GreenNode AI Lab proudly announces a high-performance in image generation with Stable Diffusion, marking a milestone in our AI innovation.

Led by the AI Lab of GreenNode, the team was tasked with building the Image Generation by Stable Diffusion system and optimizing its workflow.

Tackling Challenges with Stable Diffusion

The AI Lab embarked on a mission to enhance the inferencing system of our internal application utilizing Stable Diffusion. Despite the novelty of Stable Diffusion, we encountered challenges in scaling up to meet increasing demand and task complexity. Our primary focus was to enhance throughput and reduce latency in the inferencing system, without compromising accuracy or reliability.

Through rigorous experimentation, we optimized each component of our pipeline. To efficiently improve the model layer, we ensured that other components were not bottlenecks. Key optimizations included implementing an asynchronous mechanism for the gRPC layer and leveraging TensorRT in the model layer instead of common torch models.

We enhance performance by transitioning from Sync to Async mechanisms at the gRPC layer to handle more requests.

Remarkable Achievements

Following the implementation of optimizations on the NVIDIA A40 Tensor Core GPU, we achieved remarkable improvements. Specifically, we attained over a 3.5x increase in speed and more than a 3.8x increase in throughput compared to common torch versions.

This breakthrough underscores GreenNode's commitment to pushing the boundaries of AI innovation and delivering cutting-edge solutions to our customers and partners. The adoption of appropriate technology stacks for AI systems can lead to cost savings and a reduced carbon footprint. Improved inference speed directly translates to enhanced user experiences and cost savings for the company.

Technical Details

NVIDIA Products Utilized:

GPUs: NVIDIA H100 Tensor Core GPU, A40 and NVIDIA GeForce RTX 4090.
NVIDIA Triton Inference Server, TensorRT.

Stable Diffusion v1.5 Performance (images/min)

Other Related Products and Information:

DL frameworks: TensorRT
Protocols: gRPC
Data: Pretrained checkpoints, LORAs, and auxiliary components (controllnet-canny, controllnet-depth, etc.)
Total data volume: Approximately 300 GB in float16 datatype
Inference on GPUs: 4 x NVIDIA A40.

Introduction to Artian: Where AI Meets Art

Powered by our AI Lab team, Artian demonstrates the remarkable capabilities of Generative AI in crafting captivating game characters, maps, storylines, and promotional materials for VNGGames. This digital art platform combines AI and art, serving game studios under VNGGames.

Artian is the combination of 3 services:

ImageGeneratorSD
Triton
CMS

And here are the impressive Artian statistics until now:

116 images of game characters processed
Nearly 72,000 background images processed

By harnessing the innovative capabilities of Generative AI showcased through Artian, our AI Lab exemplifies our commitment to pushing the boundaries of AI innovation.
At GreenNode, we are committed to continuing our journey of exploration and innovation, driving impactful solutions for our customers and partners. With our dedication to excellence and relentless pursuit of innovation, GreenNode is poised to shape the future of AI and empower businesses across industries.
Subscribe to GreenNode to stay updated with the latest news!