High-performance NVIDIA Blackwell GPU with DFlash technology showcasing a 15x boost in AI inference speed for accelerated mac

Editorial illustration for DFlash speculative decoding boosts NVIDIA Blackwell inference up to 15×

DFlash speculative decoding boosts NVIDIA Blackwell...

DFlash speculative decoding boosts NVIDIA Blackwell inference up to 15×

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

June 23, 2026 • 2 min read

Why does low‑latency inference matter now? As AI moves beyond single‑turn queries toward coordinated multi‑agent workflows, every millisecond counts. Autoregressive large language models still generate tokens one after another, a pattern that leaves GPUs under‑utilized and throttles throughput when responsiveness is essential.

Speculative decoding tries to fix that by letting a lightweight “drafter” propose future tokens while the main model checks them in parallel. DFlash, an open‑source block‑diffusion drafter, pushes the idea a step further: it spits out an entire block of candidate tokens in a single forward pass, converting what was a sequential draft into block‑parallel GPU work. The result, according to the authors, is up to a 15× speed‑up for gpt‑oss‑120b on NVIDIA Blackwell hardware, while keeping interactivity unchanged.

For Llama 3.1 8B, DFlash nearly doubles interactivity at the same concurrency compared with the leading EAGLE‑3 approach. The team has already posted 20 checkpoints on Hugging Face and supplied recipes for both Blackwell and Hopper GPUs, and they’re integrating the method into TensorRT‑LLM, SGLang and vLLM pipelines. The following data plot the latency‑throughput trade‑off on an eight‑node NVIDIA DGX B300 system using the SPEED‑Bench suite.

DFlash is well matched to this architecture because it exposes more parallel work to Blackwell's 15 PFLOPS of dense NVFP4 compute, serving up to 15x more users concurrently at the same interactivity rate. DFlash also shows interactivity speedups over EAGLE-3 speculative decoding across different datasets. The gains extend to smaller models as well, with DFlash nearly doubling performance over EAGLE-3 on Llama 3.1 8B for the Speed-Bench multilingual dataset. NVIDIA ecosystem brings DFlash to developers without application refactoring Researchers at UC San Diego released the paper DFlash: Block Diffusion for Flash Speculative Decoding in February 2026 as part of ongoing work on faster, more efficient LLM inference on NVIDIA Blackwell.

Boost Inference Performance up to 15x on NVIDIA Blackwell Using DFlash Speculative Decoding - NVIDIA Developer Blog

Why this matters

We see DFlash promising up to a fifteen‑fold boost in inference speed on NVIDIA’s Blackwell GPUs, a claim that hinges on speculative decoding’s ability to parallelise token generation. By letting a lightweight model draft tokens while the larger model verifies them, DFlash aims to keep Blackwell’s 15 PFLOPS of dense NVFP4 compute busy, potentially serving many more users at unchanged latency. The open‑source nature of the block diffusion approach could lower entry barriers for developers experimenting with multi‑agent workflows that demand low‑latency responses.

Yet the report offers no data on real‑world latency under varied loads, nor does it explain how much engineering effort is required to integrate DFlash into existing stacks. Comparisons to EAGLE‑3 suggest speedups, but the underlying test conditions remain vague. For founders and researchers, the headline is attractive, but we should remain cautious until broader benchmarks confirm that the theoretical parallelism translates into consistent performance gains across typical production scenarios.