WOLF-144L
Overview
HPEC with NVIDIA RTX A4500 and ConnectX-7
The VPX3U-A4500E-CX7 HPEC module includes an NVIDIA RTX A4500 embedded GPU and a ConnectX SmartNIC. The A4500 embedded GPU provides advanced processing capabilities for high performance embedded computing (HPEC) and artificial intelligence (AI) processing. The ConnectX-7 provides the Ethernet and PCIe connectivity needed to move large datasets efficiently.
The NVIDIA Ampere architecture includes CUDA cores for HPEC, 3rd generation Tensor cores for AI and data science computations, and 2nd generation Ampere Ray Tracing (RT) cores for visually accurate rendering. The Ampere architecture introduces support for PCIe Gen 4, improvements to memory handling, and improvements to warp scheduling, which all optimize the way data is distributed for processing. The Ampere GPU is very power efficient, delivering significantly more performance per watt compared to previous generations.
The NVIDIA ConnectX-7 SmartNIC provides PCIe and Ethernet connectivity. ConnectX-7 is ideal for the high-speed, secure, data transfer capabilities required for data-heavy tasks such as sensor data processing and other C5ISR tasks. The ConnectX-7 also provides support for RDMA over Converged Ethernet (RoCE), enabling the fastest method for transferring data across the network to the GPU.
Block Diagram
The block diagram shows the RTX A4500 (Ampere) GPU integrated with a ConnectX-7 SmartNIC, emphasizing the high-throughput Ethernet-to-PCIe path for data-heavy mission workloads. It highlights how the SmartNIC provides network connectivity and routes data efficiently to GPU compute through the module’s PCIe interfaces.
Features
WOLF-144L is a rugged 3U VPX HPEC module that integrates an NVIDIA RTX embedded GPU with an NVIDIA ConnectX-7 SmartNIC to support embedded AI and high performance embedded computing workloads with high-throughput networking. It includes 16 GB GDDR6 256-bit memory with up to 512 GB/s bandwidth, and the ConnectX-7 provides up to 100GbE Ethernet plus a PCIe Gen4 switch to move large datasets efficiently. Module power is configurable from 100W to 140W, and the design supports SOSA aligned slot profiles 14.6.11 or 14.6.13.
GPU Features
- Ampere GPU parallel processing:
- CUDA Toolkit 12, CUDA Compute capability 8.6
- CUDA-X AI and CUDA-X HPEC libraries
- OpenCL 3.0, DirectX 12 Ultimate, OpenGL 4.6, OpenGL ES 3.2, Vulkan 1.2
- 184 Tensor Cores (3rd Gen)
- 46 Ray Tracing cores (2nd Gen)
- NVENC (7th Gen) and NVDEC (5th Gen) with up to 8K video encoding and hardware decoding support
Connectivity / System Management
- On-board IPMI controller for system management
- PCIe configurable switch
- Switching is offloaded from the CPU to the ConnectX with NVIDIA ASAP2 technology
- Daisy Chain options
- Support for 40/100GBASE-KR4 protocols
- 10GBASE-KR Data and Control planes
- GPUDirect RDMA and RoCE support
- Block-level hardware encryption and the use of dedicated encryption keys per user
- Linux and Windows drivers
- NVIDIA Ampere driver support requires the following host CPU: Intel E, S/H/H35 or AMD H/HS Class
Mechanical / Open Systems Architecture
- High level of ruggedization:
- Rugged conduction cooled
- Operating temperature: - 40°C to +70°C standard, operational to +85°C
- Vibration (sine wave): 10G peak, 5 - * Shock: 40G peak
- Dimensions: 160mm x 100mm x 25.4mm
- Weight (approximately): 1.3 kg
- ANSI/VITA 48, 65 (VPX- REDI, OpenVPX)
- SOSA Aligned 14.6.11 or 14.6.13 slot profile
Specifications
Key configuration and performance parameters for WOLF-144L are summarized below for quick comparison. Values shown reflect the standard module configuration; contact WOLF if you need a tailored build or a specific payload profile.
Configuration Guide
The following table defines series of common order codes for the VPX3U-A4500E-CX7 module. The asterisks denote characters of the part number that are defined based on common configuration options.
Some configuration options for this module include:
- Ethernet Configuration Options
- PCIe Configuration Options
- Variant Locked
- Default Power Threshold
- Conformal Coatings
- P2 and P2B depopulated (SOSA aligned)
Contact Sales for the latest Ordering Numbers and available options.
Representative ordering numbers
| Ordering Number | Description |
|---|---|
| 144L33-F***-***VPX3vA0 | 3U VPX, Conduction Cooled, 1”, OpenVPX, NVIDIA Ampere RTX A4500, 16GB GDDR6, PCIe up to Gen4, ConnectX 7, Rear: 4x DisplayPort output, Front Panel 2x 1000Base-T Ethernet |
| 144L33-F***-***VPX3vA0 | 3U VPX, Conduction Cooled, 1”, SOSA 14.6.11 profile with P2 and P2B depopulated, NVIDIA Ampere RTX A4500, 16GB GDDR6, PCIe up to Gen4, ConnectX 7, no video output |
Related Products
- NVIDIA RTX A4500 (GA104)
- 5888 CUDA Cores, 184 Tensor Cores
- 16 GB GDDR6 VRAM
- PCIe Gen 4
- SOSA Payload
- Module power 70–160 W
- NVIDIA RTX A4500 (GA104)
- 5888 CUDA Cores, 184 Tensor Cores
- 16 GB GDDR6 VRAM
- PCIe Gen 4
- SOSA Payload, SOSA Legacy
- Module power 100–190 W
FAQ
Choose WOLF-144L when your platform is constrained by network/fabric data movement and you want SmartNIC offloads integrated with GPU compute.
This is typically driven by:
- High-rate ingest/distribution over Ethernet fabrics
- Use of RoCE and GPUDirect RDMA to reduce CPU copies/overhead
- A desire to offload packet movement and switching tasks from the host CPU
If your workload is primarily local to the chassis and you do not need SmartNIC capabilities, a GPU-only module is usually simpler.
It is intended to improve end-to-end throughput and latency by reducing CPU involvement in moving data over the network into the processing pipeline.
It matters most for:
- Streaming sensor data pipelines
- Multi-node Ethernet fabrics
- Designs where CPU cycles are scarce or latency budgets are tight
Net: it changes how you design buffers, CPU workload, and network-to-processing handoff.
Confirm these early to prevent backplane rework:
- Slot profile alignment: SOSA 14.6.11 or 14.6.13
- Ethernet support: ensure your backplane supports the intended KR plane mapping
- PCIe routing assumptions: confirm the PCIe presentation aligns with chassis expectations
SmartNIC modules are easiest to integrate when the platform was designed with high-speed Ethernet plane mapping in mind.
Key tradeoffs are integration choices:
- P2 vs P2B depopulation (profile choice): drives routing and compatibility
- Ethernet/PCIe configuration: determines lane/port presentation
- Power threshold (100W-140W envelope): must match chassis cooling
- Program controls: conformal coatings and variant locking
Contact Sales for the latest ordering numbers and available configuration options.