T5000 is the MODULE; the Jetson AGX Thor Developer Kit is the carrier board built around one, adding a 1 TB NVMe drive, I/O breakout and cooling, and sharing every compute specification with it. All figures on this row are MAXN mode, the top performance mode, and every one of them is lower in the fixed 70W, 90W and 120W modes the datasheet also tabulates. TWO DELIBERATE REFUSALS WORTH KNOWING ABOUT. First, tensor_core_count is empty: revision v1.4 of the datasheet explicitly removed the tensor core count per its own Document History, so the widely repeated "96 fifth-generation Tensor Cores" survives only in marketing copy and is not carried here. Second, NVIDIA's product page advertises the module as "power configurable between 40 W and 130 W" while the module datasheet documents no mode below 70W; the 130W ceiling agrees exactly and is what tdp_watts and max_power_watts carry, but the 40W floor is unverifiable against the datasheet and is most likely a custom profile from NVIDIA's Jetson Power Estimator. NVIDIA prints dense and sparse in adjacent columns for both FP4 and FP8, so no dense figure below is derived, and datasheet Note 5 states "FP8 TFLOPs is equivalent to INT8 TOPs", which is where the INT8 row comes from.

NVIDIA Jetson AGX Thor T5000 128GB

Blackwell Module 2025 Spec confidence: Official
FP8 (dense)
517
TFLOPS
FP32
8.1
TFLOPS
VRAM
128 GB
LPDDR5X
Bandwidth
273 GB/s
memory
TDP
130 W
62.0 TFLOPS/kW
Download datasheet

Overview

Jetson Thor is the first Blackwell part small enough to bolt to a robot, and the jump over Orin is large enough to be worth the disruption. The T5000 module carries a 2560-core Blackwell GPU with fifth-generation tensor cores rated at 2070 sparse FP4 TFLOPS, or 1035 dense, next to the 275 sparse INT8 TOPS the Jetson AGX Orin managed. Just as important, it has 128 GB of 256-bit LPDDR5X at 273 GB/s, twice the AGX Orin's memory, which is what allows a genuinely large model to run on a mobile platform rather than a small quantised one. The CPU moves to 14 Arm Neoverse-V3AE cores at 2.6 GHz, and the GPU supports MIG, so a single module can be partitioned across several workloads. Power runs to 130 W at the ceiling, which is a real design constraint for anything battery powered. The developer kit is $3,499.

Performance

Peak theoretical throughput by precision type

PrecisionDense2:4 Sparse
FP64
No verified data available Structured sparsity is a tensor-core feature; this vector precision has no sparse form
FP32
32-bit floating point
8.1TFLOPS Structured sparsity is a tensor-core feature; this vector precision has no sparse form
TF32
No verified data available The part supports 2:4 sparsity, but the vendor publishes no sparse figure for this precision
BF16
No verified data available The part supports 2:4 sparsity, but the vendor publishes no sparse figure for this precision
FP16
No verified data available The part supports 2:4 sparsity, but the vendor publishes no sparse figure for this precision
FP8
8-bit floating point
517TFLOPS 1,035TFLOPS
FP6
No verified data available The part supports 2:4 sparsity, but the vendor publishes no sparse figure for this precision
FP4
4-bit floating point
1,035TFLOPS 2,070TFLOPS
INT8
8-bit integer
517TOPS 1,035TOPS

Specifications

Architecture

Blackwell

Form Factor

Module

Launch Year

2025

Process Node

No verified data available

Memory

128 GB LPDDR5X

Bandwidth

273 GB/s

TDP

130 W

Max Power

130 W

CUDA Cores

2,560

Spec Confidence

Official

Full Specifications

Compute Engine
CUDA Cores 2,560
Streaming Multiprocessors 20
Boost Clock 1.57 GHz
Memory
VRAM 128 GB
Memory Type LPDDR5X
Bandwidth 273 GB/s
Interface Width 256-bit
Power & Thermal
TDP 130 W
Max Board Power 130 W
Enterprise Features
MIG Support Yes
Sparsity Yes
General
Form Factor Module
Architecture Blackwell
Process Node No verified data available
Launch Year 2025

Datasheet & Resources

Data Provenance

Every figure traced to a source

Primary Source

Publisher
NVIDIA
Published
2026-06-01

Data Quality

Spec confidence
Official
Clock basis
Boost
Core precisions with figures
4 of 9
Normalization
All values in TFLOPS

Compare cloud providers offering on-demand GPU instances for AI training, inference, and HPC workloads.

Browse GPU Cloud Providers

Similar GPUs

Frequently Asked Questions

How many TFLOPS does the NVIDIA Jetson AGX Thor have?

The NVIDIA Jetson AGX Thor delivers 8.1 TFLOPS FP32 and 517 TFLOPS FP8 at peak. Flopper does not currently have a verified FP16 throughput figure for it.

What is the power consumption of the NVIDIA Jetson AGX Thor?

The NVIDIA Jetson AGX Thor has a TDP (Thermal Design Power) rating of 130 watts.

How much memory does the NVIDIA Jetson AGX Thor have?

The NVIDIA Jetson AGX Thor is equipped with 128 GB of VRAM with 273 GB/s of memory bandwidth.

What architecture is the NVIDIA Jetson AGX Thor based on?

The NVIDIA Jetson AGX Thor is based on the Blackwell architecture, launched in 2025.

Stay Updated on GPU Releases

Get notified when new GPUs are added or specifications are updated.

Loading verification...

No spam, unsubscribe anytime.

Back to GPUs