Catalyst PXIe-GPU
PXIe-GPU Compute & Graphics
Catalyst PXIe-GPU
Eliminates the Need for Separate GPU Servers — COTS PXIe-GPUs from Ampere to Blackwell for Signal, Image & Video Processing and ML / DL / LLM Inference & Training
Catalyst PXIe-GPUs are COTS products that bring cost-effective, easy-to-program, high-performance GPU compute acceleration and advanced graphics capabilities of NVIDIA professional GPUs, from Ampere to Ada Lovelace to Blackwell, to the PXIe/CPCIe platform. The Catalyst PXIe-GPU family spans entry-level RTX A1000 through high-performance Ada RTX 4000 SFF and Blackwell RTX Pro 4000 SFF, plus the NVIDIA L4 Data Center GPU for AI inference workloads.
Unmatched GPU Compute Performance — Where Data is Acquired
The Catalyst-GPU family now spans from the entry-level NVIDIA RTX A1000 (6.7 FP32 TFLOPs, 8 GB GDDR6) through the Ada Lovelace RTX 2000 SFF (12.0 TFLOPs, 16 GB) and RTX 4000 SFF (19.2 TFLOPs, 20 GB), up to the new Blackwell RTX Pro 2000 SFF (17.0 TFLOPs, 16 GB GDDR7) and RTX Pro 4000 SFF (~38.4 TF32 TFLOPs, 24 GB GDDR7). All models fit in a single PXIe/CPCIe slot at 50–70 W TGP, compatible with all NI and 3rd-party PXIe chassis at 50–72 W TGP.
FFT Performance vs. Embedded Controller
RADX Catalyst-GPU PXIe-GPU-T1000-8GB Average Performance Gains Versus Intel Xeon W-2245 8C/16T 3.9 GHz Embedded Controller Under (Windows 10) MATLAB and Python on 1k to 32M Point FP32 FFTs
ML/DL AI Acceleration
Catalyst-GPUs deliver dramatic AI inference and training acceleration over CPU-only PXIe embedded controllers. Ada and Blackwell generation GPUs add dedicated 4th-gen Tensor Cores and hardware-accelerated ray tracing. The NVIDIA L4 Data Center GPU (30.3 TFLOPs FP32, 24 GB GDDR6, 72 W) is purpose-built for AI inference, video analytics, and mixed-precision workloads in thermally constrained PXIe chassis.
RADX Catalyst-GPU Performance Gain vs. NI PXIe-8881 Intel Xeon W-2245 8C/16T 3.9 GHz Embedded Controller on MATLAB FP32 Win20 DL Inference Benchmarks
Improved Accuracy & RBW for LPI Signal Processing
Catalyst-GPU supports arbitrary-length FFT, PSD, Correlation, and other DSP algorithms. In most FPGAs, the longest practical FFT lengths are typically 8k points. In Catalyst-GPUs, 1M-point and longer FFTs are practical and may be executed in real-time or near-real-time. With longer FFTs, Low Probability of Intercept (LPI) signals become readily detectable and characterizable.
LPI Signal Detection — Arbitrary Length FFT
Same signal captured with 8k-point FFT (left) vs. 1M-point FFT (right) on Catalyst-GPU. Longer FFTs dramatically
reduce RBW, resolving signals buried in the noise floor.
8K FFT — Signal Buried in Noise
Wide RBW masks LPI signal
1M FFT — Signals Clearly Resolved
Narrow RBW reveals LPI peaks
Catalyst-GPU supports 1M+ point FFTs in real-time — enabling detection and characterization of Low Probability of Intercept (LPI) signals.
Easy-to-Program via MATLAB, Python, and C/C++
Catalyst-GPUs support programming via MATLAB™, Python, and C/C++ with NVIDIA CUDA® and OpenCL® acceleration. They support both Windows and Linux operating environments, as well as popular AI frameworks including TensorFlow, PyTorch, RAPIDS AI, and RAPIDS cuSignal. All models expose a standard PCIe Gen 4 ×16 host interface to the PXIe system controller.
- All Catalyst-GPU models are based on RADX patent-pending Catalyst-X™ design, single-slot PXIe/CPCIe form factor, BAA/TAA Compliant, Designed and Assembled in the USA, and available on GSA via TestMart. Contact RADX sales for current pricing and availability.
| RADX Catalyst-GPU Family — PXIe/CPCIe GPU Modules & SKUs | |||||||
| All models: single-slot PXIe/CPCIe · PCIe Gen 4 ×16 host interface · BAA/TAA Compliant · Designed and Assembled in the USA | |||||||
| Part Number | GPU | Architecture | VRAM | FP32 Compute | TGP | Display Outputs | Q226 FOB MSRP Excl. US Tariffs |
|---|---|---|---|---|---|---|---|
| PXIe-GPU-T400-2GB | Turing / Quadro T400 | Turing | 2 GB GDDR6 | 1.0 TFLOPS | 38 W | 3× mDP | $3,999 |
| PXIe-GPU-T600-4GB | Turing / Quadro T600 | Turing | 4 GB GDDR6 | 1.7 TFLOPS | 38 W | 4× mDP | $4,999 |
| PXIe-GPU-T1000-8GB | Turing / Quadro T1000 | Turing | 8 GB GDDR6 | 2.5 TFLOPS | 50 W | 4× mDP | $5,999 |
| PXIe-GPU-RTXA1k-8GB | Ampere RTX A1k | Ampere | 8 GB GDDR6 | 6.7 TFLOPS | 50 W | 4× mDP 1.4 | $7,999 |
| PXIe-GPU-AdaRTX2k-16GB | Ada RTX 2k | Ada Lovelace | 16 GB GDDR6 | 12.0 TFLOPS | 70 W | 4× mDP 1.4 | $8,999 |
| PXIe-GPU-AdaRTX4kSFF-20GB | Ada RTX 4k SFF | Ada Lovelace | 20 GB GDDR6 | 19.2 TFLOPS | 70 W | 4× mDP 1.4 | $11,999 |
| PXIe-GPU-BWRTXPro2k-16GB | Blackwell RTX Pro 2k | BlackwellNew | 16 GB GDDR7 | 17.0 TFLOPS | 72 W | 4× mDP 2.1 | $9,999 |
| PXIe-GPU-BWRTXPro4kSFF-24GB | Blackwell RTX Pro 4k SFF | BlackwellNew | 24 GB GDDR7 | 24.0 TFLOPS | 72 W | 4× mDP 2.1 | $15,999 |
| PXIe-GPU-AdaL4DC-24GB | Ada L4 Data Center | Ada LovelaceAI/Inference | 24 GB GDDR6 | 30.3 TFLOPS | 72 W | None (compute only) | $17,999 |
Pricing is FOB San Jose, CA and excludes US import tariffs not in effect prior to RADX PO acceptance. BAA & TAA Compliant. Available on GSA via TestMart. Delivery: ~30–60 days ARO. Contact RADX for volume and educational pricing.
Compliance & Standards
Related Products
BAA / TAA Compliant. GSA Available.
All RADX PXIe modules are COTS products that are designed and assembled in the USA and available on GSA from TestMart.
