Clarity for AI buildersSponsored by AI Routing
adoption

vLLM reports 0.87-second H3-Image 4K generation on four Vera Rubin GPUs

vLLM reports 0.87-second H3-Image 4K generation on four Vera Rubin GPUs
vLLM reports native 4K H3-Image generation in 0.87 seconds on four NVIDIA Vera Rubin GPUs. According to the vLLM engineering post, vLLM-Omni runs MiniMax’s open-weight image model at native 4096×4096 resolution, with a measured median response time of 0.866 seconds in a short-prompt serving benchmark. A separate configuration reached 6,115 images per hour. Those results represent different operating points: the fastest individual responses and the highest aggregate throughput were not achieved together. H3-Image supports text-to-image generation and image editing through one checkpoint, using six sampling steps and up to 16 reference images. It produces native high-resolution output without an upscaler. The integration extends across consumer and data-center hardware, although the post’s matched performance comparisons cover NVIDIA systems. Latency versus throughput The subsecond result used four GPUs with four-way Ulysses sequence parallelism, identified as SP4, and MXFP8 for eligible diffusion-transformer linear operations. At concurrency one, the benchmark recorded a 0.944-second p95 response time. The authors report that the median was 2.21 times faster than their GB200 configuration and 1.64 times faster than their GB300 configuration, also using four GPUs, SP4 and MXFP8. Peak throughput instead came from four independent single-GPU replicas, designated SP1×DP4. At concurrency eight, that setup delivered 6,115 images per hour with a 4.411-second median latency. SP4 reached 4,407 images per hour at concurrency two, with a 1.421-second median. For serving operators, the distinction is consequential: distributing one image’s computation across four GPUs favors response time, while assigning independent requests to separate GPUs delivered more images overall in these tests. The reported 38.7% throughput advantage for replicas should therefore not be read as an improvement in interactive latency. What the benchmark measures The Rubin serving sweep used 20 short prompts, each 122–160 characters long, across eight seeds, with 160 warm requests at each concurrency point. A separate cross-hardware comparison used a 1,292-character coffee-product prompt and reported three-seed means. Rubin averaged 1.09 seconds for that longer 4K request, rather than the headline 0.87 seconds. The post explicitly distinguishes the request corpora, HTTP clients and summary statistics. It also cautions that GPU count, precision and memory placement vary across its broader hardware results. Those comparisons describe deployed serving systems, not an isolated measurement of GPU architecture. Memory placement is particularly important for smaller deployments. A single RTX 5090 using CPU layerwise offload took 52.29 seconds for the cross-hardware 4K generation request and 61.45 seconds for a one-reference edit, including the complete PNG response. Offload makes a single-GPU configuration possible, but introduces weight-transfer work that resident data-center configurations avoid. Where the speedups come from The authors describe H3-Image as a Qwen3-VL encoder paired with a 33.1-billion-parameter dense diffusion transformer and an image VAE. Native 4K output contributes 16,384 target-image tokens across six denoising steps before decoding; reference images add encoding and conditioning work. The implementation combines eligible MXFP8 operations, sequence parallelism, cached conditioning projections and distributed VAE decoding. All six denoising steps still run. In a separate BF16 control on four GB300 GPUs, combining CUDA Graph replay with multiple CUDA streams reduced VAE decoding from 2.635 seconds to 0.213 seconds and end-to-end latency from approximately 4.0 seconds to 1.6 seconds. That control supports the decoding optimization, rather than independently establishing the Rubin headline result. Capacity and revenue remain projections At the measured peak rate, the authors extrapolate daily capacity to 146,755 images per four-GPU Rubin system. This was not a 24-hour load test. Using an assumed sale price of USD 0.06879 per image, drawn from MiniMax’s Standard 4K API pricing example before launch discounts, the post projects gross revenue of USD 105.16 per GPU-hour at full billable utilization, or USD 78.87 at 75%. These are revenue scenarios before operating costs, not profit estimates or demonstrated commercial returns. Actual hosted-API charges also vary with text-token usage. The post additionally documents community implementations for AMD Instinct, Intel Arc Pro, Huawei Ascend and Moore Threads hardware. Those validations use their own workloads and configurations, so they do not establish matched performance against the NVIDIA results.