The DynamicXcelerator is a protocol-agnostic connectivity solution,
allowing dynamically reconfigurable low-latency direct paths
between various resource units inside a data center.
How is DynamicXcelerator Different?
- Eliminates the cost of electrical packet switching at scale
- Heterogeneous fabric with homogeneous virtual pods
- Shapes and reconfigures the topology to match workloads
- Eliminates network hops, and silicon latency
- Eliminates the need for dynamic protocols (DCBx, 802.3ad)
- Removes congestion issue totally exceeding PCIe lane bandwidth
- Reduces high power and cooling costs
- Dedicated point-to-point paths through dynamic optical fabric
- RDMC (RDMA enabled Multicast) to vPOD mapping
Dynamicxcelerator Fabric Building Block Approach
Compute Block
Off the Shelf Servers (OTSS) of your choice
Resource Block
Photonic Resource Unit (PRU)
GPUs, NVME, FPGAs, DPUs, SmartNICs, etc.
Photonic Block
All photonic switch fabric scalable to 10s of thousands of rate agnostic ports
Software Block
Drut Fabric Manager (DFM)
Drut Software Platform (DSP)
Application Integration
System Block
Backend management servers, system application servers, and OoBnetwork switches
DynamicXcelerator vPOD Architecture
Each physical DynamicXceleratorpod can be composed into virtual multi-pods or vPODs
vPODsconfigurations are callable either as “# of GPU nodes” or “Data/Model (Tensor/Pipeline)” parallelism requirements
- Example: 32 GPU configuration could be created as: 4x4x2=32, 2x2x8=32, 2x4x4=32, etc., depending on parallelism needs.
Three types of fabric communication paths (could be combined into any combination)
- Type 1: CPU to PRUs (GPUs/FPGAs/NVME, etc.) (GA)
- Type 2: GPU to GPU via RDMA over Photonics (Mid-2024)
- Type 3: CPU to CPU via RDMA over Photonics (1H 2025)
Open Software – Applications (model/data) Defined Topologies
Open Hardware – Vendor Agnostic GPU and RDMA NIC
Photonic Fabric – Low Latency, Rate Agnostic