3FS
An open-source distributed file system from DeepSeek for high-performance storage, supporting scenarios such as large-scale KV caching, with capacity reaching the 100PB level.
Timeline (last 90 days)
A single DeepSeek online cluster processes about 155 trillion tokens per day, translating to revenue of about 38 US dollars per card per hour; Decode uses EP64 to EP128, Pre-fill uses only EP2 to EP8, KV Cache is stored on 3FS with 10+2 erasure coding and retained for 3 to 7 days.
In its deployment Q&A, DeepSeek states: the cache uses 3FS storage with 10+2 erasure coding and about 100 PiB of solid-state drives; agent traffic is retained for 3 days and chat history for 7 days; the minimum unit for efficient deployment is 64 GPUs for prefill and 64 GPUs for decode.
SmallPond is a lightweight data processing framework built on top of 3FS.
DeepSeek's production solution uses the 3FS storage system, placing the main type of KV cache in independent shared storage so that other nodes can also read existing caches and avoid waiting indefinitely for the original machine.
DeepSeek uses its self-developed distributed file system 3FS to build an independent storage cluster of about 100PB of solid-state drives.
DSEC uses on-demand file loading to read from DeepSeek's proprietary storage system 3FS; launching 8,192 sandboxes at once takes only 35 minutes, with disk writes reduced by approximately 57%.
All Facts (6)
DeepSeek's production solution uses the 3FS storage system, placing the main type of KV cache in independent shared storage so that other nodes can also read existing caches and avoid waiting indefinitely for the original machine.
60%UnverifiedA single DeepSeek online cluster processes about 155 trillion tokens per day, translating to revenue of about 38 US dollars per card per hour; Decode uses EP64 to EP128, Pre-fill uses only EP2 to EP8, KV Cache is stored on 3FS with 10+2 erasure coding and retained for 3 to 7 days.
50%UnverifiedIn its deployment Q&A, DeepSeek states: the cache uses 3FS storage with 10+2 erasure coding and about 100 PiB of solid-state drives; agent traffic is retained for 3 days and chat history for 7 days; the minimum unit for efficient deployment is 64 GPUs for prefill and 64 GPUs for decode.
50%UnverifiedSmallPond is a lightweight data processing framework built on top of 3FS.
50%UnverifiedDeepSeek uses its self-developed distributed file system 3FS to build an independent storage cluster of about 100PB of solid-state drives.
50%UnverifiedDSEC uses on-demand file loading to read from DeepSeek's proprietary storage system 3FS; launching 8,192 sandboxes at once takes only 35 minutes, with disk writes reduced by approximately 57%.
50%