Skip to content

Wire Briefings Search

Fetching from the wire…

01

02

03

04

05

06

07

08

Wire Briefings Search

Entity trail

NetraRuntime

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

1

Findings

1

Edges

1

Sources

1

Corpus findings

2026-08-26 / reddit-researcherAn Open-Source Kernel Stack Got Qwen3.6-35B-A3B to 78,498 Output Tokens Per Second on 8 AMD MI350X, 2.16x vLLMNetraRuntime published kernel-level optimizations for Qwen3.6-35B-A3B on AMD MI350X, reporting 11,161 output tok/s on one GPU and 78,498 tok/s mean (81,331 peak) across eight, which they measure at 2.16x vLLM throughput on the same 8-GPU benchmark. The kernels are open source at github.com/NetraRuntime/netra-kernel with a write-up of the method. Their closing note is the useful part for anyone attempting the same: once the kernels got fast enough, the bottleneck moved into scheduling, graph coverage, recurrent state, routing and HTTP serialization.

Graph relationships

BENCHMARKED AGAINST
vLLM
NetraRuntime kernel optimizations achieved 78,498 output tokens per second, 2.16x vLLM throughput.
Source finding

Source trail

published kernel-level optimizations

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues