Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0 vs MemryX MX3 M.2 module

GoogleAI accelerator (USB)No licensed photo yet

MemryXedge accelerator cardNo licensed photo yet

Showing the newest results per platform and source. 18 superseded older results are hidden.Showing all results, including 18 superseded by newer results from the same source.

MemryX MX3 M.2 module is faster in the only like-for-like comparison (largest gap: semantic segmentation, 6.59×).

	Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0	MemryX MX3 M.2 module
Best semantic segmentation	303 FPS D Older data (2020)	2002 FPS C 2026-07-15 (published)
Power	2.00 W spec	n/a
Perf / W	152 FPS/W spec	n/a
Price	n/a	n/a
Workloads with data	1 of 7	5 of 7

Best published semantic segmentation result per platform (may be from different like-for-like groups — see data). n/a = not published; never estimated.

Show data — all workloads, sources and like-for-like groups

object detection

Only MemryX MX3 M.2 module has published object detection results.

Platform	Result	Model · precision · stack	Grade	Power	Source
MemryX MX3 M.2 module	2391 FPS throughput · b=1	MediaPipe-object-Detection-3d-camera not stated (MX3 activations BF16 / weights 4-8-16 bit per vendor) · MemryX SDK 2.2 (closed-loop compiled DFP) · mobile-class CNN	C	unknown power n/a no power published	MemryX Model Explorer (SDK 2.2) 2026-07-15 (published)

Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0:

◌

No published object detection results for Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0 or MemryX MX3 M.2 module yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.

instance segmentation

Only MemryX MX3 M.2 module has published instance segmentation results.

Platform	Result	Model · precision · stack	Grade	Power	Source
MemryX MX3 M.2 module	1023 FPS throughput · b=1	YOLO26-nano-seg not stated (MX3 activations BF16 / weights 4-8-16 bit per vendor) · MemryX SDK 2.2 (closed-loop compiled DFP) · unknown	C	unknown power n/a no power published	MemryX Model Explorer (SDK 2.2) 2026-07-15 (published)

Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0:

◌

No published instance segmentation results for Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0 or MemryX MX3 M.2 module yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.

pose estimation

Only MemryX MX3 M.2 module has published pose estimation results.

Platform	Result	Model · precision · stack	Grade	Power	Source
MemryX MX3 M.2 module	2721 FPS throughput · b=1	MoveNet.SinglePose.Lightning not stated (MX3 activations BF16 / weights 4-8-16 bit per vendor) · MemryX SDK 2.2 (closed-loop compiled DFP) · mobile-class CNN	C	unknown power n/a no power published	MemryX Model Explorer (SDK 2.2) 2026-07-15 (published)

Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0:

◌

No published pose estimation results for Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0 or MemryX MX3 M.2 module yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.

image classification

Only MemryX MX3 M.2 module has published image classification results.

Platform	Result	Model · precision · stack	Grade	Power	Source
MemryX MX3 M.2 module	2317 FPS throughput · b=1	ResNet-50 (MXA-Optimized) not stated (MX3 activations BF16 / weights 4-8-16 bit per vendor) · MemryX SDK 2.2 (closed-loop compiled DFP) · mid-size CNN	C	unknown power n/a no power published	MemryX Model Explorer (SDK 2.2) 2026-07-15 (published)

Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0:

◌

No published image classification results for Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0 or MemryX MX3 M.2 module yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.

semantic segmentation

✓ Like-for-like: model-size class mid-size CNN, batch 1. Model, precision and stack may still differ.

Platform	Result	Model · precision · stack	Grade	Power	Source
MemryX MX3 M.2 module	1998 FPS throughput · b=1	Deeplab-v3 not stated (MX3 activations BF16 / weights 4-8-16 bit per vendor) · MemryX SDK 2.2 (closed-loop compiled DFP) · mid-size CNN	C	unknown power n/a no power published	MemryX Model Explorer (SDK 2.2) 2026-07-15 (published)
Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0	303 FPS inference_time · b=1	Unet Mv2 (128x128) INT8 · TensorFlow Lite C++ benchmark (Edge TPU runtime) · mid-size CNN	D	2.00 W spec platform spec envelope (not measured with this result)	Google Coral Edge TPU benchmarks page (Table 1) Older data (2020)

speech-to-text

◌

No published speech-to-text results for Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0 or MemryX MX3 M.2 module yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.

LLM generation

◌

No published LLM generation results for Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0 or MemryX MX3 M.2 module yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.

MemryX MX3 M.2 module is faster in all 5 like-for-like comparisons (largest gap: image classification, 114×).

	Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0	MemryX MX3 M.2 module
Best object detection	154 FPS D Older data (2019)	8519 FPS C 2026-07-15 (published)
Power	2.00 W spec	n/a
Perf / W	76.92 FPS/W spec	n/a
Price	n/a	n/a
Workloads with data	3 of 7	5 of 7

Best published object detection result per platform (may be from different like-for-like groups — see data). n/a = not published; never estimated.

Show data — all workloads, sources and like-for-like groups

object detection

✓ Like-for-like: model-size class mobile-class CNN, batch 1. Model, precision and stack may still differ.

