# Image Classification on ImageNet 1k (val) (Accuracy & TPU Efficiency)

Source: [Image Classification on ImageNet 1k (val) (Accuracy & TPU Efficiency)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-1k-val-accuracy-and-tpu-efficiency>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-02-26T06:06:57.076Z

Evaluation of Image Classification performance on the validation split of the ImageNet 1k dataset, reporting standard Top-1 Accuracy alongside specific TPU hardware utilization metrics (Throughput and Core-Days).

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

| Scope | Recorded value |
| --- | --- |
| Task | Image Classification |
| Dataset (including any stated split) | ImageNet 1k (val) |
| Metric | Top-1 Accuracy |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 13 of 13 |
| Coverage | All results currently returned for this benchmark |

Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol.

## Metrics

- [Top-1 Accuracy](<https://www.sota2.com/research/sota/image-classification-on-imagenet-1k-val-accuracy-and-tpu-efficiency.md?metric=Top-1%20Accuracy>): Higher is better; unit: PERCENT.

- [Throughput (img/sec/core)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-1k-val-accuracy-and-tpu-efficiency.md?metric=Throughput%20(img%2Fsec%2Fcore)>): Higher is better; unit: OTHER.

- [TPUv3 Core-Days](<https://www.sota2.com/research/sota/image-classification-on-imagenet-1k-val-accuracy-and-tpu-efficiency.md?metric=TPUv3%20Core-Days>): Lower is better; unit: SCALAR.

## Recorded results

| Method | Top-1 Accuracy | Result date | Report paper |
| --- | --- | --- | --- |
| MPL | 90 | Not recorded | 68fa6a201e209131cea26207 |
| NFNet-F4+ | 89.2 | Not recorded | 68fa6a201e209131cea26207 |
| ALIGN | 88.64 | Not recorded | 68fa6a201e209131cea26207 |
| ViT-H/14 | 88.55 | Not recorded | 68fa6a201e209131cea26207 |
| Mixer-H/14 | 87.94 | Not recorded | 68fa6a201e209131cea26207 |
| BiT-R152x4 | 87.54 | Not recorded | 68fa6a201e209131cea26207 |
| HaloNet | 85.8 | Not recorded | 68fa6a201e209131cea26207 |
| HaloNet | 85.5 | Not recorded | 68fa6a201e209131cea26207 |
| BiT-R152x4 | 85.39 | Not recorded | 68fa6a201e209131cea26207 |
| ViT-L/16 | 85.3 | Not recorded | 68fa6a201e209131cea26207 |
| ViT-L/16 | 85.15 | Not recorded | 68fa6a201e209131cea26207 |
| Mixer-L/16 | 84.15 | Not recorded | 68fa6a201e209131cea26207 |
| Mixer-L/16 | 82.84 | Not recorded | 68fa6a201e209131cea26207 |

### MPL

Top-1 Accuracy: 90. Result ID: 698e11bdffc3901de09a8f82.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa69731e209131cea2227c.md>)

[Implementation](<https://github.com/google-research/google-research/tree/master/meta_pseudo_labels>)

Recorded evaluation configuration: {"Pre-training dataset":"unlabelled or weakly labelled data (proprietary)","Resolution":"448","Evaluation protocol":"Fine-tuned"}. This record does not assert independent reproduction.

### NFNet-F4+

Top-1 Accuracy: 89.2. Result ID: 698e11bdffc3901de09a8f7e.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

Recorded evaluation configuration: {"Pre-training dataset":"JFT-300M (proprietary)","Resolution":"448","Evaluation protocol":"Fine-tuned"}. This record does not assert independent reproduction.

### ALIGN

Top-1 Accuracy: 88.64. Result ID: 698e11bdffc3901de09a8f83.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea261e7.md>)

Recorded evaluation configuration: {"Pre-training dataset":"unlabelled or weakly labelled data (proprietary)","Resolution":"448","Evaluation protocol":"Fine-tuned"}. This record does not assert independent reproduction.

