Five approaches to compute, networking and visualization at the edge
At the edge, GPU performance is only part of the design. Data must reach the processor, results must move where they are needed, and the system must fit within the chassis’ power and cooling limits.
Those constraints shape module selection. A platform processing multiple sensor streams may need fast network access. A display-heavy system may require direct video output. A chassis with limited slots may benefit from combining switching and compute.
The useful question is: What role does the board need to play in the complete system?
This blog will investigate five of WOLF’s Blackwell VPX modules to show how they offer different answers, from dedicated 3U GPU processing to integrated switching and dual-GPU 6U architectures.
What Changes Beyond the GPU
The WOLF-1636, WOLF-163L and WOLF-163S each integrate one NVIDIA RTX 5000 Blackwell embedded GPU with 10,496 CUDA cores, 320 fifth-generation Tensor Cores and 24 GB of GDDR7 memory with ECC support. The 6U WOLF-2638 uses two of these GPUs.
The WOLF-167S uses an RTX 2000 Blackwell embedded GPU with 3,328 CUDA cores, 104 fifth-generation Tensor Cores and 8 GB of GDDR7 memory with ECC support.
Beyond GPU capacity, the key differences are connectivity, switching, video output and power requirements.
| Product | VPX format and GPU | Architectural focus | Best suited for |
|---|---|---|---|
| WOLF-1636 | 3U; one RTX 5000 Blackwell | GPU compute and DisplayPort output | GPU acceleration with direct video output |
| WOLF-163L | 3U; one RTX 5000 Blackwell | GPU compute with NVIDIA ConnectX-7 SmartNIC | Networked processing and high-rate sensor ingest |
| WOLF-163S | 3U; one RTX 5000 Blackwell | GPU compute with PCIe Gen4 and 200GbE switching | Combined switching and RTX 5000-class processing |
| WOLF-167S | 3U; one RTX 2000 Blackwell | PCIe Gen4 and 200GbE switching with GPU compute | Switch-slot integration with a smaller GPU and power envelope |
| WOLF-2638 | 6U; two RTX 5000 Blackwell GPUs | Dual-GPU compute, PCIe Gen4 switching and up to eight DisplayPort outputs | Parallel workloads and multi-output visualization |
WOLF-1636: Compute and Direct Video Output
The WOLF-1636 provides RTX 5000 Blackwell processing for AI, high-performance computing and visualization, without an onboard network or fabric switch.
The GPU connects to the VPX backplane through configurable PCIe x8 or x16 interfaces. DisplayPort outputs can be available at the front panel or, in selected OpenVPX configurations, routed through the rear backplane.
For platforms that handle networking and switching elsewhere, the 1636 lets this slot concentrate on workloads such as AI inference, image enhancement, sensor processing and high-resolution graphics.
Module power is configurable from 90 W to 150 W.
WOLF-163L: Compute Close to the Network
In a sensor-rich system, moving data to the GPU can be as consequential as processing it.
The WOLF-163L pairs its RTX 5000 Blackwell GPU with an NVIDIA ConnectX-7 SmartNIC. It supports up to 100GbE, RDMA over Converged Ethernet and NVIDIA GPUDirect RDMA.
Depending on the system design, these capabilities can reduce intermediate data copies and CPU involvement as information moves between the network and GPU. This makes the 163L a useful option for multi-sensor fusion, real-time imaging and distributed processing where data movement may limit performance.
Module power is configurable from 80 W to 140 W.
WOLF-163S: Compute and Switching in a Single Slot
The WOLF-163S combines an RTX 5000 Blackwell GPU, a configurable PCIe Gen4 switch and a 200GbE network switch on one 3U VPX module. It can serve as both a compute node and part of the chassis fabric.
The Ethernet switch supports up to eight 25GbE ports, with Time-Sensitive Networking support. Consolidating these functions can free a slot for another payload, depending on the backplane and system architecture.
That integration is most useful when slot efficiency is a priority. Separate compute and switching cards may better support requirements for independent upgrades, redundancy or fault containment.
WOLF-167S: A Switch Slot with Onboard Compute
The WOLF-167S also combines GPU compute, PCIe switching and Ethernet switching in one 3U VPX slot, but uses an RTX 2000 Blackwell embedded GPU. It suits systems that need switching alongside processing workloads that do not require RTX 5000-class capacity.
The GPU has a dedicated PCIe x8 data path. Configurable PCIe Gen4 and 200GbE switches route chassis traffic, with up to eight 25GbE ports, one 1000BASE-T port and Time-Sensitive Networking support.
The 167S aligns with SOSA switch-slot profile 14.4.15 and has no DisplayPort or HDMI output. Its configurable power range is 75 W to 140 W, making switch-slot compatibility, workload and thermal budget central to the choice.
WOLF-2638: Two GPUs for 6U Compute and Visualization
The WOLF-2638 brings two RTX 5000 Blackwell embedded GPUs into one 6U VPX module. Each GPU has 24 GB of GDDR7 memory with ECC support, with connectivity managed through a configurable PCIe Gen4 switch.
This architecture supports workloads that can be divided into parallel sensor pipelines or run AI processing and visualization concurrently. Up to eight DisplayPort outputs, with HDMI options by configuration, also support substantial display requirements.
Module power is configurable from 190 W to 320 W, with conduction-cooled, air-cooled and liquid flow-through options. Designers should establish backplane compatibility, display routing and sustained cooling capacity early, and confirm that the workload can make effective use of both GPUs.
Choosing for the Complete System
Before selecting a module, consider:
- Data movement: Does the application need direct video output, high-rate network ingest or both?
- Switching: Does the chassis already provide Ethernet and PCIe switching, or would integration free a needed slot?
- Compute: Will an RTX 2000-class GPU meet the workload, or is RTX 5000-class or dual-GPU processing required?
- Integration: Does the backplane support the required slot profile, and can the chassis sustain the module’s power and cooling needs?
In rugged embedded computing, GPU specifications are the starting point. Connectivity, heat removal, mechanical design and system management determine how effectively that performance serves the finished platform.
As an NVIDIA OEM Partner, WOLF integrates Blackwell GPUs into VPX architectures designed for demanding operating environments. Start with the workload and chassis, then choose the module whose architecture supports both.
Discuss your compute, networking, video and integration requirements with a WOLF expert to identify the right Blackwell VPX module for your system.
About Wolf Advanced Technology
Wolf Advanced Technology (WOLF) delivers rugged, high-performance embedded computing, AI, and video processing solutions for aerospace and defense. Leveraging NVIDIA® GPUs and AMD®/Xilinx® FPGAs, WOLF delivers SWaP-optimized VPX (SOSA® aligned), XMC, MXM/MXC, VNX+, SFF, and custom solutions, enabling real-time video, AI inference, and high-speed data in mission-critical environments.