Unverified50% confidenceFactExact time
GDS establishes a direct storage-to-GPU channel via direct memory access (DMA), making the data path shorter and more predictable.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedGPU Direct Storage（GDS）通过建立NVMe SSD与GPU显存之间的直接DMA通道，绕过CPU内存，在实际测试中可将大文件加载带宽提升2-5倍79% similarVerifiedThe LLM decoding process is memory bound, with the GPU spending most of its time waiting for data transfers during token-by-token generation.70% similarUnverifiedGDS's benefits are most evident in scenarios such as large single-read data blocks, datasets far exceeding host memory capacity, the CPU concurrently handling other intensive tasks, and multiple GPUs loading the same storage pool in parallel.68% similarUnverifiedIn the GDS scenario, storage devices can write directly into GPU memory (BAR space) via point-to-point (P2P) transactions on the PCIe bus, bypassing system memory.68% similarUnverifiedWhen using GDS, the target buffer must reside in GPU memory and meet alignment requirements (typically 4KB or larger), otherwise it falls back to a compatibility path.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1116814API
curl https://kongchang.com/api/v1/knowledge/claims/1116814MCP
get_claim(id=1116814)