Optical server connection via chiplet with 4 Terabit/s
Intel's "Optical Compute Interconnect Chiplet" can be integrated into computing accelerators to network AI servers quickly and efficiently.
Intel's Optical Compute Interconnect chiplet next to a pencil for size comparison.
(Image: Intel)
Intel has been developing optoelectronic semiconductor function blocks for decades. Now the company is showing the prototype of an "Optical Compute Interconnect" (OCI) chiplet that transmits a total of 4 terabits per second (Tbit/s) over eight wavelengths, i.e. 2 Tbit/s per direction. It is intended for integration into chiplet processors and computing accelerators. It is intended to connect several server nodes in a rack or across several racks via optical waveguides.
The OCI chiplet communicates with the computing units of the respective processor via up to 64 lanes of the PCI Express 5.0 generation, i.e. at 64 × 32 Gigabit/s. On the optical side, it uses eight different wavelengths. Intel specifies 5 picojoules (5 pJ) as the energy requirement per bit. According to Intel, this is significantly less than if you were to connect an Ethernet adapter with SFP ports via PCIe and then convert the data in the SFP transceivers from electrical to optical signals.
However, 5 trillionths of a joule per bit at 4 trillion bits per second leads to a power consumption of 20 watts (1 watt = 1 joule/s).
Alternative to NVLink
(Image: Intel)
AI market leader Nvidia networks its fast computing accelerators with its own NVLink, which previously relied on copper cables to save the energy required to convert electrical signals into optical signals. The latest generation NVLink 5 transmits 900 GByte/s per direction via an x18 connection, i.e. 1.8 TByte/s in total – this corresponds to 14.4 Tbit/s.
Nvidia has also been working on "co-packaged photonics" for years, and reported on a concept at GTC China 2020 that should achieve 4 pJ/bit. At the time, Nvidia quoted around 8 pJ/bit for NVLink 2.0; NVLink 5.0 is likely to be far more efficient.
Intel is cooperating with other companies on the Ultra Accelerator Link (UALink), an open interconnect based on Ultra Ethernet.
(Image: Intel)
FPGA cooperation with Ayar
Intel is also cooperating with AyarLabs on the integration of optical networking directly into logic chips. In November 2023, the cooperation partners demonstrated the combination of an Intel FPGA (now more likely Altera FPGA again) and optical transceivers for 4 Tbit/s. There was also talk of less than 5 pJ/bit.
(ciw)
