AI clusters are pushing datacenter optics into a new architectural era. At CIOE 2026 in Shenzhen, Semtech announced a 224G linear TIA and driver portfolio aimed squarely at Near-Packaged Optics (NPO) and Co-Packaged Optics (CPO) in AI/ML clusters and hyperscale data centers.
The new family consists of the GN1838L and GN42T380 linear octal TIAs, plus the GN42M380 linear octal Mach-Zehnder Modulator (MZM) driver. By moving high-speed optical I/O right next to – or even into the same package as – the switch or ASIC, NPO/CPO tackles power, thermal and signal-integrity limits in AI back-end networks.
Render of Semtech GN1838L, GN42T380 and GN42M380 224G optical connectivity chips — Image: Semtech
224G linear optics for 1.6T to 12.8T engines
Semtech’s 224G parts are optimized for CEI-224G-Linear and Open CPX interfaces and are targeted at 1.6T, 3.2T, 6.4T and 12.8T optical engines.
GN1838L and GN42T380 TIAs provide:
- 500µm channel pitch (GN1838L) and 375µm (GN42T380)
- Placement side‑by‑side with the photonics IC (PIC) or directly on top of it
- An Automatic Gain Control (AGC) stage and output driver with programmable CTLE
- Support for both MGC and AGC modes
- Bandwidth tuned for low peaking, low IRN and good group delay with minimal distortion
The GN42M380 MZM driver offers:
- 375µm channel pitch
- Support for SiPho, InP MZM and TFLN modulators
- Low group delay and minimal THD, with programmable CTLE to address ISI
- Flexible biasing for multi‑vendor MZM compatibility
- Configurable output swing and on‑chip equalization for precise electrical/optical tuning
Diagnostics and performance tuning across the TIA and driver family are accessible via device pads and an I²C interface.
Why it matters
According to Dell’Oro Group, NPO and CPO are entering their first real deployment ramp in 2026 as hyperscalers hit power and scaling bottlenecks in AI fabrics. Semtech’s 224G linear portfolio is positioned as one of the enabling building blocks for those optical I/O architectures, giving switch and optics designers a path to higher bandwidth density without resorting to power‑hungry retimed links.
Source: Semtech, Electronics Weekly
