
Fine-tuning BERT Large on a GPU Workstation
For this post, we measured fine-tuning performance (training and inference) for the BERT implementation of TensorFlow on NVIDIA RTX A4500 GPUs. For testing we used an Exxact Valence Workstation fitted with 8x RTX A4500 GPUs with 20GB GPU memory per GPU.
Benchmark scripts we used for evaluation were the finetune_train_benchmark.sh and finetune_inference_benchmark.sh from NVIDIA NGC Repository BERT for TensorFlow. We made slight modifications to the training benchmark script to get the larger batch size numbers.
The script runs multiple tests on the SQuAD v1.1 dataset using batch sizes 1, 2, 4, and 8. Inferencing tests were conducted using a 1x GPU configurations on BERT Large. In addition, we ran all benchmarks using TensorFlow's XLA across the board.
Key Points and Observations
- Scenarios that are not typically used in real-world training, such as single GPU throughput are illustrated in the table below, and provided for reference as an indication of single chip throughput of the platform.
- For those interested in training BERT Large, a 2x RTX A4500 system may be a great choice to start with, giving the opportunity to add additional cards as budget/scaling needs increase.
- NOTE: In order to run these benchmarks, or be able to fine-tune BERT Large with 4x GPUs you'll need a system with at least 64GB RAM.
Interested in getting faster results?
Learn more about Exxact AI workstations to do NLP training on starting around $5,500
Exxact Workstation System Specs:
Nodes | 1 |
Processor / Count | 2x AMD EPYC 7552 |
Total Logical Cores | 48 |
Memory | DDR4 512GB |
Storage | NVMe 3.84TB |
OS | Ubuntu 18.04 |
CUDA Version | 11.4 |
BERT Dataset | squad v1 |
TensorFlow | 2.40 |
GPU Benchmark Overview
FP = Floating Point Precision, Seq = Sequence Length, BS = Batch Size
1x Quadro RTX A4500 BERT LARGE Inference Benchmark
Raw Data
Model | Sequence-Length | Batch-size | Precision | Total-Inference-Time | Throughput-Average(sent/sec) | Latency-Average(ms) | Latency-50%(ms) | Latency-90%(ms) | Latency-95%(ms) | iLatency-99%(ms) | Latency-100%(ms) |
base | 128 | 1 | fp16 | 25.98 | 186.13 | 8.94 | 5.56 | 6.34 | 6.46 | 6.8 | 6387.51 |
base | 128 | 1 | fp32 | 20.34 | 188.55 | 8.38 | 5.48 | 6.32 | 6.48 | 6.85 | 5501.44 |
base | 128 | 2 | fp16 | 20.47 | 364.38 | 8.84 | 5.67 | 6.2 | 6.32 | 6.78 | 5679.04 |
base | 128 | 2 | fp32 | 19.65 | 389.55 | 8.47 | 5.3 | 5.89 | 5.99 | 6.18 | 5654.98 |
base | 128 | 4 | fp16 | 26.39 | 695.64 | 11.55 | 5.79 | 6.24 | 6.36 | 6.9 | 5774.98 |
base | 128 | 4 | fp32 | 26.41 | 696.75 | 11.54 | 5.76 | 6.22 | 6.41 | 6.86 | 5767.47 |
base | 128 | 8 | fp16 | 34.2 | 964.18 | 13.57 | 8.27 | 8.72 | 8.94 | 9.52 | 5876.52 |
base | 128 | 8 | fp32 | 34.42 | 965.47 | 13.6 | 8.28 | 8.64 | 8.89 | 9.44 | 5899.06 |
