ArcHPC GPU Optimization Software – Technical Exchange Meetings

Date: 02/21/2024

Vendor: Arc Compute

Technology/Topic: ArcHPC GPU Optimization Software

TEM Video (milTube): https://www.milsuite.mil/video/watch/video/68037

______________________________________________

Welcome to the Technical Exchange Meeting (TEM)!

Arc Compute is a research and development company building optimization software for accelerated hardware, revolutionizing the HPC space. We aim to fully optimize the performance and utilization of the most prominent accelerated hardware technologies starting with NVIDIA GPUs. Contrary to popular belief GPUs do not instinctively operate at peak performance given the task/computation it is completing at any given moment. ArcHPCs managing, directing and reconfiguring of kernels, optimizing thread arrangement and warp scheduling in flight outperform the same hardware by a factor 1.4x to 3x while only using halve/half as much of the hardware and completes 2x the amount of tasks at any given time. This results in a power savings of 13% to 38% and reduced compute time of 29% to 65.7% of code that was not optimized for ArcHPC but it itself was optimized code.

Using ArcHPC AI/ML model training and inference time can be significantly reduced while performing 2x amount of tasks, this is not limited to just AI/ML as ArcHPC is beneficial to all GPU dependent workloads because it is optimizing the execution of instructions which they all depend to complete tasks on GPUs. Whether a code is fully optimized, for example NVIDIA CuBLAS, there are deadtimes and suboptimal optimization being performed by the GPU itself as it handles the execution/completion of the task presented. What would take 6 months to train could take you 2 months to 4 months and what could take an AI-powered fighter .300 seconds could be reduced to 0.1029 to 0.213 seconds with unoptimized code for ArcHPC meaning this time can be substantially reduced even further.

Arc Compute’s ArcHPC is building the “air flight controller” and “Google maps” of computer instructions/threads compute environment so that at any given moment the tasks executing and tasks computing are completed through the most optimal tuned environment starting with accelerated hardware. This will increase the performance per wattage and efficiency of environments leading to the most minimal amount of hardware required to scale in the digital world because we are maximizing pipelines across hardware. As AI/ML advances the demand for hardware exponentially increases while the ability to service the demand is linearly based so organizations must be able to fully optimize all aspects of their compute environment.

______________________________________________

To join the DISA TEM mailing list, please contact: disa.tem@mail.mil

______________________________________________

Disclaimer:

— TEMs do not serve as a marketing venue or request for proposal actions.

— TEMs shall not be interpreted as a commitment by the Government to issue a solicitation or ultimately award a contract.

— TEMs do not serve as an endorsement of any presented technologies or capabilities

— Presentations will not be considered as proposals nor will any awards be made as a result of a TEM session.

— TEMs are public open forums – no proprietary or sensitive information should be presented during TEM sessions. Only publicly facing content is permissible in DISA TEM sessions.