AMD has 'supercharged' its software development by using AI agents. These agents run in automated loops, constantly analyzing and optimizing customer models for AMD's hardware. This turns a slow, manual process into a scalable, nonstop operation, dramatically improving out-of-the-box performance for developers.

A "roofline analysis" reveals that LLM performance is limited by the slower of two factors: the time it takes to fetch model parameters from memory (memory-bound) or the time it takes to perform matrix multiplications (compute-bound). Optimizing performance requires identifying and addressing the correct bottleneck.

For long queries, Baseten first checks for cached inputs. It then uses disaggregated GPUs—one set for pre-fill (processing input) and another for decode (generating tokens), often with a speculative decoder to accelerate output.

OpenAI achieved a major reduction in the cost of running its models through purely software and algorithmic improvements, such as quantization and smarter caching. This demonstrates that efficiency innovation can be as impactful as acquiring more hardware, suggesting a path to overcoming compute bottlenecks without relying solely on expensive chips.

Recent tests on NVIDIA B200 GPUs show that open-source models like China's GLM 5.2 can match or exceed the performance of proprietary models for tasks like coding. This performance threatens the moats of large, closed AI labs.

Model architecture decisions directly impact inference performance. AI company Zyphra pre-selects target hardware and then chooses model parameters—such as a hidden dimension with many powers of two—to align with how GPUs split up workloads, maximizing efficiency from day one.

LLMs can both generate code analysis tools (measuring metrics like cognitive complexity) and then act on those results. This creates a powerful, objective feedback loop where you can instruct an LLM to refactor code specifically to improve a quantifiable metric, then validate the improvement afterward.

The gap between a basic and a highly optimized inference setup is massive. Stacking techniques like quantization, custom speculative decoders, and KV-aware routing can yield performance improvements of 4x to 10x over a standard baseline for the same model and hardware.

The popular PyTorch Profiler only shows the 'tip of the iceberg.' To achieve meaningful performance gains, engineers must move beyond it and analyze 50-60 low-level GPU metrics related to streaming multiprocessors, instruction pipelines, and specialized function units. Most of the PyTorch community stops too early.