Eliovp reported that Paiton runs FLUX.2 klein 4B at around 1 second per 1024 × 1024 image on a single Radeon AI PRO R9700, four steps, after warmup, achieving roughly a third less peak Torch allocation down to 12.9 GiB, with maximum sampled driver VRAM at 14.6 GiB, and the whole model pipeline stays on the GPU with no CPU offload.
Public source
Publisher name
Public post
We’re still building quietly. This time, we’ll let the images do the talking ;-) Paiton runs FLUX.2 klein 4B at around 1 second per 1024 × 1024 image on a single Radeon…
Company
Eliovp
Speeding Up Machines & Accelerating People
- Industry
- IT Services and IT Consulting
- Location
- Sint Gillis Waas, BE
- Company size
- 2–10 employees
About Eliovp
Try to reimagine high performance computing. Now open your eyes. You´re here. If you’re into high-end hardware, this is your happy place. Automated, optimized and fit for the future.
See moreLatest activity
Latest activity from Eliovp
5 signals
Research & Knowledge
Eliovp reported that Paiton runs FLUX.2 klein 4B at around 1 second per 1024 × 1024 image on a single Radeon AI PRO R9700, four steps, after warmup, achieving around 19% more images per euro than the strongest stock configuration qualified.
Research & Knowledge
Eliovp reported that Paiton runs FLUX.2 klein 4B at around 1 second per 1024 × 1024 image on a single Radeon AI PRO R9700, four steps, after warmup, achieving 16% lower generation latency and 19% more projected images per hour than the strongest stock configuration qualified.
Research & Knowledge
Eliovp announced that Paiton runs FLUX.2 klein 4B at around 1 second per 1024 × 1024 image on a single Radeon AI PRO R9700, four steps, after warmup.
Discover more
Similar signals
Similar public activity from other companies.
Research & Knowledge
Labyrinth Labs
Labyrinth Labs demonstrated layer streaming on an RTX 3050 Laptop GPU to run Llama 3.1 8B Instruct at 119.6 tok/s with 3.32 GB peak VRAM.
Research & Knowledge
Yantrion Inc
Yantrion Inc measured 975.5 aggregate decode tokens per second across 56 concurrent requests on one 8× AMD Instinct MI350X node using Kimi-K3.
Research & Knowledge
Amazon Web Services (AWS)
Amazon Web Services (AWS) benchmarked model loading on p5.48xlarge and found the bottleneck flips with model size.
Research & Knowledge
Sustainable AI Group
Sustainable AI Group published data on pre-training and model development phases for GPT-5.6, noting that 700,000 A100-equivalent GPU hours for automated red teaming equates to about 308 MWh, a quarter of the energy needed to pre-train GPT-3.
Research & Knowledge
Cloudian Inc