Decode capacity · Jetson Thor

Jetson Thor T5000 AV1 decode capacity

Dataset 2026-09-07 · verified 2026-09-07 · last updated September 2026

NVIDIA rates the Jetson Thor T5000 NVDEC at 4 × 8K30 · 12 × 4K60 · 24 × 4K30 · 48 × 1080p60 · 96 × 1080p30 AV1 (3120 MP/s) at 120 W.

Verbatim datasheet figures with their source. Anything not in the table below is an EdgeAIStack estimate and is labelled as such. Jetson Thor T5000 platform page.

Published decode points (class A)

Jetson Thor T5000: 2 NVDEC instances, AV1 profile Main, 8-bit 4:2:0 unless the datasheet says otherwise. Stream counts are per module (per-instance counts × 2, as the datasheet’s total column does).

120 W mode (datasheet default)

ProfilePixels @ fpsPer instanceStreams (module)MP/s per streamStreams × MP/sEvidence
8K307680×4320 @ 3024995.33,981A datasheet
4K603840×2160 @ 60612497.75,972A datasheet
4K303840×2160 @ 301224248.85,972A datasheet
1080p601920×1080 @ 602448124.45,972A datasheet
1080p301920×1080 @ 30489662.25,972A datasheet
Rated throughput (up to)3,120 MP/sA datasheet

Source

  • NVIDIA · DS-11945-001 v1.5 §2.2 Multi-Standard Video Decoder, Table 2-2 (T5000), 120W Mode · verified 2026-09-07 · Jetson Thor Series Modules Data Sheet · class A
    “AV1 Main Profile 48x 24x 12x 6x 2x 1560 3120”

MAXN mode

ProfilePixels @ fpsPer instanceStreams (module)MP/s per streamStreams × MP/sEvidence
8K307680×4320 @ 3024995.33,981A datasheet
4K603840×2160 @ 60612497.75,972A datasheet
4K303840×2160 @ 301224248.85,972A datasheet
1080p601920×1080 @ 602652124.46,470A datasheet
1080p301920×1080 @ 305410862.26,718A datasheet
Rated throughput (up to)3,400 MP/sA datasheet

Source

  • NVIDIA · DS-11945-001 v1.5 §2.2 Multi-Standard Video Decoder, Table 2-2 (T5000), MAXN Mode · verified 2026-09-07 · Jetson Thor Series Modules Data Sheet · class A
    “AV1 Main Profile 54x 26x 12x 6x 2x 1700 3400”

v1.5 dated June 2026. Tables are per NVDEC/NVENC instance and per power mode; totals here multiply by the instance count as the datasheet's "Total Mpix/sec (2x)" column does. T4000 rows are marked preliminary by NVIDIA.

Between the published points (class D)

Camera profiles NVIDIA does not print are estimated by comparing the pixel rate against the rated 3120 MP/s (120 W table): max streams = ⌊3120 ÷ (width × height × fps ÷ 10⁶)⌋. These are EdgeAIStack numbers, not NVIDIA's; they assume roughly linear NVDEC throughput inside the published range and are not asserted outside it. Rows that coincide with a published point show the datasheet count instead.

ProfilePixels @ fpsMP/s per streamMax streams (est.)EvidenceFull pipeline
720p301280×720 @ 3027.6112D pixel-ratecheck the full pipeline
1080p151920×1080 @ 1531.1100D pixel-ratecheck the full pipeline
1080p201920×1080 @ 2041.575D pixel-ratecheck the full pipeline
1080p251920×1080 @ 2551.860D pixel-ratecheck the full pipeline
1440p152560×1440 @ 1555.356D pixel-ratecheck the full pipeline
1440p202560×1440 @ 2073.742D pixel-ratecheck the full pipeline
1440p302560×1440 @ 30110.628D pixel-ratecheck the full pipeline
4K153840×2160 @ 15124.425D pixel-ratecheck the full pipeline
4K203840×2160 @ 20165.918D pixel-ratecheck the full pipeline
4K253840×2160 @ 25207.415D pixel-ratecheck the full pipeline

Power mode

On Thor the NVDEC/NVENC clock follows the power mode (1.56 GHz at 120 W / 70 W, 1.69 GHz at MAXN) and the datasheet publishes separate decode tables per mode.

Source

  • NVIDIA · DS-11945-001 v1.5 §1 Overview, HD Video table · verified 2026-09-07 · Jetson Thor Series Modules Data Sheet · class A
    “Jetson T5000: 2x NVDEC @ 120W: 1.56 GHz | MAXN: 1.69 GHz”

Encode

The datasheet publishes no hardware AV1 encode row for Jetson Thor T5000. Published encoder rows: H.264, H.265.

What these numbers do and do not cover

  • Hardware limit, not software. JetPack, DeepStream, GStreamer or Frigate may cap concurrent decoder sessions below the datasheet count.
  • Decode in isolation. Simultaneous NVENC use shares memory bandwidth; expect real decode + encode totals below the sum of the two tables.
  • Profile and chroma. Counts are for Main at 8-bit 4:2:0; 4:4:4 halves throughput (datasheet note); Main10 shares the row.
  • Decode is rarely the bottleneck. Inference, memory and preprocessing usually fail first on a detection pipeline; run the full pipeline check for your workload.
  • Method. Formulas, evidence classes and the confidence rule are on the methodology page.

Other codecs on this module

AV1 on other modules

Is decode really the limit?

Run Jetson Thor T5000 with your camera count, model and frame rate: every stage as a constraint, every number with its source, a permanent link.

CHECK THE FULL PIPELINE →