Unverified50% confidenceSolutionExact time

After all 31 layers of the 26B model are offloaded to the GPU, ollama ps still reports 14%/86% CPU/GPU, which originates from mmap-mapped CPU-side address space tensor buffers rather than actual CPU computation.

1

Sources

50%

Confidence

Medium-term (~90 days)

Relevance

8/16/2026

First Seen

Valid until: 11/14/2026

Sources

Running a 26B LLM on AMD Integrated Graphics: Full GPU Inference and mmap Optimization Guide for Unified Memory Architecture

Related Claims

UnverifiedGPU上下文切换需保存完整GPU状态（包括寄存器文件、共享内存和L1缓存），数据量可达数百MB，而CPU上下文切换仅需保存几KB72% similar UnverifiedCPU offload places some Transformer layers in system main memory to be computed by the CPU, with data transferred over the PCIe bus (PCIe 4.0 x16 theoretical bandwidth is approximately 32 GB/s), while GPU memory bandwidth is typically in the range of hundreds to over a thousand GB/s, making the bandwidth gap the fundamental cause of the performance bottleneck.71% similar UnverifiedGPU每个SM的共享内存通常仅有48-96KB68% similar UnverifiedThe KV cache for a 128K context window of a GPT-4-level model can occupy tens of gigabytes of VRAM67% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/756814

API

curl https://kongchang.com/api/v1/knowledge/claims/756814

MCP

get_claim(id=756814)