Platform	Result	Model · precision · stack	Grade	Power	Source
MemryX MX3 M.2 module	2391 FPS throughput · b=1	MediaPipe-object-Detection-3d-camera not stated (MX3 activations BF16 / weights 4-8-16 bit per vendor) · MemryX SDK 2.2 (closed-loop compiled DFP) · mobile-class CNN	C	unknown power n/a no power published	MemryX Model Explorer (SDK 2.2) 2026-07-15 (published)
Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0	154 FPS inference_time · b=1	MobileNet v1 SSD (224x224) INT8 · TensorFlow Lite C++ benchmark (Edge TPU runtime) · mobile-class CNN	D	2.00 W spec platform spec envelope (not measured with this result)	Google Coral Edge TPU benchmarks page (Table 1) Older data (2019)

instance segmentation

Only MemryX MX3 M.2 module has published instance segmentation results.

Platform	Result	Model · precision · stack	Grade	Power	Source
MemryX MX3 M.2 module	1023 FPS throughput · b=1	YOLO26-nano-seg not stated (MX3 activations BF16 / weights 4-8-16 bit per vendor) · MemryX SDK 2.2 (closed-loop compiled DFP) · unknown	C	unknown power n/a no power published	MemryX Model Explorer (SDK 2.2) 2026-07-15 (published)

Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0:

◌

pose estimation

Only MemryX MX3 M.2 module has published pose estimation results.

Platform	Result	Model · precision · stack	Grade	Power	Source
MemryX MX3 M.2 module	2721 FPS throughput · b=1	MoveNet.SinglePose.Lightning not stated (MX3 activations BF16 / weights 4-8-16 bit per vendor) · MemryX SDK 2.2 (closed-loop compiled DFP) · mobile-class CNN	C	unknown power n/a no power published	MemryX Model Explorer (SDK 2.2) 2026-07-15 (published)

Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0:

◌

image classification

✓ Like-for-like: model-size class mid-size CNN, batch 1. Model, precision and stack may still differ.

Platform	Result	Model · precision · stack	Grade	Power	Source
MemryX MX3 M.2 module	2317 FPS throughput · b=1	ResNet-50 (MXA-Optimized) not stated (MX3 activations BF16 / weights 4-8-16 bit per vendor) · MemryX SDK 2.2 (closed-loop compiled DFP) · mid-size CNN	C	unknown power n/a no power published	MemryX Model Explorer (SDK 2.2) 2026-07-15 (published)
Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0	20.41 FPS inference_time · b=1	ResNet-50 V1 (299x299) INT8 · TensorFlow Lite C++ benchmark (Edge TPU runtime) · mid-size CNN	D	2.00 W spec platform spec envelope (not measured with this result)	Google Coral Edge TPU benchmarks page (Table 1) Older data (2019)

✓ Like-for-like: model-size class mobile-class CNN, batch 1. Model, precision and stack may still differ.

Platform	Result	Model · precision · stack	Grade	Power	Source
MemryX MX3 M.2 module	8312 FPS throughput · b=1	MobileNet not stated (MX3 activations BF16 / weights 4-8-16 bit per vendor) · MemryX SDK 2.2 (closed-loop compiled DFP) · mobile-class CNN	C	unknown power n/a no power published	MemryX Model Explorer (SDK 2.2) 2026-07-15 (published)
Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0	417 FPS inference_time · b=1	MobileNet v1 (224x224) INT8 · TensorFlow Lite C++ benchmark (Edge TPU runtime) · mobile-class CNN	D	2.00 W spec platform spec envelope (not measured with this result)	Google Coral Edge TPU benchmarks page (Table 1) Older data (2019)

✓ Like-for-like: model-size class unknown, batch 1 (size or batch not stated by source — unverified). Model, precision and stack may still differ.

Platform	Result	Model · precision · stack	Grade	Power	Source
MemryX MX3 M.2 module	10057 FPS throughput · b=1	Vehicle-Attributes-recog-barrier-0039 not stated (MX3 activations BF16 / weights 4-8-16 bit per vendor) · MemryX SDK 2.2 (closed-loop compiled DFP) · unknown	C	unknown power n/a no power published	MemryX Model Explorer (SDK 2.2) 2026-07-15 (published)
Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0	476 FPS inference_time · b=1	SqueezeNet (224x224) INT8 · TensorFlow Lite C++ benchmark (Edge TPU runtime) · unknown	D	2.00 W spec platform spec envelope (not measured with this result)	Google Coral Edge TPU benchmarks page (Table 1) Older data (2019)

semantic segmentation

✓ Like-for-like: model-size class mid-size CNN, batch 1. Model, precision and stack may still differ.

Platform	Result	Model · precision · stack	Grade	Power	Source
MemryX MX3 M.2 module	1998 FPS throughput · b=1	Deeplab-v3 not stated (MX3 activations BF16 / weights 4-8-16 bit per vendor) · MemryX SDK 2.2 (closed-loop compiled DFP) · mid-size CNN	C	unknown power n/a no power published	MemryX Model Explorer (SDK 2.2) 2026-07-15 (published)
Google Coral USB Accelerator (Edge TPU) on Xeon Gold 6154 host, USB 3.0	303 FPS inference_time · b=1	Unet Mv2 (128x128) INT8 · TensorFlow Lite C++ benchmark (Edge TPU runtime) · mid-size CNN	D	2.00 W spec platform spec envelope (not measured with this result)	Google Coral Edge TPU benchmarks page (Table 1) Older data (2020)

speech-to-text

◌

LLM generation

◌

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.