Platforms · NVIDIA

Jetson Thor

NVIDIA's next-generation embedded AI module built on the Blackwell GPU architecture. Thor is designed as the next platform for physical AI and advanced robotics applications and enables edge AI such as LLM/VLM and foundation models.

14x Arm Neoverse V3AE @ 2.6 GHz Blackwell 128 GB LPDDR5X 40-130 W

Max Power at 130 W

CPU available: 50.2%; 7.0 of 14 cores free
Missions completed: 10; in 900 s
Control loop misses: 1.58/s; against a 30 Hz target
VLM queries completed: 25 of 35; scene-understanding queries

CPU

Total load, its distribution across cores, and the clock behind it.

#CPU utilization

Total CPU utilization across all cores, sampled once a second for the length of the run.

Data table (4 rows)

Metric	Mean	p50	p95	Max
CPU total (%)	49.77	49.90	53.30	60.60
Peak core (%)	90.95	90.90	93.30	100.00
Mean core (%)	50.76	50.90	54.30	61.30
Active cores (%)	99.84	100.00	100.00	100.00

Source: max_power/jetson_thor · CSV

Supporting charts

Per-core utilization over time

#Per-core utilization over time

One row per core. Persistent dark bands are cores that stay pinned for the whole run.

Data table (14 rows)

Core	Mean (%)
0	51.9
1	46.9
2	51.0
3	50.7
4	50.3
5	47.4
6	48.1
7	47.9
8	58.7
9	56.8
10	34.4
11	54.5
12	61.2
13	50.9

Source: max_power/jetson_thor · CSV

Per-core utilization distribution

#Per-core utilization distribution

Every per-core sample in the run, bucketed into 5% bands.

Data table (20 rows)

Utilization band (%)	Samples
0–5	18
5–10	10
10–15	30
15–20	141
20–25	447
25–30	870
30–35	1381
35–40	1866
40–45	1883
45–50	1388
50–55	931
55–60	505
60–65	290
65–70	186
70–75	123
75–80	102
80–85	97
85–90	1362
90–95	950
95–100	20

Source: max_power/jetson_thor · CSV

CPU clock frequency

#CPU clock frequency

Sustained CPU clock frequency under load.

Data table (1 rows)

Metric	Mean	p50	p95	Max
CPU frequency (MHz)	1,977.29	1,970.35	2,333.60	2,592.00

Source: max_power/jetson_thor · CSV

GPU and memory

Accelerator load from the VLM workload, and the memory it leaves behind.

#GPU utilization

GPU utilization while the VLM issues queries continuously.

Data table (3 rows)

Metric	Mean	p50	p95	Max
GPU utilization (%)	89.14	97.00	98.00	98.00
GPU clock (MHz)	1,558.34	1,575.00	1,575.00	1,575.00
Effective throughput (GHz)	140.17	152.78	154.35	154.35

Source: max_power/jetson_thor · CSV

Supporting charts

System memory usage

#System memory usage

System memory in use across the run, as a share of the total fitted.

Data table (3 rows)

Metric	Mean	p50	p95	Max
RAM (%)	63.95	65.00	66.00	66.80
RAM used (MB)	78,639.75	79,990.60	81,163.41	82,250.30
Swap (%)	2.90	2.90	2.90	2.90

Source: max_power/jetson_thor · CSV

Power and thermal

Board draw and the thermal margin left underneath it.

#Board power

Board power draw across the run, against the configured TDP.

Data table (1 rows)

Metric	Mean	p50	p95	Max
Board power (W)	91.85	96.08	105.89	111.55

Source: max_power/jetson_thor · CSV

Supporting charts

CPU and GPU temperature

#CPU and GPU temperature

CPU and GPU package temperature under sustained load, against a 100 °C throttle point.

Data table (2 rows)

Metric	Mean	p50	p95	Max
CPU temperature (°C)	64.94	66.85	71.00	72.90
GPU temperature (°C)	65.52	67.20	71.80	74.10

Source: max_power/jetson_thor · CSV

Clock against temperature

#Clock against temperature

A clock that falls as temperature climbs is thermal throttling; a flat clock means the platform held its performance for the whole run.

Data table (2 rows)

Metric	Mean	p50	p95	Max
CPU temperature (°C)	64.94	66.85	71.00	72.90
CPU clock (MHz)	1,977.29	1,970.35	2,333.60	2,592.00

Source: max_power/jetson_thor · CSV

Compute delivered per watt

#Compute delivered per watt

CPU utilization divided by instantaneous board power, showing how efficiency moves as the workload shifts between planning and perception.

Data table (3 rows)

Metric	Value
Mean available compute per watt (GHz-cores/W)	0.151
Mean GPU utilization per watt (%/W)	0.971
Mean board power (W)	91.9

Source: max_power/jetson_thor · CSV

System behaviour

Stability, I/O, and process load, the context around the headline counters.

