AI BriefingDistributed AI Inference Speed Improved 15x
Key point
The Shard project dramatically improved distributed inference speed for large-scale models by leveraging speculative decoding and other techniques in a WAN environment.
Details
The Shard project, released by researcher leyten, successfully achieved a speed of approximately 30 tokens per second (tok/s) by distributing the GLM-5.2 (744B) model across 6 RTX Pro 6000 GPUs in a WAN (wide area network) environment.
Compared to the existing Petals (2022) approach, which recorded 1-2 tok/s on smaller models, this represents an approximately 15-20x improvement. The key technical achievements are as follows:
- Speculative decoding over WAN: A small draft model proposes multiple tokens, and the distributed large model verifies them in a single network round trip, minimizing latency.
- Ring pipelining with direct return: The final node sends results directly to the coordinator without passing through all stages, increasing efficiency.
- CUDA-graphed draft model: The draft model is precompiled as a CUDA graph, achieving a 3.8-5.3x speed improvement.
Shard is planned to serve as the underlying infrastructure for the c0mpute.ai network, where anyone can contribute GPUs and receive rewards in USDC. All execution results are verifiable through public receipts that include GPU UUID, IP address, latency, and more.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
vLLM Publishes Performance Benchmarks for 5 Speculative Decoding Methods on AMD GPUs
Optimizing Gemma 4 Serving with vLLM on Amazon EKS [Part 2: Throughput and Latency SLOs on a Single GPU]Optimizing Gemma 4 Serving with vLLM on Amazon EKS: Cold Start Reduced from 428s to 226s