HWBenchedge-AI hardware data, cited & graded

Upcoming edge-AI hardware

20 announced or expected edge-AI products, newest first. Not benchmarked unless it has a platform page. unconfirmed = date or status not firmly stated by the source.

NVIDIA Jetson Orin Nano 2

NVIDIA
announced
Expected: H1 2027 unconfirmed
78 TOPS
Source

NVIDIA Jetson / IGX T3000

NVIDIA
announced
Expected: Q1 2027
865 TOPS (FP4 TFLOPS)
Source

NVIDIA Jetson T2000

NVIDIA
announced
Expected: Q1 2027
400 TOPS (FP4 TFLOPS)
Source

Intel Core Series 3 (edge/value)

Intel
shipping unconfirmed
Expected: Edge systems from Q2 2026
Source

Apple M5 Pro / M5 Max

Apple
shipping
Expected: Shipping
Source

AMD Versal AI Edge Series Gen 2 VEK385 Evaluation Kit

AMD/Xilinx
shipping
Expected: Orderable; 16-20 week lead time listed
Source

Raspberry Pi AI HAT+ 2 (Hailo-10H)

Raspberry Pi / Hailo
shipping
Expected: On sale
40 TOPS (INT4)
Source

ASUS UGen300 USB AI accelerator (Hailo-10H)

ASUS / Hailo
announced unconfirmed
Expected: TBA
Source

NVIDIA Jetson T4000

NVIDIA
shipping
Expected: Available to order (Jan 2026)
1200 TOPS (FP4 sparse TFLOPS)
Source

Intel Core Ultra Series 3 (Panther Lake)

Intel
shipping
Expected: Launched; edge systems from Q1 2026
50 TOPS
Source

AMD Ryzen AI Embedded X100 Series

AMD
sampling unconfirmed
Expected: Sampling expected H1 2026
50 TOPS
Source

AMD Ryzen AI Embedded P100 Series

AMD
sampling unconfirmed
Expected: 4-6 core: production shipments expected Q2 2026; 8-12 core sampling from Q1 2026
50 TOPS
Source

AMD Ryzen AI 400 / PRO 400 (Gorgon Point)

AMD
shipping
Expected: OEM systems from Q1 2026
60 TOPS
Source

Qualcomm Dragonwing IQ10 Series

Qualcomm
sampling unconfirmed
Expected: Sampling June 2026; commercial availability scheduled Sept 2026 (per HotHardware) unconfirmed
700 TOPS (INT8 sparse (350 dense))
Source

DEEPX DX-M2

DEEPX
announced
Expected: Engineering samples H1 2027 (per Japanese distributor) unconfirmed
80 TOPS
Source

MemryX MX4

MemryX
announced
Expected: Sampling 2027 / production 2028 per announcement; MX4 page now says 2029 unconfirmed
Source

Axelera Europa AIPU (Edge 232p / Server 250p)

Axelera AI
shipping
Expected: Shipping (launch 2026-09-15; originally H1 2026)
629 TOPS (INT8 (also INT4/INT16))
Source

Apple M5

Apple
shipping
Expected: Shipping
Source

NVIDIA Jetson AGX Thor Dev Kit / T5000

NVIDIA
shipping
Expected: Generally available (2025-08-25)
2070 TOPS (FP4 sparse TFLOPS)
Source

Hailo-10H

Hailo
shipping
Expected: Generally available (2025-07-22); automotive SOP 2026
40 TOPS (INT4 (20 INT8))
Source
How to read these numbers
Every number is labelled. Confidence: A MLPerf closed, audited · B documented third-party method · C community / review · D vendor-published, not independently verified · E estimate. Power basis: measured with the result · spec vendor spec/envelope · power n/a not published.

Machine-readable: /find.json (see llms.txt); full tables are in “Show data”. Confidence grades: A audited MLPerf closed · B documented third-party · C community · D vendor-published · E estimate. Unknown values are shown as unknown, never estimated. Generated 2026-10-16 from public sources.

Latest: data and hardware feed

Subscribe: Atom · JSON Feed. Data entries use the real ingest date of each source.

