SYSTEM Cited by 2 sources

Gumby (Dropbox 7th-gen flexible GPU platform)

Definition

Gumby is Dropbox's 7th-generation flexible GPU platform — a Crush compute server with PCIe GPU slots added on top. Positioned for mixed inference and embedding workloads where the right accelerator size varies widely across tenants. Rolled out 2025 alongside systems/godzilla (the dense multi-GPU tier).

Form-factor envelope

• TDP range: 75 W → 600 W
• Form factors: HHHL (half-height half-length) and FHFL (full-height full-length)

The wide envelope is the whole point — Gumby's name "signals flexibility" (Dropbox's framing). Rather than one GPU SKU per server, Gumby can host anything from a low-end inference card to a high-TDP training-class GPU in a CPU-dominant chassis.

Target workloads

• Intelligent previews (document understanding) for Dash
• Video processing / transcoding
• Embedding generation
• Service-side ML enhancements — see the linked bringing-AI- powered-answers-and-summaries-to-file-previews-on-the-web post.
• Lightweight inference generally — anywhere a CPU is insufficient but a full multi-GPU dense server is overkill.

Relationship to Godzilla

• Gumby — one or few GPUs, variable TDP, CPU-dominant chassis, inference-leaning.
• Godzilla — up to 8 interconnected GPUs, GPU-dominant chassis, training-leaning.

Together they form Dropbox's 7th-gen accelerator tier; neither one alone serves the full range of workloads that Dash and related ML features demand.

Software axis — low-bit inference

Gumby's inference workloads are the substrate for Dropbox's low-bit inference strategy: multiple quantization formats (pre-MXFP AWQ/HQQ- style A16W4 or A8W8; MXFP / NVFP on Tensor Core-capable silicon) run across Gumby depending on whether each workload is memory-bound vs compute-bound. No single strategy fits all Dash workloads — conversational AI, multimodal search, document understanding, and speech processing each hit different points on the latency vs throughput frontier (Source: sources/2026-02-12-dropbox-how-low-bit-inference-enables-efficient-ai).

Seen in

• sources/2025-08-08-dropbox-seventh-generation-server-hardware — 7th-gen hardware introduction; Gumby/Godzilla split between flexibility and density.
• sources/2026-02-12-dropbox-how-low-bit-inference-enables-efficient-ai — landscape survey of the quantization stack running on top of Gumby-class inference hardware for Dash workloads.