical chiplet designed to completely replace copper interconnections with light -based alternatives. The L200 offers 32 Total Band Tbpswhile the version L200X atinge 64 Tbps. When integrated into a single package, the values can exceed 200 TBPS.
The differential is in the ability to offer “No Border” connectivity – That is, data channels can be positioned anywhere on the chip surface, not just the edges as occurs in conventional architectures. This improves in large scale the internal communication between components.
The L200 uses photonic circuits developed by Lightmatter itself and integra ps the chipples end allowaveadopting the interface UCIe die-to-die and support for 320 SerDes multiprotocolo and WDM multiplexing of 16 wavelengths per fiberreaching up to 1.6 tbps per fiber.
Lightmatter acts as a company fablessthat is, it has no factories. The production of silicon components is made by Globalfroundrossusing the platform Fotonixwhich integrates CMOS and photonic logic in the same chip. Encapsulation services are performed by big names like Amkor e ASE.
O Passage M1000 is expected to reach the market in the Summer of 2025while the L200 should be released in 2026.
Future scenarios: End of bottlenecks and a new interconnection paradigm?
The advance proposed by Lightmatter can represent a Timing point in chips design – Especially those focused on artificial intelligence and high performance computing (HPC).
Today, copper -based interconnections already face severe physical limitationshow latency increase, high energy consumption e heating As chips become larger and more complex.
Companies such as NVIDIAwhich currently dominates the AI accelerator market with the family Hopper (H100, GH200) and interconnectors like the NVLinkface the same challenge: maintain communication between thousands of nuclei and chips with minimal loss of performance. Although NVLINK and NVSwitch offers high bandwidth, they are still electricity -based and therefore consume more and have a limited scale.
Lightmatter, when adopting optical communication with integrated photoniche wants overcome these barrierswith much higher speeds and more controlled thermal consumption. If successful, the company may be Indirect competitor of giants like Nvidia, AMD e Intelnot necessarily in the chips itself, but In the infrastructure that connects these chips – An increasingly vital component in the performance equation.
Comparative: Current interconnections vs. photonic
| Technology | Maximum bandwidth | Sign type | Energy efficiency | Scalability |
|---|---|---|---|---|
| NVLink (NVIDIA) | Up to 900 GB/s per GPU (in clusters) | Electric | High (but with great thermal consumption) | Limited by physical design |
| Infinity Fabric (AMD) | Until ~ 800 Gb/s by socket | Electric | High (with optimizations) | Limited to internal systems |
| Passage M1000 (Lightmatter) | 114 Tbps (14,25 TB/s) | Optic | Very high | Extremely scalable |
| Passage L200X | Atular 64 TBPS Por Chiplet | Optic | Alta | “No Border” connectivity |
Also read:
The future of AI and high performance computing
Although Lightmatter is technically entering a space disputed by companies like Nvidia, it is more likely that, At least at first, she acts as a supplier or technological partner. After all, big industry players have sought alternatives to climb your architectures And they already work with chipple technologies and 3D stacking.
Companies like Nvidia can, in the future, adopt optical third party solutions – or even acquire startups like Lightmatter – if technology is viable in production. The same goes for Intel e AMDwhich already invest in next generation interconnections with standards such as Teaching (Universal Chplet Interconnect Express), which is compatible with what Lightmatter is proposing.
Therefore, with solutions such as the M1000 and the L200, Lightmatter bets on photônica as a means of interconnection exceed performance limits and energy efficiency that have waged advances in modular architectures.
If technology proves viable on a large scale, we may be facing a radical change in the way processors are designed, built and connected.
Fonte: Lightmatter
Source: https://www.adrenaline.com.br/tech/lightmatter-passage-m1000-superchicp-fotonico-ia/
