> This page is for OpenFold2, version 2.4.0.
> For other versions, use one of these documentation indexes:
> - 2.6.0 (Latest) (default): https://docs.nvidia.com/nim/bionemo/openfold2/2.6.0/llms.txt
> - 2.5.0: https://docs.nvidia.com/nim/bionemo/openfold2/2.5.0/llms.txt
> - 2.4.0: https://docs.nvidia.com/nim/bionemo/openfold2/2.4.0/llms.txt
> - 2.3.0: https://docs.nvidia.com/nim/bionemo/openfold2/2.3.0/llms.txt
> - 2.2.0: https://docs.nvidia.com/nim/bionemo/openfold2/2.2.0/llms.txt
> - 2.1.0: https://docs.nvidia.com/nim/bionemo/openfold2/2.1.0/llms.txt
> - 2.0.0: https://docs.nvidia.com/nim/bionemo/openfold2/2.0.0/llms.txt
> - 1.2.0: https://docs.nvidia.com/nim/bionemo/openfold2/1.2.0/llms.txt
> - 1.1.0: https://docs.nvidia.com/nim/bionemo/openfold2/1.1.0/llms.txt
> - 1.0.0: https://docs.nvidia.com/nim/bionemo/openfold2/1.0.0/llms.txt

> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/nim/bionemo/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/nim/bionemo/_mcp/server.

# Performance

## Version 2.4.0

This section reports OpenFold2 v2.4.0 performance using internal benchmark artifacts.

The benchmark set contains four protein chains:

| Test ID | Seq Length |
|---------|------------|
| 7WBN_A  | 98         |
| 7ONG_A  | 304        |
| 7ZHT_A  | 562        |
| 7Y4I_A  | 914        |

### Benchmark Configuration

| Parameter | Setting |
|-----------|---------|
| selected_models | [1,2,3,4,5] |
| num_trials | 2 |

### Table 1: Performance Across the Supported NVIDIA Hardware Units

The table below reports `pipeline_time_mean` (seconds) with TensorRT backend and no structural templates.

| Hardware | 7WBN_A (98) | 7ONG_A (304) | 7ZHT_A (562) | 7Y4I_A (914) |
|----------|-------------|--------------|--------------|--------------|
| NVIDIA A100 80GB | 7.08 | 34.16 | 73.82 | 171.75 |
| NVIDIA B200 | 5.68 | 22.56 | 48.49 | 104.95 |
| NVIDIA H100 80GB HBM3 | 4.60 | 20.03 | 42.66 | 97.04 |
| NVIDIA GB200 | 7.15 | 18.16 | 39.93 | 83.19 |
| NVIDIA H200 | 4.35 | 16.20 | 37.60 | 85.28 |
| NVIDIA L40S | 12.67 | 36.40 | 81.56 | 183.18 |
| NVIDIA GB10 (DGX Spark) | 29.70 | 137.10 | 402.87 | 1131.11 |
| NVIDIA RTX 6000 Ada | 11.28 | 33.56 | 80.64 | 187.23 |
| NVIDIA RTX PRO 6000 Blackwell | 5.68 | 27.59 | 66.81 | 154.09 |
| NVIDIA GH200 | 5.63 | 13.81 | 29.65 | 69.50 |

### Table 2: Performance Across Optimization Backends

The table below compares H100 performance between PyTorch and TensorRT backends without structural templates.

| Test ID | Seq Length | torch (s) | trt (s) | trt-speedup-over-torch |
|---------|------------|-----------|---------|------------------------|
| 7WBN_A  | 98         | 29.74     | 4.60    | 6.47x                  |
| 7ONG_A  | 304        | 53.34     | 20.03   | 2.66x                  |
| 7ZHT_A  | 562        | 98.38     | 42.66   | 2.31x                  |
| 7Y4I_A  | 914        | 199.99    | 97.04   | 2.06x                  |

### Table 3: Performance Impact From Structural Templates

The table below reports H100 TensorRT performance with and without structural templates.

| Test ID | Seq Length | Without structural templates (s) | With structural templates (s) |
|---------|------------|----------------------------------|-------------------------------|
| 7WBN_A  | 98         | 4.60                             | 5.29                          |
| 7ONG_A  | 304        | 20.03                            | 19.53                         |
| 7ZHT_A  | 562        | 42.66                            | 43.06                         |
| 7Y4I_A  | 914        | 97.04                            | 97.77                         |

## Version 2.3.0

Version 2.3.0 adds support for GB10 (DGX Spark) GPU architecture with optimized performance for this platform.

### Performance on GB10 (DGX Spark)

Below are benchmark times, measured for each input chain, in sequential execution on a single **NVIDIA GB10 (DGX Spark)** device.

| protein chain id | metric                      | 7WBN_A    |  7ONG_A  |  7ZHT_A |  7Y4I_A |
|------------------|-----------------------------|-----------|----------|---------|---------|
| sequence length  |  sequence length            |   98      |  304     |  562    |  914    |
| Version 2.3.0    |   pipeline_time             |   46.51   |  170.78  |  441.02 |  1079.12|
| Version 2.3.0    |   pipeline_time_per_model   |   9.30    |  34.16   |  88.20  |  215.82 |

*pipeline_time is defined as the time to load parameter sets, compute features, and the sum of the time
to complete the forward pass for each model in [model_1, model_2, model_3, model_4, model_5]

**pipeline_time divided by 5

### Accuracy on GB10

Accuracy metrics for GB10 (DGX Spark) remain consistent with other supported GPU architectures. For detailed accuracy benchmarks, refer to the [Accuracy Metrics](#accuracy-metrics-on-h100) in the Version 2.0.0 section below.

