Ming-Hung Chen

Talia Gershon, Seetharami Seelam, Brian Belgodere et al.

AI Infrastructure plays a key role in the speed and cost-competitiveness of developing and deploying advanced AI models. The current demand for powerful AI infrastructure for model training is driven by the emergence of generative AI and foundational models, where on occasion thousands of GPUs must cooperate on a single training job for the model to be trained in a reasonable time. Delivering efficient and high-performing AI training requires an end-to-end solution that combines hardware, software and holistic telemetry to cater for multiple types of AI workloads. In this report, we describe IBM's hybrid cloud infrastructure that powers our generative AI model development. This infrastructure includes (1) Vela: an AI-optimized supercomputing capability directly integrated into the IBM Cloud, delivering scalable, dynamic, multi-tenant and geographically distributed infrastructure for large-scale model training and other AI workflow steps and (2) Blue Vela: a large-scale, purpose-built, on-premises hosting environment that is optimized to support our largest and most ambitious AI model training tasks. Vela provides IBM with the dual benefit of high performance for internal use along with the flexibility to adapt to an evolving commercial landscape. Blue Vela provides us with the benefits of rapid development of our largest and most ambitious models, as well as future-proofing against the evolving model landscape in the industry. Taken together, they provide IBM with the ability to rapidly innovate in the development of both AI models and commercial offerings.

Rajesh Bordawekar, Bulent Abali, Ming-Hung Chen

In a large class of deep learning models, including vector embedding models such as word and database embeddings, we observe that floating point exponent values cluster around a few unique values, permitting entropy based data compression. Entropy coding compresses fixed-length values with variable-length codes, encoding most probable values with fewer bits. We propose the EFloat compressed floating point number format that uses a variable field boundary between the exponent and significand fields. EFloat uses entropy coding on exponent values and signs to minimize the average width of the exponent and sign fields, while preserving the original FP32 exponent range unchanged. Saved bits become part of the significand field increasing the EFloat numeric precision by 4.3 bits on average compared to other reduced-precision floating point formats. EFloat makes 8-bit and even smaller floats practical without sacrificing the exponent range of a 32-bit floating point representation. We currently use the EFloat format for saving memory capacity and bandwidth consumption of large vector embedding models such as those used for database embeddings. Using the RMS error as metric, we demonstrate that EFloat provides higher accuracy than other floating point formats with equal bit budget. The EF12 format with 12-bit budget has less end-to-end application error than the 16-bit BFloat16. EF16 with 16-bit budget has an RMS-error 17 to 35 times less than BF16 RMS-error for a diverse set of embedding models. When making similarity and dissimilarity queries, using the NDCG ranking metric, EFloat matches the result quality of prior floating point representations with larger bit budgets.