Press Release

ThirdAI’s BOLT Deep Learning Engine Achieves Exceptional Performance on AMD EPYC™ Processors, Surpassing State-of-the-Art GPU Models

June 13, 2023 (Houston, TX) – ThirdAI, a pioneering startup dedicated to democratizing
machine learning capabilities, has successfully completed benchmarking of its BOLT deep
learning engine on AMD EPYC™ 9004 Series processors as well as the latest 9004 Series
AMD EPYC processors based on the “Zen 4c” architecture. These benchmarking experiments
showcase BOLT’s exceptional performance acceleration on cutting-edge AMD EPYC CPUs
versus current state-of-the-art models trained on NVIDIA GPUs.

In a series of rigorous tests, ThirdAI BOLT engine’s acceleration on AMD EPYC CPU-based
machines outperformed well-established baselines across diverse machine learning tasks on
NVIDIA A100s GPUs.

Graph node classification (a technique to detect fraud or scams within social media): using
the open-source Yelp-Chi fraud detection dataset, BOLT surpassed accuracy measurements
against well-established graph neural network baselines GCN (Graph Convolutional Networks)
and GAT (Graph Attention Networks).

Trained on Yelp-Chi fraud dataset:
BOLT: 91.1% accuracy with training time of 10s on AMD EPYC 9754 CPU
GCN: 63.6% accuracy with training time of 150s
*training time 4.51s on NVIDIA A100 GPU
GAT: 81.4% accuracy with training time of 200s
*training time 22.5s on NVIDIA A100 GPU

Sequence 2 Sequence (models trained for purposes such as language translation): evaluation
on Multi-30K Translation Dataset from English to German, BOLT far surpassed translation
accuracy against standard LSTM (Long Short-Term Memory) GPU-trained model.

Trained on Multi-30K Translation Dataset from English to German:
BOLT: 39% accuracy with inference of 10ms (AMD EPYC 9754 CPU)
LSTM Seq2Seq: 20.31% accuracy with inference 29.3ms (AMD EPYC 9654 CPU)
*LSTM Seq2Seq: 20.31% accuracy with inference of 12ms (NVIDIA A100 GPU)

Text Classification (technique for purposes such as sentiment analysis or intent prediction):
evaluating ThirdAI BOLT against state-of-the-art pre-trained GPU-based model RoBERTa,
BOLT achieved similar accuracy on two representative datasets, but at a fraction of the training
time. This represents nearly a 200x speedup over GPUs.

Trained on Yelp Polarity and Amazon Polarity:
BOLT (AMD EPYC 9654 CPU): 92.3% accuracy (not pre-trained) with 130s total training
time
RoBERTa (AMD EPYC 9654 CPU): 94.5% accuracy (pre-trained/tuned) with 9.1 hrs
training time
*RoBERTa (NVIDIA A100 GPU): 1.77 hrs total training time

“These benchmarking experiments underscore the remarkable performance and acceleration
delivered by ThirdAI’s BOLT engine on the latest 4 th Gen AMD EPYC processors. With the
power of ThirdAI’s dynamic sparsity and AMD EPYC high performance cores and memory, we
can deliver the fastest AI acceleration in the industry,” says Anshumali Shrivastava, CEO of ThirdAI.

“The results reinforce ThirdAI’s commitment to democratizing machine learning capabilities and
enabling cost-effective training and deployment of large language models.”

About ThirdAI:
ThirdAI is on a mission to make sophisticated large language models (LLMs) and other cutting-
edge AI technologies accessible for everyone. Our goal is to build customized, private AI that is
trained on commodity hardware with ultra-low latency inference for every organization. ThirdAI’s
innovative technology is the result of 10 years of research and development in finding
fundamental ways to make deep learning more efficient. ThirdAI does not require GPUs, TPUs,
or custom ASIC to build its AI solutions. Our technology has applications in search,
recommendations, chatbots, sentiment analysis, and more.

For more information visit https://www.thirdai.com/
AMD, the AMD arrow logo, EPYC and combinations thereof are trademarks of Advanced Micro
Devices, Inc.