  1. new data 36 benchmark rows added: MLPerf Inference v6.0 (closed, edge)
    10 platforms (Dell Pro Max with GB10, Dell PowerEdge XE7745 (8x NVIDIA L40S), Dell PowerEdge XE7745 (16x NVIDIA L4-PCIe-24GB), Lenovo IdeaPad Gaming 3 15ARH05 (1x NVIDIA GeForce GTX 1650 Ti), …); object detection, image classification, speech-to-text, medical segmentation. Result dates not stated by the source. · source
  2. new data 2 benchmark rows added: MLPerf Inference v6.0 (open, edge)
    1 platform (NVIDIA Jetson AGX Thor); llm. Result dates not stated by the source. · source
  3. new data 18 benchmark rows added: NVIDIA Jetson Benchmarks
    1 platform (NVIDIA Jetson AGX Thor); llm, vlm. Result dates not stated by the source. · source
  4. new data 24 benchmark rows added: NVIDIA Jetson Benchmarks (MLPerf v3.1 results table)
    4 platforms (NVIDIA Jetson AGX Orin 64GB, NVIDIA Jetson AGX Orin (MaxQ), NVIDIA Jetson Orin NX 16GB, NVIDIA Jetson Orin NX 16GB (MaxQ)); image classification, object detection, medical segmentation, speech-to-text. Results dated 2026-10. · source
  5. new data 68 benchmark rows added: Ultralytics docs benchmark table
    6 platforms (NVIDIA Jetson AGX Thor, NVIDIA Jetson AGX Orin 64GB, NVIDIA Jetson Orin Nano Super, NVIDIA Jetson Orin NX 16GB, …); object detection. Result dates not stated by the source. · source
  6. new data 80 benchmark rows added: Hailo Model Zoo public models (Hailo-10H)
    1 platform (Hailo-10H (M.2/PCIe, x86 host)); image classification, instance segmentation, object detection, oriented object detection. Result dates not stated by the source. · source
  7. new data 88 benchmark rows added: Hailo Model Zoo public models (Hailo-8L)
    1 platform (Hailo-8L (M.2/PCIe, x86 host)); speech-to-text, image classification, instance segmentation, object detection. Result dates not stated by the source. · source
  8. new data 88 benchmark rows added: Hailo Model Zoo public models (Hailo-8)
    1 platform (Hailo-8 (M.2/PCIe, x86 host)); speech-to-text, image classification, instance segmentation, object detection. Result dates not stated by the source. · source
  9. new data 161 benchmark rows added: Qualcomm AI Hub Models perf.yaml (BSD-3)
    4 platforms (Qualcomm Dragonwing IQ-9075 EVK, Qualcomm Snapdragon X Elite (CRD), Qualcomm Snapdragon X2 Elite (CRD), Qualcomm Dragonwing RB3 Gen 2 (QCS6490)); semantic segmentation, pose estimation, image classification, speech-to-text. Results dated 2026 (year only). · source
  10. new data 36 benchmark rows added: llama.cpp discussion #4167 scoreboard (via GitHub API)
    9 platforms (Apple M4 (10-core GPU, 120 GB/s), Apple M4 Pro (16-core GPU, 273 GB/s), Apple M4 Pro (20-core GPU, 273 GB/s), Apple M4 Max (32-core GPU, 410 GB/s), …); llm prefill, llm generation. Results dated 2026-10-01. · source
  11. new data 36 benchmark rows added: llama.cpp discussion #10879 Vulkan scoreboard (via GitHub API)
    10 platforms (Apple M3 Ultra iGPU, Apple M4 Max iGPU, NVIDIA DGX Spark, AMD Ryzen AI Max+ (Strix Halo) iGPU, …); llm prefill, llm generation. Results dated 2026-10-10. · source
  12. new data 6 benchmark rows added: vminstall.com review
    3 platforms (Apple Mac mini M4 Pro 64GB, NVIDIA Jetson AGX Thor, NVIDIA DGX Spark); llm generation, llm prefill. Results dated 2026 (year only). · source
  13. new data 2 benchmark rows added: Open Navigation robotics workload benchmark (PDF)
    2 platforms (AMD Ryzen AI Max+ (Strix Halo) X100, NVIDIA Jetson AGX Thor); vlm robotics requests (warehouse scenes). Results dated 2026 (year only). · source
  14. new data 4 benchmark rows added: StrixHaloGuide evidence (llama-bench, 20 repeats)
    1 platform (AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)); llm prefill, llm generation. Results dated 2026-10. · source
  15. new data 1 benchmark row added: strix-halo-guide GitHub releases
    1 platform (AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)); llm generation. Results dated 2026 (year only). · source
  16. new data 10 benchmark rows added: Ultralytics docs: AMD ROCm/MIGraphX integration benchmark
    1 platform (AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)); object detection. Results dated 2026 (year only). · source
