ONNX model very slow for Sentence transformer model when max_seq_len of tokens = 512

I am using a Roberta based Sentence transformer model for a chatbot. For inference I have exported the model to ONNX. When the max_length of the tokens <= 128 , the perfomance of the model is fine. But once the token length increases to 512, the ONNX model's output decreases drastically.
Even if I quantize the base model and then export to ONNX, there is no change in preformance for seq length of 512 tokens

Specs

• OS Platform and Distribution (Mac OS Monterey)
• ONNX Runtime installed using pip
• ONNX Runtime version: 1.8.1
• Python version: 3.6

I am running the model on CPU.

The code which was used to export from Sentence transformer model to ONNX model is given below

torch.onnx.export(model, (inputs['input_ids']), "roberta_model.onnx", opset_version=11, do_constant_folding=True, input_names=['input_ids'], output_names=['start', 'end'], dynamic_axes={'input_ids': [1]})

And this is how I am performing inference

options = onnxruntime.SessionOptions() options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL options.intra_op_num_threads = 1 self.onnxmodel = onnxruntime.InferenceSession(model_path, options)

Even if I remove options from the above snippet. The results seem to be similar.

The avg response time for inferencing on 512 tokens is around 1.3-1.4 seconds on an instance with 8vCPUs.

Can something be done about this?