base | 384 | 1 | fp16 | 17.03 | 177.1 | 11.46 | 5.65 | 6.11 | 6.25 | 6.7 | 6059.77 |
base | 384 | 1 | fp32 | 17.2 | 176.11 | 11.49 | 5.66 | 6.15 | 6.29 | 6.77 | 6051.43 |
base | 384 | 2 | fp16 | 25.64 | 239.89 | 19.3 | 8.31 | 8.84 | 9.13 | 9.84 | 6456.22 |
base | 384 | 2 | fp32 | 25.6 | 240.18 | 19.26 | 8.29 | 8.8 | 9.06 | 9.77 | 6447.57 |
base | 384 | 4 | fp16 | 31.28 | 305.32 | 24.29 | 13.09 | 13.62 | 13.85 | 14.97 | 6660.95 |
base | 384 | 4 | fp32 | 31.11 | 304.43 | 24.28 | 13.13 | 13.6 | 13.8 | 14.98 | 6639.34 |
base | 384 | 8 | fp16 | 43.89 | 358.82 | 34.29 | 22.27 | 22.79 | 23.1 | 24.51 | 7002.83 |
base | 384 | 8 | fp32 | 43.9 | 358.76 | 34.33 | 22.26 | 22.8 | 23.11 | 24.35 | 7031.13 |
large | 128 | 1 | fp16 | 47.42 | 103.62 | 15.96 | 9.42 | 10.95 | 11.16 | 11.67 | 11340.21 |
large | 128 | 1 | fp32 | 36.28 | 108.7 | 15.52 | 8.91 | 10.12 | 10.4 | 10.96 | 11350.15 |
large | 128 | 2 | fp16 | 38.82 | 180.4 | 17.77 | 11.09 | 11.91 | 12.05 | 12.44 | 11356.99 |
large | 128 | 2 | fp32 | 39.12 | 177.62 | 17.97 | 11.45 | 11.98 | 12.13 | 12.79 | 11389.43 |
large | 128 | 4 | fp16 | 58.85 | 253.71 | 27.84 | 15.94 | 16.48 | 16.65 | 17.52 | 11607 |
large | 128 | 4 | fp32 | 57.8 | 261.94 | 27.35 | 15.01 | 16.17 | 16.35 | 17.01 | 11606.73 |
large | 128 | 8 | fp16 | 78.49 | 348.55 | 34.15 | 22.89 | 23.82 | 24.05 | 24.71 | 11839.75 |
large | 128 | 8 | fp32 | 79.35 | 345.57 | 34.53 | 23.22 | 23.96 | 24.19 | 24.88 | 12047.05 |
large | 384 | 1 | fp16 | 34.49 | 77.09 | 24.99 | 12.98 | 13.76 | 13.89 | 14.59 | 12534.02 |
large | 384 | 1 | fp32 | 33.95 | 79.81 | 24.49 | 12.34 | 13.35 | 13.51 | 14.09 | 12472.18 |
large | 384 | 2 | fp16 | 54.4 | 97.77 | 44.04 | 20.34 | 21.21 | 21.38 | 22.2 | 13302.25 |
large | 384 | 2 | fp32 | 53.88 | 98.22 | 43.83 | 20.12 | 21.18 | 21.38 | 22.19 | 13331.07 |
large | 384 | 4 | fp16 | 68.97 | 118.22 | 57.65 | 33.95 | 34.44 | 34.62 | 35.31 | 13713.5 |
large | 384 | 4 | fp32 | 69.14 | 117.88 | 57.75 | 34.07 | 34.59 | 34.78 | 35.73 | 13693.18 |
large | 384 | 8 | fp16 | 104.65 | 126.65 | 88.76 | 63.22 | 63.87 | 64.17 | 65.32 | 14498.6 |
large | 384 | 8 | fp32 | 104.84 | 126.41 | 89 | 63.33 | 64.11 | 64.4 | 65.53 | 14539.54 |
Data Chart

8x RTX A4500 Benchmarks BERT for TensorFlow 2 FineTuning Training
Raw Data
Training Time Hours | Throughput sentences/sec | |
---|---|---|
FP16, Seq 128, BS1 | 2247.97 | 87.04 |
FP32, Seq 128, BS1 | 1859.29 | 107.77 |
FP16, Seq 128, BS2 | 996.29 | 228.01 |
FP32, Seq 128, BS2 | 994.02 | 228.55 |
FP16, Seq 128, BS4 | 617.1 | 449.09 |
FP32, Seq 128, BS4 | 618.85 | 449.5 |
FP16, Seq 128, BS8 | 392.02 | 807.31 |
FP32, Seq 128, BS8 | 391.99 | 809.54 |
FP16, Seq 384, BS1 | 1778.13 | 113.63 |
FP32, Seq 384, BS1 | 1781.29 | 113.46 |
FP16, Seq 384, BS2 | 1086.58 | 205.96 |
