CONCEPT Cited by 1 source
Edge inference¶
Definition¶
Running ML/AI model inference on hardware physically located at or near the point of use, rather than in a centralized cloud data center. In the generative AI context, this means serving LLM inference on edge devices with constrained GPU memory (typically 16-96 GB VRAM), requiring deliberate trade-offs between model capability, hardware constraints, and operational complexity.
Hardware-driven deployment strategies¶
When deploying models to multi-GPU edge devices, two fundamental strategies exist:
-
Model Replication — a full model copy on each GPU, supporting N concurrent requests without queuing. Maximizes throughput at the cost of per-GPU memory consumed by weights (e.g., 81% for a 13 GB model on a 16 GiB T4).
-
Tensor Parallelism — the model is sharded across all GPUs, leaving remaining memory for KV cache. Supports full context windows for longer interactions but limits concurrent request capacity.
The choice is workload-profile-driven: replication suits high-concurrency short-query environments; tensor parallelism suits fewer concurrent users with complex, long-form interactions.
Key constraints¶
- VRAM ceiling: Model must fit within device memory after quantization (GGUF/GPTQ/AWQ formats)
- No elastic scaling: Fixed hardware means capacity planning is static
- Thermal/power: Industrial environments may have power/cooling constraints
- Update velocity: Models updated via deployment cycles, not hot-swapped
Seen in¶
- sources/2026-07-22-aws-architecting-offline-first-generative-ai-applications-for-edge-deployments — gpt-oss-20b on g4dn.12xlarge (4× T4, 16 GiB each): 13 GB/GPU replicated (4 concurrent) or 3.2 GB/GPU tensor-parallel (full 128K context); Ollama as the serving runtime.