For performance benchmarks on other GPU architectures, refer to the [Version 2.0.0](#version-200) section below.

## Version 2.2.0

Version 2.2.0 maintains the same performance characteristics as Version 2.1.0. All performance benchmarks and accuracy metrics from Version 2.1.0 apply to Version 2.2.0.

## Version 2.1.0

Version 2.2.0 maintains the same performance characteristics as version 2.1.0. All performance benchmarks and accuracy metrics from version 2.1.0 apply to version 2.2.0.

For detailed performance comparisons and benchmarks, see the [Version 2.0.0](#version-200) section below.

## Version 2.0.0

Compared to Version 1.0.0, Version 2.0.0
  - Removes HHR-based template processing and
  - Adds TensorRT support for a wide set of GPU architectures. Refer to the Release Notes.
  - Adds support for 'explicit_templates'. Refer to the [Migrate from HHR-based Templates to Explicit mmCIF Templates](/nim/bionemo/openfold2/2.4.0/migrating-from-hhr-to-explicit-templates).

### Performance and Accuracy

- **Faster startup**: Reduced initialization time due to removal of large template database loading (~300GB)
- **Enhanced GPU support**: TensorRT optimization for L40S, B200, and RTX 6000 Ada Generation
- **Reduced storage footprint**: Container and cache requirements significantly reduced (from 380GB to 80GB total)

#### Performance Comparison (Time in seconds) on H100

- Below are benchmark times, measured for each input chain, in sequential execution on a single **NVIDIA H100 80GB HBM3** device.

| protein chain id | metric                      | 7WBN_A    |  7ONG_A  |  7ZHT_A |  7Y4I_A |
|------------------|-----------------------------|-----------|----------|---------|---------|
| sequence length  |  sequence length            |   98      |  304     |  562    |  914    |
| Version 1.0.0    |   pipeline_time             |   26.6    |  50.9    |  98.2   |  205.9  |
| Version 1.0.0    |   pipeline_time_per_model   |   5.3     |  10.2    |  19.6   |  41.2   |
| Version 2.0.0    |   pipeline_time             |   4.92    |  19.4    |  44.9   |  113    |
| Version 2.0.0    |   pipeline_time_per_model   |   0.984   |  3.90    |  8.99   |  22.6   |
| Version 2.0.0    |   speed-up vs 1.0.0         |   5.4x    |  2.6x    |  2.2x   |  1.8x   |


*pipeline_time is defined as the time to load parameter sets, compute features, and the sum of the time
to complete the forward pass for each model in [model_1, model_2, model_3, model_4, model_5]

**pipeline_time divided by 5


#### Accuracy Metrics on H100

| Metric  | Version 1.0.0 | Version 2.0.0 |
|---------|---------------|---------------|
| CADS    | 0.744         |    0.740 |
| LDDT    | 0.861         |    0.860    |
| STRIDE  | 4.24          |    4.24  |
| MP      | 0.882         |    0.878   |


## Version 1.0.0

Performance will also vary significantly depending on:

- The type of NVIDIA GPUs that are attached and available to the NIM
- The CPU type
- System RAM available

The following section details some performance expectations and provides general tips. These are not meant to be indicative of expected performance
 and performance on your system will vary from these values.

### Recommended System Requirements

 - Refer to [Supported Hardware](/nim/bionemo/openfold2/2.4.0/support-matrix) for requirements.

### Performance Benchmarks

The following are performance benchmarks for OpenFold2 version 1.0.0.

- Structure prediction performance is mostly dependent on GPU capability and memory. If you find structure prediction to be a bottleneck, consider using a higher memory device.
- The time required for structure prediction grows with sequence length.
- The time required for structure prediction grows with the total number of sequences in the alignments.
- Below are benchmark times, measured for each input chain, in sequential execution on a single **NVIDIA H100 80GB HBM3** device.
- The average value of LDDT-CA, for these protein chains, averaged over 2 runs, is 0.86.

| protein chain id             |  7WBN_A  |  7ONG_A  |  7ZHT_A |  7Y4I_A  |
|------------------------------|----------|----------|---------|----------|
|  sequence length             |    98    |  304     |  562    |   914    |
|  pipeline_time*              |    26.6  |  50.9    |  98.2   |  205.9    |
|  pipeline_time_per_model**   |    5.3   | 10.2    | 19.6  | 41.2   |

 *pipeline_time is defined as the time to load parameter sets, compute features, and the sum of the time
to complete the forward pass for each model in [model_1, model_2, model_3, model_4, model_5]

 **pipeline_time divided by 5

#### Version 1.0.0 Configuration

| algo feature / parameter                  |   setting    |
--------------------------------------------|--------------|
|      use_templates                        |    False     |
|     selected_models                       | [1,2,3,4,5]  |
|    relax_prediction                       |    False     |
| deepspeed evoformer kernel                |    active    |
| precision for deepspeed evoformer kernel  |    bf16      |
| precision for the rest of the model       |    fp32      |