FP32, Seq 384, BS2 | 1081.24 | 207.12 |
FP16, Seq 384, BS4 | 742.6 | 346.99 |
FP32, Seq 384, BS4 | 743.12 | 346.3 |
FP16, Seq 384, BS8 | 520.45 | 521.61 |
FP32, Seq 384, BS8 | 519.85 | 522.77 |
Data Chart

NVIDIA RTX A4500 Series GPUs
GPU Features | NVIDIA RTX A4500 |
---|---|
GPU Memory | 20GB GDDR6 with error-correction code (ECC) |
Display Ports | 4x DisplayPort 1.4 |
Max Power Consumption | 2000 W |
Graphics Bus | PCI Express Gen 4 x 16 |
Form Factor | 4.4” (H) x 10.5” (L) Dual Slot |
Thermal | Active |
NVLink | 2-way Low-profile (2-slot and 3-slot bridges) |
VR Ready | Yes |
Additional GPU Benchmarks
Exxact Workstation System Specs
Nodes | 1 |
Processor / Count | 2x AMD EPYC 7552 |
Total Logical Cores | 48 |
Memory | DDR4 512GB |
Storage | NVMe 3.84TB |
OS | Ubuntu 18.04 |
CUDA Version | 11.2 |
BERT Dataset | squad v1 |
Additional GPU Benchmarks
- NVIDIA RTX 3080 Ti BERT Large Fine Tuning Benchmarks in TensorFlow
- NVIDIA RTX A4000 BERT Large Fine Tuning Benchmarks in TensorFlow
- NVIDIA RTX A5000 BERT Large Fine Tuning Benchmarks in TensorFlow
- NVIDIA A5000 Deep Learning Benchmarks for TensorFlow
- NVIDIA A30 Deep Learning Benchmarks for TensorFlow
- NVIDIA RTX A6000 Deep Learning Benchmarks for TensorFlow
- NVIDIA A100 Deep Learning Benchmarks for TensorFlow
Have any questions?
Contact Exxact Today

NVIDIA RTX A4500 BERT Large Fine Tuning Benchmarks in TensorFlow
Fine-tuning BERT Large on a GPU Workstation
For this post, we measured fine-tuning performance (training and inference) for the BERT implementation of TensorFlow on NVIDIA RTX A4500 GPUs. For testing we used an Exxact Valence Workstation fitted with 8x RTX A4500 GPUs with 20GB GPU memory per GPU.
Benchmark scripts we used for evaluation were the finetune_train_benchmark.sh and finetune_inference_benchmark.sh from NVIDIA NGC Repository BERT for TensorFlow. We made slight modifications to the training benchmark script to get the larger batch size numbers.
The script runs multiple tests on the SQuAD v1.1 dataset using batch sizes 1, 2, 4, and 8. Inferencing tests were conducted using a 1x GPU configurations on BERT Large. In addition, we ran all benchmarks using TensorFlow's XLA across the board.
Key Points and Observations
- Scenarios that are not typically used in real-world training, such as single GPU throughput are illustrated in the table below, and provided for reference as an indication of single chip throughput of the platform.
- For those interested in training BERT Large, a 2x RTX A4500 system may be a great choice to start with, giving the opportunity to add additional cards as budget/scaling needs increase.
- NOTE: In order to run these benchmarks, or be able to fine-tune BERT Large with 4x GPUs you'll need a system with at least 64GB RAM.
Interested in getting faster results?
