Credo expands from DSP into optical chips with 1.6T full-DSP module

The push follows Credo's reading of the market. At a September product launch, Yang Xuexian, Credo's vice president of sales in China, said the performance growth of a single GPU can no longer keep up with AI demand, and the growing compute gap must be filled by connecting more GPUs. As the industry focus shifts from how fast one GPU is to how more GPUs can be connected into a larger computing system, end-to-end integration of optical modules has become a concrete expression of Credo's system-level interconnect strategy.

Credo's optical effort builds on its DSP business. Over eight years, millions of Credo DSPs have entered large data centers, supported by seven generations of architecture and long-term shipments across 50G/lane, 100G/lane and 200G/lane products. Chris Collins, Credo's vice president of sales and optical product marketing, said 50G single-channel DSPs remain popular, while the company is moving toward 400G/lane. In an optical module, the DSP acts as a repair station for high-speed signals. When high-speed electrical signals from upstream chips arrive after long transmission and are about to be converted to light, they may be attenuated or distorted; the DSP at the electrical input cleans them up before optical components process them. In the reverse direction, electrical signals recovered from optical signals pass through the DSP before returning to the electrical chip.

At 800G and 1.6T bandwidths, optical module architectures have diverged under power pressure. LRO modules keep only the transmit-side DSP, while LPO modules remove all DSPs from the module. As speeds and architectures fragment, a few DSP specifications cannot cover every scenario. At 200G/lane, an 8x200G DSP fits a 1.6T full-DSP module, but in a 1.6T LRO its receive-side capability becomes a clear power burden; in a 4x200G 800G module, the cost is too high. Credo's latest full-DSP module, Cardinal 802, uses a 4x200G design rather than the industry's common 8-channel single-die approach, covering both 800G and 1.6T full-DSP scenarios. As one of the earliest DSP vendors to bet on LRO, Credo also introduced a DSP designed for transmit-side processing needs. Collins explained that at 200G/lane, the signal itself is already difficult to transmit, and optical component damage and distortion from long-distance channels require equalization and compensation. If the DSP is removed entirely as in LPO, these problems must be borne by the entire link, leaving little margin for system error.

The DSP is also taking on diagnostic work. FEC statistics, multipath interference and bit error rate data can be continuously collected by the DSP and sent to Credo's PILOT diagnostic platform for analysis. Even with added functions, the DSP remains a chip that must trade off power, area and performance. After noticing tightening 5nm supply two years ago, Credo began discussing whether some products could move back to 7nm while maintaining power and performance. The 100G/lane Robin, launched alongside Cardinal this year, is a result of that approach. Compared with the previous 100G/lane Lark, which used 5nm, Robin does not use 5nm but integrates the Driver, which amplifies the DSP's high-speed electrical output, directly into the DSP, reducing package board area by nearly half. According to Collins, a single Lark DSP consumed about 5.5W, and with an external Driver the total reached at least about 6.5W; after integration in Robin, overall power remains in a similar range.

Credo is also extending from DSP into PICs. The key piece came from this year's acquisition of DustPhotonics, a company that has long developed PIC and laser integration technology for high-speed optical interconnect in data centers. The acquisition gives Credo two silicon photonics options: it can integrate the laser directly onto the PIC, placing the light source closer to the silicon photonics chip, or keep a standard PIC that lets module makers choose their own light source. In either case, light must enter the silicon photonics circuit, and some light is lost in the coupling. Credo's L3C low-loss laser coupling technology is intended to reduce that loss. With one laser, more light can be split across multiple channels. In an 8-channel design, for example, Credo can use two lasers each feeding four channels instead of eight single-channel lasers, reducing laser count, power, cost and heat pressure. The capability also applies to larger optical engines. Collins showed expansion plans for 16-channel and 32-channel PICs and said Credo's target scenarios include traditional pluggable optical modules, near-packaged optics and co-packaged optics. For the first two, where modules or optical engines can still be replaced independently and there is less concern about an embedded laser affecting overall operation, L3C supports more aggressive choices to put the laser inside the optical engine and maintain a low-loss connection between the internal light source and the silicon photonics circuit. L3C also leaves more room for laser selection. Collins said Credo has completed compatibility validation for existing lasers from several suppliers in the United States, China and Japan, giving customers more alternative sources when supply is tight.

The DustPhotonics acquisition means more than adding optical design and manufacturing capability. In Collins's view, optical module makers previously bought DSPs and PICs separately, but two chips placed together may not work directly; interfaces, signals, power and firmware all need joint tuning, and problems require troubleshooting across different suppliers. With both DSP and PIC, Credo can complete some of this adaptation and validation in advance. Originally separate chips can thus be integrated into an electrical-optical combination that has already been tested together. Together with its existing SerDes IP and active electrical cable capabilities, Credo is aiming at longer AI data center interconnect links beyond optical modules.