2x3090 hits 250W sweet spot for Qwen3.6-27B

// 47d agoBENCHMARK RESULT

2x3090 hits 250W sweet spot for Qwen3.6-27B

Benchmarks for Qwen3.6-27B on dual RTX 3090s identify 250W as the optimal power-to-performance balance for local inference. The testing highlights diminishing returns at higher power limits while maintaining peak throughput for long-context workloads.

// ANALYSIS

Local inference enthusiasts facing the 3090 "power wall" should prioritize 250W for efficiency, though 275W offers a measurable latency win for single-stream requests.

–Speculative decoding via Multi-Token Prediction (MTP) is the primary performance driver, delivering a massive 2x throughput boost for coding tasks.
–Tensor parallelism (TP=2) is mandatory to mitigate PCIe bandwidth bottlenecks on dual-GPU consumer setups, outperforming single-card configurations.
–The combination of FP8 KV caching and INT4 AutoRound quantization enables a massive 200K context window within the 48GB VRAM limit.
–Benchmark results confirm that LLM inference remains memory-bandwidth bound; aggressive 350W stock power limits generate excess heat with negligible token gains.

// TAGS

qwen3.6-27bllmvllmgpubenchmarkself-hostedinference

DISCOVERED

47d ago

2026-04-28

PUBLISHED

47d ago

2026-04-28

RELEVANCE

8/ 10

AUTHOR

JC1DA

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

OPEN SOURCE3m ago

ByteDance open-sources Bernini video framework

ByteDance has open-sourced Bernini, a unified framework that combines an MLLM-based semantic planner with a DiT-based renderer for consistent video generation and editing. By decoupling structural semantic planning from pixel generation, the system enables highly controllable, reference-guided edits while mitigating frame flickering and background drift.

OPEN SOURCE3m ago

MiniMax open-sources MSA sparse attention kernel

MiniMax has open-sourced MiniMax Sparse Attention (MSA), a blockwise sparse attention kernel designed to handle million-token context windows efficiently. By combining a two-branch architecture with a co-designed GPU execution path, MSA reduces per-token compute by 28.4×, achieving a 14.2× prefill speedup and 7.6× decoding speedup on H800 GPUs.

OPEN SOURCE3m ago

ghostty-blackhole is a custom Ghostty terminal shader that gravitationally distorts text and visualizes Claude Code context usage as a live, ray-traced black hole.

ghostty-blackhole is a custom terminal shader for the Ghostty terminal emulator that visualizes LLM token usage for Claude Code. As the context window fills up, a simulated black hole on the screen grows, using ray-traced gravitational lensing, an accretion disk, and a photon ring to warp and distort the terminal text. This visual cues developers before their context window overflows, and it displays key metadata such as fill percentage, current model, and session costs.