Exabits has demonstrated its capability to train large language models (LLMs), partnering with MyShell to dramatically reduce training costs from billions to under $100,000.
JetMoE-8B is trained at less than a $0.1 million cost but outperforms LLaMA2-7B from Meta AI (multi-billion dollar compute cost)
MyShell: Achieving LlaMA2 performance with the $100,000 JetMoE model, inspired by the sparse activation architecture of ModuleFormer, signifies a remarkable milestone in machine learning. The JetMoE-8B, with its 8 billion parameters and sophisticated structure of 24 blocks, each housing two MoE layers (Attention Head Mixture and MLP Experts Mixture), showcases advanced efficiency and computational intelligence. Each layer’s selective activation of 2 out of 8 experts per input token demonstrates a refined utilization of the Sparse Mixture of Experts (SMoE) framework, enhancing the model’s responsiveness and resource management.
The efficiency of JetMoE-8B, with its 2.2 billion activation parameters, significantly lowered training costs while delivering robust performance. The model’s effectiveness is illustrated in the subsequent figure: JetMoE-8B achieved state-of-the-art results in five categories on eight evaluation benchmarks, outperforming competitors like LLaMA-13B, LLaMA2-7B, and DeepseekMoE-16B.
On the MT-Bench benchmark, JetMoE-8B scored 6.681, surpassing models with larger capacities, such as LLaMA2 and Vicuna, which possess 13 billion parameters.
But what superpowers this architectural sophistication is Exabits’ contribution of an accelerated and stabilized cluster of 12 H100 GPU nodes (96 GPUs). Exabits’ platform played a pivotal role in powering the JetMoE model, ensuring stable, ultra-available and robust performance at a fraction of the cost of “big compute.” This synergy between JetMoE’s innovative design and Exabits’ cutting-edge GPU technology not only exemplifies a leap in machine learning capabilities but also highlights the effectiveness of combining advanced model architectures with Exabits’ cloud compute infrastructure.
Breaking the Myth: Decentralized GPU Platform for LLM Training
Exabits has disproved the skepticism that decentralized GPU platforms are unsuitable for LLM training. With a sophisticated technical stack, efficient middleware, and a robust supply chain of computational resources, Exabits has demonstrated that LLM training and inference are not only possible but also efficient and deeply cost-effective on such a platform.
Exabits, a decentralized cloud compute platform, overcomes the limitations of standard decentralized platforms by serving as the infrastructure base layer of AI computing and offering a full-stack solution. It does this by aggregating, accelerating, and stabilizing consumer-grade GPUs to match enterprise-grade GPU performance to almost parity. This approach taps into a vast, yet largely idle reserve of consumer GPUs, easing the GPU shortage crisis. Also, Exabits’ extensive experience in the data center sector provides unique access to coveted enterprise-grade H100 and A100 GPUs, and soon the B200s, further advancing the democratization of AI development. Partnerships with major projects in decentralized cloud compute have helped Exabits to seed and establish a widespread, interconnected decentralized compute network. This super-network has the potential to stand against the giants of centralized, traditional cloud compute, making AI accessible to anyone who wants to build in the space.
The Future of LLM Training with Exabits
Exabits is not just a technological platform; it is a beacon for the future of LLM training, embodying affordability, accessibility, and environmental consciousness. The success of JetMoE-8B underlines the feasibility of this platform in executing high-end model training, paving the way for more sustainable and inclusive advancements in AI research and development.
In conclusion, Exabits stands as a revolutionary force in the AI domain, challenging big compute and proving that cloud compute platforms in the web3 space can indeed support real LLM training efficiently and cost-effectively. This not only opens up new avenues for AI research and application but also sets a new standard in the computational economy, heralding a new era of innovation and collaboration in the field of web3 and artificial intelligence.
Disclaimer: This press release may contain certain forward-looking statements. Forward-looking statements describe expectations, plans, outcomes, or strategies for the future (including product offerings, regulatory plans, and business plans) and are subject to change without prior notice. Please be advised that such statements are influenced by various uncertainties, which may result in future circumstances, events, or outcomes differing from those predicted in the forward-looking statements.
Media Contact
Company Name: ExaBITs Network LTD.
Contact Person: Roy Evans
Email: Send Email
Phone: +1 650 642 8104
Country: Singapore
Website: https://www.exabits.ai