Learn more about Exxact AI workstations to do NLP training on starting around $5,500
Exxact Workstation System Specs:
Nodes | 1 |
Processor / Count | 2x AMD EPYC 7552 |
Total Logical Cores | 48 |
Memory | DDR4 512GB |
Storage | NVMe 3.84TB |
OS | Ubuntu 18.04 |
CUDA Version | 11.4 |
BERT Dataset | squad v1 |
TensorFlow | 2.40 |
GPU Benchmark Overview
FP = Floating Point Precision, Seq = Sequence Length, BS = Batch Size
1x Quadro RTX A4500 BERT LARGE Inference Benchmark
Raw Data
Model | Sequence-Length | Batch-size | Precision | Total-Inference-Time | Throughput-Average(sent/sec) | Latency-Average(ms) | Latency-50%(ms) | Latency-90%(ms) | Latency-95%(ms) | iLatency-99%(ms) | Latency-100%(ms) |
base | 128 | 1 | fp16 | 25.98 | 186.13 | 8.94 | 5.56 | 6.34 | 6.46 | 6.8 | 6387.51 |
base | 128 | 1 | fp32 | 20.34 | 188.55 | 8.38 | 5.48 | 6.32 | 6.48 | 6.85 | 5501.44 |
base | 128 | 2 | fp16 | 20.47 | 364.38 | 8.84 | 5.67 | 6.2 | 6.32 | 6.78 | 5679.04 |
base | 128 | 2 | fp32 | 19.65 | 389.55 | 8.47 | 5.3 | 5.89 | 5.99 | 6.18 | 5654.98 |
base | 128 | 4 | fp16 | 26.39 | 695.64 | 11.55 | 5.79 | 6.24 | 6.36 | 6.9 | 5774.98 |
base | 128 | 4 | fp32 | 26.41 | 696.75 | 11.54 | 5.76 | 6.22 | 6.41 | 6.86 | 5767.47 |
base | 128 | 8 | fp16 | 34.2 | 964.18 | 13.57 | 8.27 | 8.72 | 8.94 | 9.52 | 5876.52 |
base | 128 | 8 | fp32 | 34.42 | 965.47 | 13.6 | 8.28 | 8.64 | 8.89 | 9.44 | 5899.06 |
base | 384 | 1 | fp16 | 17.03 | 177.1 | 11.46 | 5.65 | 6.11 | 6.25 | 6.7 | 6059.77 |
base | 384 | 1 | fp32 | 17.2 | 176.11 | 11.49 | 5.66 | 6.15 | 6.29 | 6.77 | 6051.43 |
base | 384 | 2 | fp16 | 25.64 | 239.89 | 19.3 | 8.31 | 8.84 | 9.13 | 9.84 | 6456.22 |
base | 384 | 2 | fp32 | 25.6 | 240.18 | 19.26 | 8.29 | 8.8 | 9.06 | 9.77 | 6447.57 |
base | 384 | 4 | fp16 | 31.28 | 305.32 | 24.29 | 13.09 | 13.62 | 13.85 | 14.97 | 6660.95 |
base | 384 | 4 | fp32 | 31.11 | 304.43 | 24.28 | 13.13 | 13.6 | 13.8 | 14.98 | 6639.34 |
base | 384 | 8 | fp16 | 43.89 | 358.82 | 34.29 | 22.27 | 22.79 | 23.1 | 24.51 | 7002.83 |
base | 384 | 8 | fp32 | 43.9 | 358.76 | 34.33 | 22.26 | 22.8 | 23.11 | 24.35 | 7031.13 |
large | 128 | 1 | fp16 | 47.42 | 103.62 | 15.96 | 9.42 | 10.95 | 11.16 | 11.67 | 11340.21 |
large | 128 | 1 | fp32 | 36.28 | 108.7 | 15.52 | 8.91 | 10.12 | 10.4 | 10.96 | 11350.15 |
large | 128 | 2 | fp16 | 38.82 | 180.4 | 17.77 | 11.09 | 11.91 | 12.05 | 12.44 | 11356.99 |
large | 128 | 2 | fp32 | 39.12 | 177.62 | 17.97 | 11.45 | 11.98 | 12.13 | 12.79 | 11389.43 |
large | 128 | 4 | fp16 | 58.85 | 253.71 | 27.84 | 15.94 | 16.48 | 16.65 | 17.52 | 11607 |
large | 128 | 4 | fp32 | 57.8 | 261.94 | 27.35 | 15.01 | 16.17 | 16.35 | 17.01 | 11606.73 |
large | 128 | 8 | fp16 | 78.49 | 348.55 | 34.15 | 22.89 | 23.82 | 24.05 | 24.71 | 11839.75 |