  17. new data 20 benchmark rows added: AMD technical article: AI inference acceleration on Ryzen AI with Quark (Table 5)
    1 platform (AMD Ryzen AI processor (SKU not stated)); object detection. Results dated 2025 (year only). · source
  18. new data 1 benchmark row added: AMD technical article: Deploying object detection model on AMD AI PC
    1 platform (AMD Ryzen AI processor (SKU not stated)); open-vocabulary object detection. Results dated 2026 (year only). · source
  19. new data 28 benchmark rows added: Ultralytics docs: Apple Core AI integration benchmark
    1 platform (Apple M4 (10-core GPU, 120 GB/s)); object detection, instance segmentation, semantic segmentation, depth estimation. Results dated 2026 (year only). · source
  20. new data 60 benchmark rows added: Ultralytics docs: Intel OpenVINO integration benchmark
    3 platforms (Intel Core Ultra X7 358H (Panther Lake), Intel Core Ultra 7 258V (Lunar Lake), Intel Core Ultra 7 155H (Meteor Lake)); object detection. Results dated 2026 (year only). · source
  21. new data 8 benchmark rows added: Ultralytics docs: Hailo integration benchmark
    2 platforms (Raspberry Pi 5 (CPU only), Raspberry Pi 5 + AI HAT+ (Hailo-8)); object detection, semantic segmentation, depth estimation, image classification. Results dated 2026 (year only). · source
  22. new data 10 benchmark rows added: Seeed Studio wiki: multistream inference on RPi5 + Hailo-8
    1 platform (Raspberry Pi 5 + AI HAT+ (Hailo-8)); object detection. Results dated 2025-12 (updated). · source
  23. new data 106 benchmark rows added: MLPerf Inference v5.0 (open)
    1 platform (TRIGKEY Mini PC (AMD Ryzen 5 5560U, CPU)); image classification. Results dated 2025 (year only). · source
  24. new data 4 benchmark rows added: MLPerf Inference v5.1 (open)
    1 platform (TRIGKEY Mini PC (AMD Ryzen 5 5560U, CPU)); image classification. Results dated 2025 (year only). · source
  25. new data 974 benchmark rows added: Vitis AI 3.0 Model Zoo performance table
    4 platforms (AMD Versal AI Core VCK190 (VC1902), AMD Zynq UltraScale+ ZCU102 (ZU9EG), AMD Zynq UltraScale+ ZCU104 (ZU7EV), AMD Kria KV260 Vision AI Starter Kit (K26 SOM)); semantic segmentation, object detection, super resolution, depth estimation. Results dated 2023 (year only). · source
  26. new data 269 benchmark rows added: Vitis AI 3.5 Model Zoo performance table
    2 platforms (AMD Versal AI Edge VEK280 (VE2802), AMD Alveo V70 (Versal AI Core, XDNA AIE-ML)); object detection, depth estimation, image classification, pose estimation. Results dated 2023 (year only). · source
  27. new data 410 benchmark rows added: MLPerf Inference v6.1 (closed, datacenter)
    77 platforms (8xMI350X_2xEPYC_9575F (8x AMD Instinct MI350X 288GB HBM3e), Quanta Grand Teton 1.5 MI350 (8x AMD Instinct MI350X 288GB HBM3e), smci355-ccs-aus-m13-13 (8x AMD Instinct MI355X 288GB HBM3e), 9x8xMI355X_2xEPYC_9575F (8x AMD Instinct MI355X 288GB HBM3e), …); recommendation, wan-2.2-t2v-a14b, gpt-oss-120b, llama2-70b. Result dates not stated by the source. · source
  28. new data 16 benchmark rows added: MLPerf Inference v6.1 (closed, edge)
    9 platforms (NVIDIA DGX Spark, AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S), Dell Pro Max with GB10, HPE ProLiant DL145 Gen11 (1x NVIDIA RTX PRO 4500 Blackwell Server Edition), …); qwen3.6-27b, object detection, speech-to-text. Result dates not stated by the source. · source
  29. new data 10 benchmark rows added: MLPerf Inference v6.1 (open, datacenter)
    6 platforms (8xMI355X_2xEPYC_9575F (8x AMD Instinct MI355X), NVIDIA GB200 NVL72 by HPE (4x NVIDIA GB200), HPE Cray XD670 (8x NVIDIA H200-SXM-141GB), Orrick Optimized 6x NVIDIA B200 (6x NVIDIA B200-SXM-180GB), …); recommendation, deepseek-r1, qwen3-vl-235b-a22b, llama3.1-8b. Result dates not stated by the source. · source
  30. new data 5 benchmark rows added: MLPerf Inference v6.1 (open, edge)
    3 platforms (HPE ProLiant DL345 Gen12 (4x NVIDIA RTX PRO 4500 Blackwell-PCIe-32GB), HPE ProLiant Compute DL380 Gen12 (1x NVIDIA RTX PRO 4500 Blackwell-PCIe-32GB), Lambda 8x B200-SXM-180GB (8x NVIDIA B200-SXM-180GB)); llm, image classification, qwen3.6-27b. Result dates not stated by the source. · source