Five Jetson production modules have a listed count of 64 or more Tensor Cores. Each of them has exactly 64, and no module lists more than 64.
Production modules with ≥64 Tensor Cores nvidia.com +4nvidia.com ↗“2048 CUDA + 64 Tensor cores, 12-core Cortex-A78AE, 64GB LPDDR5”nvidia.com ↗“Industrial-grade Ampere SoM; up to 248 INT8 TOPS, 2048 CUDA + 64 Tensor cores”developer.nvidia.com ↗“512 CUDA + 64 Tensor cores, 8-core Carmel, 64GB LPDDR4x”developer.nvidia.com ↗“512 CUDA + 64 Tensor cores, 8-core Carmel, 32GB LPDDR4x, 10-30W”developer.nvidia.com ↗“Industrial Volta SoM; up to 30 INT8 TOPS, 512 CUDA + 64 Tensor cores”
| Module | GPU architecture | Tensor Cores | CUDA cores | Memory | Max AI (INT8 TOPS) | Max FP32 TFLOPS | Power |
|---|---|---|---|---|---|---|---|
| Jetson AGX Orin 64GB | Ampere | 64 | 2048 | 64 GB LPDDR5 | 275 (sparse) | 5.3 | 15–60 W |
| Jetson AGX Orin Industrial | Ampere | 64 | 2048 | 64 GB LPDDR5 | 248 (sparse) | 4.8 | 15–75 W |
| Jetson AGX Xavier 64GB | Volta | 64 | 512 | 64 GB LPDDR4x | 32 (dense) | 1.41 | 10–30 W |
| Jetson AGX Xavier 32GB | Volta | 64 | 512 | 32 GB LPDDR4x | 32 (dense) | 1.41 | 10–30 W |
| Jetson AGX Xavier Industrial | Volta | 64 | 512 | 32 GB LPDDR4x | 30 (dense) | 1.24 | 20–40 W |
Notes:
- Different TOPS bases: The Orin figures are sparse INT8 and the Xavier figures are dense INT8. That means the TOPS numbers can't be compared directly across the two generations.
- Lifecycle: The Jetson AGX Xavier 64GB is listed as EOL. No lifecycle status is listed for the other four. developer...developer.nvidia.com ↗“identity.lifecycle_status=EOL”
- Near miss: The Jetson AGX Orin 32GB has 56 Tensor Cores, so it falls just short of 64. What I couldn't find: There's no Tensor Core count listed for the three Blackwell-based Thor modules. Those are the Jetson AGX Thor T5000, Jetson T4000 and IGX T5000. I've left them out because I can't confirm they meet the ≥64 threshold. Their mixed-precision figures are given in FP4 instead: developer... +2developer.nvidia.com ↗“compute.gpu_cuda_cores=2560, compute.gpu_architecture=Blackwell”developer.nvidia.com ↗“compute.gpu_cuda_cores=1536, compute.gpu_architecture=Blackwell”nvidia.com ↗“IGX Thor system-on-module” nvidia.comnvidia.com ↗“1792 CUDA + 56 Tensor cores”
- AGX Thor T5000: up to 2070 FP4 TFLOPS developer...developer.nvidia.com ↗“up to 2070 FP4 TFLOPS”
- T4000: up to 1200 FP4 TFLOPS developer...developer.nvidia.com ↗“up to 1200 FP4 TFLOPS”
- IGX T5000: 2070 FP4 TFLOPS nvidia.comnvidia.com ↗“performance.max_fp4_tflops=2070 TFLOPS”
If you're deciding on raw low-precision throughput rather than Tensor Core count, the Thor modules are worth considering.