“Object storage is cheap — until everyone asks for a different cache key.”
What happened: Object storage is cheap — until everyone asks for a different cache key.
This Neural Toon is based on the story “ObjectCache: Layerwise Object-Storage Retrieval for KV Cache Reuse”.
Read our summary of the story · original source
Panel 1 – Alex: We swapped GPU RAM for S3. Genius!
Panel 1 – Priya: And latency? You told no one.
Panel 2 – Alex: 5.6% overhead for 64K contexts. Told you.
Panel 2 – Priya: At 4K, it's 56 to 75ms. Not zero.
Panel 3 – Alex: Saves millions on DRAM. Cheaper inference.
Panel 3 – Priya: Under bandwidth caps, TTFT drops 1.8x.
Panel 4 – Alex: We'll scale. ObjectCache handles it.
Panel 4 – Priya: What happens with 10,000 unique prefixes?
More Neural Toons · Neural Snaps · The Neural Feed
One email a day, no spam, unsubscribe anytime.