Video · IBM Technology ·
Why AI Models Run Faster on Some Chips: The KV Cache Effect
In short
The video explains how KV cache lets LLMs reuse previously computed keys and values during token generation, while creating a capacity-versus-locality memory challenge. It compares GPU, TPU, and IBM Spire accelerator memory paths and orchestration, showing why data movement can determine inference efficiency and throughput as much as raw compute.
Key points
- KV cache stores attention keys and values from earlier tokens so each new token does not recompute the full history.
- Large memories provide capacity, while smaller fast memories provide the locality needed for efficient attention computation.
- GPUs typically use HBM with runtime- and kernel-managed allocation and access patterns.
- TPUs use compiler-directed tiling and pipelining between HBM and on-chip vector memory.
- IBM Spire uses 128 GB LPDDR5 per card, 32 cores, and a 2 MB SRAM scratchpad per core.
- Spire's compiler explicitly schedules scratchpad loads and stores instead of relying on automatic hardware caching.
- Accelerator comparisons should examine how memory state moves, not only how much math the chip can perform.
Written by an AI model from the video’s captions and description.
The same AI model can run differently across chips. Kaoutar El Maghraoui explains how GPUs, TPUs, and AI accelerators store and move KV cache during LLM inference. Learn why memory architecture and data movement can shape AI performance.



