Stephen J. Tarsa

Robert Porter, Adam Diehl, Benjamin Pastel et al.

We introduce LLMD, a large language model designed to analyze a patient's medical history based on their medical records. Along with domain knowledge, LLMD is trained on a large corpus of records collected over time and across facilities, as well as tasks and labels that make nuanced connections among them. This approach is critical to an accurate picture of patient health, and has distinctive advantages over models trained on knowledge alone, unlabeled records, structured EHR data, or records from a single health system. The recipe for LLMD continues pretraining a foundational model on both domain knowledge and the contents of millions of records. These span an average of 10 years of care and as many as 140 care sites per patient. LLMD is then instruction fine-tuned on structuring and abstraction tasks. The former jointly identify and normalize document metadata, provenance information, clinical named-entities, and ontology mappings, while the latter roll these into higher-level representations, such a continuous era of time a patient was on a medication. LLMD is deployed within a layered validation system that includes continual random audits and review by experts, e.g. based on uncertainty, disease-specific rules, or use-case. LLMD exhibits large gains over both more-powerful generalized models and domain-specific models. On medical knowledge benchmarks, LLMD-8B achieves state of the art accuracy on PubMedQA text responses, besting orders-of-magnitude larger models. On production tasks, we show that LLMD significantly outperforms all other models evaluated, and among alternatives, large general purpose LLMs like GPT-4o are more accurate than models emphasizing medical knowledge. We find strong evidence that accuracy on today's medical benchmarks is not the most significant factor when analyzing real-world patient data, an insight with implications for future medical LLMs.'

Stephen J Tarsa, Chit-Kwan Lin, Gokce Keskin et al.

CPU branch prediction has hit a wall--existing techniques achieve near-perfect accuracy on 99% of static branches, and yet the mispredictions that remain hide major performance gains. In a companion report, we show that a primary source of mispredictions is a handful of systematically hard-to-predict branches (H2Ps), e.g. just 10 static instructions per SimPoint phase in SPECint 2017. The lost opportunity posed by these mispredictions is significant to the CPU: 14.0% in instructions-per-cycle (IPC) on Intel SkyLake and 37.4% IPC when the pipeline is scaled four-fold, on par with gains from process technology. However, up to 80% of this upside is unreachable by the best known branch predictors, even when afforded exponentially more resources. New approaches are needed, and machine learning (ML) provides a palette of powerful predictors. A growing body of work has shown that ML models are deployable within the microarchitecture to optimize hardware at runtime, and are one way to customize CPUs post-silicon by training to customer applications. We develop this scenario for branch prediction using convolutional neural networks (CNNs) to boost accuracy for H2Ps. Step-by-step, we (1) map CNNs to the global history data used by existing branch predictors; (2) show how CNNs improve H2P prediction in SPEC 2017; (3) adapt 2-bit CNN inference to the constraints of current branch prediction units; and (4) establish that CNN helper predictors are reusable across application executions on different inputs, enabling us to amortize offline training and deploy ML pattern matching to improve IPC.