Unverified50% confidenceBenchmarkExact time
In a toy simulation with a 1B-class model and 16K to 64K token prompts, PD disaggregation reduced p99 inter-token latency from 5.0ms to 1.0ms and the worst-case stall from 931ms to 1ms.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/19/2026
First Seen
Valid until: 1/17/2027
Sources
Related Entities
Related Claims
UnverifiedOn the same machine and model, the decoding speed difference between low power mode and high power mode is 1.4 to 3.8 times; for example, running Qwen 3.5 35B on Ollama gives 51.9 tokens/s in low power and 107.1 tokens/s in high power.67% similarUnverifiedAt a scale of 1 million vectors, PRISM achieves 99% recall and 15 milliseconds median latency on a 2GB c8g.medium; HNSW on a 4GB c8g.large is 392 milliseconds, with building taking about 21.7 hours and exhibiting memory thrashing.67% similarUnverifiedThrough pure CPU inference and the 4-bit GGUF quantization format, llama.cpp enables a 7B-parameter model to achieve a generation speed of 20–40 tokens per second on an ordinary server without a GPU.67% similarUnverifiedThe Q8 main model occupies 162 GB, preventing the drafter from fitting into unified memory, rendering DSpark acceleration ineffective; after moving three layers of MoE experts back to the CPU, output only improved from 22.87 to 23.49 tokens/second (2.7% faster), while prefill dropped from 95.38 to 76.65.67% similarUnverifiedIn the experiment, the two size points of 2.8KB and 4.1KB were executed 8 times in total, with empty-output rates of 0% each, average output tokens of 9,096 and 7,913 respectively, and convention hit rates of 4/4 each.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1070127API
curl https://kongchang.com/api/v1/knowledge/claims/1070127MCP
get_claim(id=1070127)