large | 128 | 8 | fp32 | 79.35 | 345.57 | 34.53 | 23.22 | 23.96 | 24.19 | 24.88 | 12047.05 |
large | 384 | 1 | fp16 | 34.49 | 77.09 | 24.99 | 12.98 | 13.76 | 13.89 | 14.59 | 12534.02 |
large | 384 | 1 | fp32 | 33.95 | 79.81 | 24.49 | 12.34 | 13.35 | 13.51 | 14.09 | 12472.18 |
large | 384 | 2 | fp16 | 54.4 | 97.77 | 44.04 | 20.34 | 21.21 | 21.38 | 22.2 | 13302.25 |
large | 384 | 2 | fp32 | 53.88 | 98.22 | 43.83 | 20.12 | 21.18 | 21.38 | 22.19 | 13331.07 |
large | 384 | 4 | fp16 | 68.97 | 118.22 | 57.65 | 33.95 | 34.44 | 34.62 | 35.31 | 13713.5 |
large | 384 | 4 | fp32 | 69.14 | 117.88 | 57.75 | 34.07 | 34.59 | 34.78 | 35.73 | 13693.18 |
large | 384 | 8 | fp16 | 104.65 | 126.65 | 88.76 | 63.22 | 63.87 | 64.17 | 65.32 | 14498.6 |
large | 384 | 8 | fp32 | 104.84 | 126.41 | 89 | 63.33 | 64.11 | 64.4 | 65.53 | 14539.54 |
Data Chart

8x RTX A4500 Benchmarks BERT for TensorFlow 2 FineTuning Training
Raw Data
Training Time Hours | Throughput sentences/sec | |
---|---|---|
FP16, Seq 128, BS1 | 2247.97 | 87.04 |
FP32, Seq 128, BS1 | 1859.29 | 107.77 |
FP16, Seq 128, BS2 | 996.29 | 228.01 |
FP32, Seq 128, BS2 | 994.02 | 228.55 |
FP16, Seq 128, BS4 | 617.1 | 449.09 |
FP32, Seq 128, BS4 | 618.85 | 449.5 |
FP16, Seq 128, BS8 | 392.02 | 807.31 |
FP32, Seq 128, BS8 | 391.99 | 809.54 |
FP16, Seq 384, BS1 | 1778.13 | 113.63 |
FP32, Seq 384, BS1 | 1781.29 | 113.46 |
FP16, Seq 384, BS2 | 1086.58 | 205.96 |
FP32, Seq 384, BS2 | 1081.24 | 207.12 |
FP16, Seq 384, BS4 | 742.6 | 346.99 |
FP32, Seq 384, BS4 | 743.12 | 346.3 |
FP16, Seq 384, BS8 | 520.45 | 521.61 |
FP32, Seq 384, BS8 | 519.85 | 522.77 |
Data Chart

NVIDIA RTX A4500 Series GPUs
GPU Features | NVIDIA RTX A4500 |
---|---|
GPU Memory | 20GB GDDR6 with error-correction code (ECC) |
Display Ports | 4x DisplayPort 1.4 |
Max Power Consumption | 2000 W |
Graphics Bus | PCI Express Gen 4 x 16 |
Form Factor | 4.4” (H) x 10.5” (L) Dual Slot |
Thermal | Active |
NVLink | 2-way Low-profile (2-slot and 3-slot bridges) |
VR Ready | Yes |
Additional GPU Benchmarks
Exxact Workstation System Specs
Nodes | 1 |
Processor / Count | 2x AMD EPYC 7552 |
Total Logical Cores | 48 |
Memory | DDR4 512GB |
Storage | NVMe 3.84TB |
OS | Ubuntu 18.04 |
CUDA Version | 11.2 |
BERT Dataset | squad v1 |
Additional GPU Benchmarks
- NVIDIA RTX 3080 Ti BERT Large Fine Tuning Benchmarks in TensorFlow
- NVIDIA RTX A4000 BERT Large Fine Tuning Benchmarks in TensorFlow
- NVIDIA RTX A5000 BERT Large Fine Tuning Benchmarks in TensorFlow
- NVIDIA A5000 Deep Learning Benchmarks for TensorFlow
- NVIDIA A30 Deep Learning Benchmarks for TensorFlow
- NVIDIA RTX A6000 Deep Learning Benchmarks for TensorFlow
- NVIDIA A100 Deep Learning Benchmarks for TensorFlow
Have any questions?
Contact Exxact Today