#CPU against GPU utilization

Each point is one second of the run. A tight cluster means the two resources move together; a wide spread means one is waiting on the other.

Data table (2 rows)

Metric	Mean	p50	p95	Max
CPU total (%)	49.77	49.90	53.30	60.60
GPU utilization (%)	89.14	97.00	98.00	98.00

Source: max_power/jetson_thor · CSV

Supporting charts

Rolling stability

#Rolling stability

Standard deviation over a 30-sample window. Spikes are moments the load became bursty rather than steady.

Data table (2 rows)

Series	Window (samples)
CPU total	30
GPU utilization	30

Source: max_power/jetson_thor · CSV

Network throughput

#Network throughput

Sensor data arriving over the wired DDS link from the simulation machine. Every platform is offered the same load.

Data table (3 rows)

Metric	Mean	p50	p95	Max
Network received (MB/s)	66.31	66.40	67.20	68.80
Network sent (MB/s)	0.28	0.30	0.30	0.40
Network errors	1.00	1.00	1.00	1.00

Source: max_power/jetson_thor · CSV

Disk throughput

#Disk throughput

Disk read and write throughput across the run.

Data table (2 rows)

Metric	Mean	p50	p95	Max
Disk read (MB/s)	0.00	0.00	0.00	0.00
Disk write (MB/s)	0.06	0.00	0.20	1.10

Source: max_power/jetson_thor · CSV

Load average

#Load average

Run-queue depth. A load average well above the core count means threads are waiting for a core rather than running on one.

Data table (3 rows)

Metric	Mean	p50	p95	Max
Load average (1m)	8.68	8.88	10.03	10.53
Load average (5m)	6.55	7.02	8.72	8.94
Load average (15m)	4.19	4.27	6.02	6.19

Source: max_power/jetson_thor · CSV

Process count

#Process count

Total processes on the machine while the benchmark runs.

Data table (1 rows)

Metric	Mean	p50	p95	Max
Process count	537.84	537.00	542.00	543.00

Source: max_power/jetson_thor · CSV

All metrics for Max Power

Mean / p95 / max for every captured metric, across every run in this category.
Metric	NVIDIA Jetson Thor
Metric	Mean	p95	Max
CPU Total (%)	49.77	53.30	60.60
CPU Frequency (MHz)	1,977.29	2,333.60	2,592.00
Peak Core Utilization (%)	90.95	93.30	100.00
Mean Core Utilization (%)	50.76	54.30	61.30
Active Cores (%)	99.84	100.00	100.00
RAM Usage (%)	63.95	66.00	66.80
RAM Used (MB)	78,639.75	81,163.41	82,250.30
RAM Total (MB)	125,771.70	125,771.70	125,771.70
Swap Usage (%)	2.90	2.90	2.90
GPU Utilization (%)	89.14	98.00	98.00
GPU Clock (MHz)	1,558.34	1,575.00	1,575.00
GPU Effective Throughput (GHz)	140.17	154.35	154.35
Board Power (W)	91.85	105.89	111.55
CPU Temperature (°C)	64.94	71.00	72.90
GPU Temperature (°C)	65.52	71.80	74.10
Disk Read (MB/s)	0.00	0.00	0.00
Disk Write (MB/s)	0.06	0.20	1.10
Disk Usage (%)	13.60	13.60	13.60
Net Recv (MB/s)	66.31	67.20	68.80
Net Sent (MB/s)	0.28	0.30	0.40
Network Errors	1.00	1.00	1.00
Load Average (1m)	8.68	10.03	10.53
Load Average (5m)	6.55	8.72	8.94
Load Average (15m)	4.19	6.02	6.19
Process Count	537.84	542.00	543.00

Raw samples: CSV · max_power/jetson_thor

Max Power (AI Workload Optimized) at 130 W

CPU available: 52.8%; 7.4 of 14 cores free
Missions completed: 10; in 900 s
Control loop misses: 1.85/s; against a 30 Hz target
VLM queries completed: 67 of 71; scene-understanding queries

CPU

Total load, its distribution across cores, and the clock behind it.

#CPU utilization

Total CPU utilization across all cores, sampled once a second for the length of the run.

Data table (4 rows)

Metric	Mean	p50	p95	Max
CPU total (%)	47.17	47.40	50.20	53.50
Peak core (%)	89.65	89.90	93.30	100.00
Mean core (%)	47.76	48.00	50.90	54.20
Active cores (%)	99.65	100.00	100.00	100.00

Source: max_power_optimized/jetson_thor_optimized · CSV

Supporting charts

Per-core utilization over time

#Per-core utilization over time

One row per core. Persistent dark bands are cores that stay pinned for the whole run.