### ViT-H/14

Top-1 Accuracy: 88.55. Result ID: 698e11bdffc3901de09a8f81.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6a0b1e209131cea25b04.md>)

[Implementation](<https://github.com/google-research/vision_transformer>)

Recorded evaluation configuration: {"Pre-training dataset":"JFT-300M (proprietary)","Resolution":"448","Evaluation protocol":"Fine-tuned"}. This record does not assert independent reproduction.

### Mixer-H/14

Top-1 Accuracy: 87.94. Result ID: 698e11bdffc3901de09a8f7f.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

[Implementation](<https://github.com/google-research/vision_transformer>)

Recorded evaluation configuration: {"Pre-training dataset":"JFT-300M (proprietary)","Resolution":"448","Evaluation protocol":"Fine-tuned"}. This record does not assert independent reproduction.

### BiT-R152x4

Top-1 Accuracy: 87.54. Result ID: 698e11bdffc3901de09a8f80.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa677a1e209131cea16a7e.md>)

[Implementation](<https://github.com/tensorflow/tpu/tree/master/models/official/retinanet>)

Recorded evaluation configuration: {"Pre-training dataset":"JFT-300M (proprietary)","Resolution":"448","Evaluation protocol":"Fine-tuned"}. This record does not assert independent reproduction.

### HaloNet

Top-1 Accuracy: 85.8. Result ID: 698e11bdffc3901de09a8f7a.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6a211e209131cea26344.md>)

Recorded evaluation configuration: {"Pre-training dataset":"ImageNet-21k (public)","Resolution":"448","Evaluation protocol":"Fine-tuned"}. This record does not assert independent reproduction.

### HaloNet

Top-1 Accuracy: 85.5. Result ID: 698e11bdffc3901de09a8f86.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6a211e209131cea26344.md>)

Recorded evaluation configuration: {"Pre-training dataset":"ImageNet-21k (public)","Resolution":"384","Evaluation protocol":"Fine-tuned"}. This record does not assert independent reproduction.

### BiT-R152x4

Top-1 Accuracy: 85.39. Result ID: 698e11bdffc3901de09a8f7d.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa677a1e209131cea16a7e.md>)

[Implementation](<https://github.com/tensorflow/tpu/tree/master/models/official/retinanet>)

Recorded evaluation configuration: {"Pre-training dataset":"ImageNet-21k (public)","Resolution":"448","Evaluation protocol":"Fine-tuned"}. This record does not assert independent reproduction.

### ViT-L/16

Top-1 Accuracy: 85.3. Result ID: 698e11bdffc3901de09a8f7c.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6a0b1e209131cea25b04.md>)

[Implementation](<https://github.com/google-research/vision_transformer>)

Recorded evaluation configuration: {"Pre-training dataset":"ImageNet-21k (public)","Resolution":"448","Evaluation protocol":"Fine-tuned"}. This record does not assert independent reproduction.

### ViT-L/16

Top-1 Accuracy: 85.15. Result ID: 698e11bdffc3901de09a8f85.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6a0b1e209131cea25b04.md>)

[Implementation](<https://github.com/google-research/vision_transformer>)

Recorded evaluation configuration: {"Pre-training dataset":"ImageNet-21k (public)","Resolution":"384","Evaluation protocol":"Fine-tuned"}. This record does not assert independent reproduction.

### Mixer-L/16

Top-1 Accuracy: 84.15. Result ID: 698e11bdffc3901de09a8f7b.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

[Implementation](<https://github.com/google-research/vision_transformer>)

Recorded evaluation configuration: {"Pre-training dataset":"ImageNet-21k (public)","Resolution":"448","Evaluation protocol":"Fine-tuned"}. This record does not assert independent reproduction.

### Mixer-L/16

Top-1 Accuracy: 82.84. Result ID: 698e11bdffc3901de09a8f84.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6a201e209131cea26207.md>)

[Implementation](<https://github.com/google-research/vision_transformer>)

Recorded evaluation configuration: {"Pre-training dataset":"ImageNet-21k (public)","Resolution":"224","Evaluation protocol":"Fine-tuned"}. This record does not assert independent reproduction.

Page 1 of 1. This is an index page, not the full collection.