Need a freelancer for running AI benchmarks on CPU & GPU

During the next few years AI will be heavily deployed in the Cloud, the Edge and the Fog. NeuReality is an innovative system and semiconductor startup focusing on the future challenges of this growing demand by designing the next evolution of AI-as-a-Service Infrastructure.

To achieve best AI inference performance, AI developers convert Deep Learning models to target HW (CPU, GPU, ASIC).

Neureality defines AI pipeline for inference as a graph of pre-processed functions + pretrained DL model + (possibly) post-processing functions.
For example and AI pipeline for object classification: JPEG decoding -> resize -> normalization -> RenNet50. ( 4 pipeline stages).

Project Scope:
AI pipeline benchmarks on CPU and GPU for:
• 4 AI pipelines for: GNMT (machine translation), BERT (NLP), RS50 (image processing), and DLRM (recommendation systems). All models are publicly available.
• Based on Python + TensorFlow/Pytorch DL/ML packages.
For each pipeline:
• Theoretical calculations of compute operations (e.g., math operations) and amount of memory accesses (total amount of read/write bytes ) per stage.
• Performance measurements and analysis per pipeline stage on CPU (by default Xeon) and GPU (by default Nvidia T4) wrt to i) total execution time ii) amount of memory read/write and iii) CPU/GPU utilization

Deliverables:
Full codebase (Git) + .xls with performance measurements.

The person we are looking for should have:
• Extensive proven experience working with AI frameworks (TensorFlow, PyTorch etc.)
• Extensive proven experience building and deploying AI models on various platforms and architectures (cloud, edge, GPU, TPU)
• Hands-on Python coding. C/C++ experience is a big advantage.
• Advantage: Experience working with AI models formats (SavedModel, ONNX, Pytorch).
• Advantage: Experience with AI cloud services (AWS SageMaker, Google CloudML, etc.)
• Problem solver with good communication skills.