Data table (14 rows)

Core	Mean (%)
0	47.5
1	54.2
2	45.3
3	61.5
4	50.8
5	47.1
6	48.1
7	46.7
8	45.2
9	47.0
10	36.2
11	53.7
12	37.4
13	48.0

Source: max_power_optimized/jetson_thor_optimized · CSV

Per-core utilization distribution

#Per-core utilization distribution

Every per-core sample in the run, bucketed into 5% bands.

Data table (20 rows)

Utilization band (%)	Samples
0–5	42
5–10	59
10–15	164
15–20	421
20–25	770
25–30	1192
30–35	1495
35–40	1678
40–45	1463
45–50	1014
50–55	750
55–60	542
60–65	306
65–70	161
70–75	137
75–80	142
80–85	314
85–90	1380
90–95	545
95–100	25

Source: max_power_optimized/jetson_thor_optimized · CSV

CPU clock frequency

#CPU clock frequency

Sustained CPU clock frequency under load.

Data table (1 rows)

Metric	Mean	p50	p95	Max
CPU frequency (MHz)	1,880.34	1,865.55	2,225.66	2,561.10

Source: max_power_optimized/jetson_thor_optimized · CSV

GPU and memory

Accelerator load from the VLM workload, and the memory it leaves behind.

#GPU utilization

GPU utilization while the VLM issues queries continuously.

Data table (3 rows)

Metric	Mean	p50	p95	Max
GPU utilization (%)	91.57	98.00	98.00	98.00
GPU clock (MHz)	1,519.61	1,575.00	1,575.00	1,575.00
Effective throughput (GHz)	144.20	154.35	154.35	154.35

Source: max_power_optimized/jetson_thor_optimized · CSV

Supporting charts

System memory usage

#System memory usage

System memory in use across the run, as a share of the total fitted.

Data table (3 rows)

Metric	Mean	p50	p95	Max
RAM (%)	81.59	81.60	81.70	81.80
RAM used (MB)	101,367.66	101,399.75	101,546.81	101,685.80
Swap (%)	0.10	0.10	0.10	0.10

Source: max_power_optimized/jetson_thor_optimized · CSV

Power and thermal

Board draw and the thermal margin left underneath it.

#Board power

Board power draw across the run, against the configured TDP.

Data table (1 rows)

Metric	Mean	p50	p95	Max
Board power (W)	83.76	86.67	100.48	129.49

Source: max_power_optimized/jetson_thor_optimized · CSV

Supporting charts

CPU and GPU temperature

#CPU and GPU temperature

CPU and GPU package temperature under sustained load, against a 100 °C throttle point.

Data table (2 rows)

Metric	Mean	p50	p95	Max
CPU temperature (°C)	61.51	63.00	66.40	67.70
GPU temperature (°C)	62.41	63.90	67.40	72.80

Source: max_power_optimized/jetson_thor_optimized · CSV

Clock against temperature

#Clock against temperature

A clock that falls as temperature climbs is thermal throttling; a flat clock means the platform held its performance for the whole run.

Data table (2 rows)

Metric	Mean	p50	p95	Max
CPU temperature (°C)	61.51	63.00	66.40	67.70
CPU clock (MHz)	1,880.34	1,865.55	2,225.66	2,561.10

Source: max_power_optimized/jetson_thor_optimized · CSV

Compute delivered per watt

#Compute delivered per watt

CPU utilization divided by instantaneous board power, showing how efficiency moves as the workload shifts between planning and perception.

Data table (3 rows)

Metric	Value
Mean available compute per watt (GHz-cores/W)	0.166
Mean GPU utilization per watt (%/W)	1.093
Mean board power (W)	83.8

Source: max_power_optimized/jetson_thor_optimized · CSV

System behaviour

Stability, I/O, and process load, the context around the headline counters.

#CPU against GPU utilization

Each point is one second of the run. A tight cluster means the two resources move together; a wide spread means one is waiting on the other.

Data table (2 rows)

Metric	Mean	p50	p95	Max
CPU total (%)	47.17	47.40	50.20	53.50
GPU utilization (%)	91.57	98.00	98.00	98.00

Source: max_power_optimized/jetson_thor_optimized · CSV

Supporting charts

Rolling stability

#Rolling stability

Standard deviation over a 30-sample window. Spikes are moments the load became bursty rather than steady.

Data table (2 rows)

Series	Window (samples)
CPU total	30
GPU utilization	30

Source: max_power_optimized/jetson_thor_optimized · CSV

Network throughput

#Network throughput

Sensor data arriving over the wired DDS link from the simulation machine. Every platform is offered the same load.

Data table (3 rows)

Metric	Mean	p50	p95	Max
Network received (MB/s)	66.25	66.40	66.91	71.00
Network sent (MB/s)	5.67	7.80	7.90	8.00
Network errors	0.00	0.00	0.00	0.00

Source: max_power_optimized/jetson_thor_optimized · CSV

Disk throughput

#Disk throughput

Disk read and write throughput across the run.

