Break Through the Rack!
Photowave™ optical communications hardware is the first of its kind designed for PCIe and Compute Express Link (CXL) connectivity. Leveraging the significant latency and energy efficiency benefits of photonics, it enables data center managers to scale resources within or across server racks.
Request Documantation
Form Factors
Active Optical Cables
CXL Support
2.0
Latency
<1ns + TOF
Power Consumption
15W
PCIe Generation
5.0
Speed
32Gbps x16
BER
1e-12
Interface
CDFP
Fiber Type
OM3 Multi-mode Fiber
Bifurcation Support
Single x16, Dual x8 and Quad x4
Active Optical Cables
CXL Support
2.0
Latency
<1ns + TOF
Power Consumption
15W
PCIe Generation
5.0
Speed
32Gbps x16
BER
1e-12
Interface
CDFP
Fiber Type
OM3 Multi-mode Fiber
Bifurcation Support
Single x16, Dual x8 and Quad x4
OCP 3.0 SFF Card
CXL Support
2.0
Latency
<20ns + TOF
Power Consumption
30W
PCIe Generation
5.0
Speed
32Gbps x16
BER
1e-12
Interface
OCP 3.0
Fiber Type
OM3 Multi-mode Fiber
Bifurcation Support
Single x16, Dual x8 and Quad x4
OCP 3.0 SFF Card
CXL Support
2.0
Latency
<20ns + TOF
Power Consumption
30W
PCIe Generation
5.0
Speed
32Gbps x16
BER
1e-12
Interface
OCP 3.0
Fiber Type
OM3 Multi-mode Fiber
Bifurcation Support
Single x16, Dual x8 and Quad x4
Low Profile PCle Card
CXL Support
2.0
Latency
<20ns + TOF
Power Consumption
30W
PCIe Generation
5.0
Speed
32Gbps x16
BER
1e-12
Interface
PCIe CEM
Fiber Type
OM3 Multi-mode Fiber
Bifurcation Support
Single x16, Dual x8 and Quad x4
Low Profile PCle Card
CXL Support
2.0
Latency
<20ns + TOF
Power Consumption
30W
PCIe Generation
5.0
Speed
32Gbps x16
BER
1e-12
Interface
PCIe CEM
Fiber Type
OM3 Multi-mode Fiber
Bifurcation Support
Single x16, Dual x8 and Quad x4
Case Study: Memory Expansion for Text Summarization
At Flash Memory Summit 2023 in Santa Clara, California, Lightelligence demonstrated the benefits of memory expansion using CXL to improve workload efficiency for Large Language Model OPT-66B, an AI model used for text summarization. Using two Photowave™ PCIe cards to connect an AMD Genoa server and Samsung Memory Expansion Module, CXL memory expansion over optics demonstrated a 2.4x throughput advantage compared to SSD/NVMe disk offloading with performance remaining constant.
AI Workload:
OPT-66B
GPU:
NVIDIA A10
Server:
2U Supermicro
Memory Expansion:
Samsung CXL 128GB
CPU:
AMD Genoa
Software:
MemVerge MemoryViewer
OPT-66B MODEL PROGRESS 99%
Decode throughput was higher for optical CXL memory, resulting in a 2.4x faster time to
completion.
Performance of the CXL Memory Offload remained constant while Disk / NVMe deteriorated.
Decode throughput was higher for optical CXL memory, resulting in a 2.4x faster time to completion.
Key Technology: Optical Networking (oNET)
Using a CXL optical interconnect, latency is kept low via direct connection to the CPU/GPU and memory.
Copyright © 2022 LIGHTELLIGENCE Co., Ltd.