Data table (2 rows)

Metric	Mean	p50	p95	Max
Disk read (MB/s)	0.00	0.00	0.00	0.70
Disk write (MB/s)	0.04	0.00	0.10	5.70

Source: max_power_optimized/jetson_thor_optimized · CSV

Load average

#Load average

Run-queue depth. A load average well above the core count means threads are waiting for a core rather than running on one.

Data table (3 rows)

Metric	Mean	p50	p95	Max
Load average (1m)	10.75	10.95	12.38	13.76
Load average (5m)	9.38	9.97	10.95	11.09
Load average (15m)	7.34	7.51	8.82	8.87

Source: max_power_optimized/jetson_thor_optimized · CSV

Process count

#Process count

Total processes on the machine while the benchmark runs.

Data table (1 rows)

Metric	Mean	p50	p95	Max
Process count	465.33	465.00	470.00	471.00

Source: max_power_optimized/jetson_thor_optimized · CSV

All metrics for Max Power (AI Workload Optimized)

Mean / p95 / max for every captured metric, across every run in this category.
Metric	NVIDIA Jetson Thor (optimized)
Metric	Mean	p95	Max
CPU Total (%)	47.17	50.20	53.50
CPU Frequency (MHz)	1,880.34	2,225.66	2,561.10
Peak Core Utilization (%)	89.65	93.30	100.00
Mean Core Utilization (%)	47.76	50.90	54.20
Active Cores (%)	99.65	100.00	100.00
RAM Usage (%)	81.59	81.70	81.80
RAM Used (MB)	101,367.66	101,546.81	101,685.80
RAM Total (MB)	125,748.60	125,748.60	125,748.60
Swap Usage (%)	0.10	0.10	0.10
GPU Utilization (%)	91.57	98.00	98.00
GPU Clock (MHz)	1,519.61	1,575.00	1,575.00
GPU Effective Throughput (GHz)	144.20	154.35	154.35
Board Power (W)	83.76	100.48	129.49
CPU Temperature (°C)	61.51	66.40	67.70
GPU Temperature (°C)	62.41	67.40	72.80
Disk Read (MB/s)	0.00	0.00	0.70
Disk Write (MB/s)	0.04	0.10	5.70
Disk Usage (%)	27.00	27.00	27.00
Net Recv (MB/s)	66.25	66.91	71.00
Net Sent (MB/s)	5.67	7.90	8.00
Network Errors	0.00	0.00	0.00
Load Average (1m)	10.75	12.38	13.76
Load Average (5m)	9.38	10.95	11.09
Load Average (15m)	7.34	8.82	8.87
Process Count	465.33	470.00	471.00

Raw samples: CSV · max_power_optimized/jetson_thor_optimized

Verdict

Production-capable for this workload, Thor has state of the art GPU and memory performance, and is more efficient in GPU-only compute. It has a strong and established NVIDIA software backing and Isaac SDK support.

Strong

Excellent GPU compute and memory bandwidth efficiency.
Unified memory: free system RAM is directly available to the GPU for larger models.
Completed every mission and kept the VLM fed throughout.
The optimized AI workload configuration nearly tripled throughput and cut query latency by roughly 60%.
The Jetson ecosystem, including Isaac SDK, is available and mature.

Weakness

Control-loop misses roughly 3.5x those of Strix Halo at max power.
About half the CPU is consumed by the benchmark workload, leaving materially less for the rest of an application.

Specification

The vendor's numbers on the left; what this platform's sensor drivers actually cost it, measured on real hardware, on the right.

Vendor specification
CPU	14x Arm Neoverse V3AE @ 2.6 GHz
GPU architecture	Blackwell
GPU cores	2560 CUDA + 96 Tensor
Memory bandwidth	273 GB/s
NPU / DLA	None
RAM	128 GB LPDDR5X
Power (TDP)	40-130 W
Product page	NVIDIA

Measured sensor-driver load (fraction of one core, per instance)
3D LiDAR	0.220
2D safety LiDAR	0.040
RGBD camera	1.300
Total for the robot's sensor suite	4.64 cores

Setup instructions for this platform are in the platform setup guide.

Results by TDP or Configuration

Each category fixes a TDP power budget or optimized configuration and runs the workload on every platform. Open one to see how this platform behaved under those conditions, and how it compared against the others.

Max Power

Each platform configured to its maximum rated TDP, so the comparison reflects the most compute each board can deliver.

See the results

Max Power (AI Workload Optimized)

Maximum rated TDP with the AI inference stack tuned for the platform according to OEM best practices.

See the results

Balanced Power

Each platform set to a balanced power profile (except Orin AGX, max TDP 65W).

See the results