NVIDIA cuEquivariance and | News and Insights | ITZKXY | NVIDIA Network Adapters and InfiniBand Network Solutions

AI is a networking product designed for AI clusters.

ENN is a networking product designed for enterprise and data center networking.

On the top left corner, there is a sketch of a molecule. An arrow indicates it is an input to the neural network below. Another arrow from the network from a vector below (in the shape of a large arrow) shows that it is the neural network's output. This shows the usual input-network-output pipeline. This same pipeline appears on the right side, this time with the same molecule in another orientation, rotated from the left. It also connects to the network below and its output, which is rotated correspondingly. These two pipelines are connected with the words “Rotated input” at the top and “Rotated output” at the bottom. This shows that with equivariant networks, the rotation or network operation order does not matter; one can either rotate and pass through the network or pass through the network and then rotate; the result is the same as in the bottom right corner of the image. — This product is a networking product designed for enterprise and data center networking.

This product is a networking product designed for enterprise and data center networking.

ENNs is a NVIDIA networking product designed for AI clusters.

Product Information

AI (including Tensor Field Networks, LieConv, Cormorant, SE(3)-Transformer, NequIP and DiffDock Equiformer is a networking product designed for enterprise and data center networking.

irreps is a networking product designed for enterprise and data center networking.

, irreps irreps, Product Information irreps is a networking product designed for enterprise and data center networking.

Product Information

This product category is a NVIDIA networking product designed for enterprise and data center networking.

cuEquivariance is a networking product designed for enterprise and data center networking.

cuEquivariance is a storage product designed for storage networks.

This product category is a NVIDIA networking product designed for enterprise and data center networking.

cuEquivariance is a networking product designed for AI clusters.

The figure shows six panels titled “Speed up of NVIDIA cuEquivariance vs Baseline for Different Workloads on H100.” Each panel is dedicated to the comparison of a different model and operation. The top left panel shows cuEquivariance acceleration of 3.1x for DiffDock TP forward operation, and the top right model shows 15.6x for backward of the same. The middle left panel shows the speedup of MACE-OFF Large model symmetric contraction forward operation, with the acceleration of 17x concerning the baseline. The middle right panel shows 9x for the backward of the same. The bottom left model shows a 2.2x speedup for MACE-OFF small model tensor product forward operation. The bottom right shows MACE-OFF medium model tensor product forward operation with a 3.3x speedup. — 2 DiffDock is a networking product designed for enterprise and data center networking.

irreps is a networking product designed for enterprise and data center networking.

cuEquivariance is a NVIDIA networking product designed for enterprise and data center networking.

The figure shows four panels titled “Training and Inference time for MACE-OFF and MACE-MP Large on A100. The top row shows results for MACE-OFF Large model. The top left shows training time to be 6.1x accelerated for baseline, while inference acceleration is 7.2x. The bottom row shows the results for MACE-MP Large model. The bottom left shows training time to be accelerated by 5.9x, and the bottom right shows inference to be accelerated by 5.9x with respect to baseline. — MACE-OFF is a networking product designed for enterprise and data center networking.

Two panels titled “MACE-OFF Large Symmetric Contraction on Different GPUs.” MACE-OFF Large SC forward operation timings on the left panel are shown, where timing reduces from 6 milliseconds on NVIDIA A100 architecture to approximately four milliseconds on H100. In between, the timing gradually reduces for the L40 architecture. The same trend is also visible in the right panel for MACE-OFF Large model SC backward operation: the timings reduce approximately 25 milliseconds for A100 to further down for NVIDIA L40 and then NVIDIA H100 being approximately 20 milliseconds. — 4. MACE-OFF NVIDIA GPU and, 10K, FP32

Product Information

cuEquivariance is a networking product designed for AI clusters. DiffDock MACE is a networking product designed for enterprise and data center networking.

cuEquivariance is a networking product designed for AI clusters, enterprise networks.

NVIDIA

Extreme

Aruba

Ruckus

Optical Transceivers

HUAWEI

H3C

Wi-Fi Authentication

Sangfor

NetentSec

Esafenet

Panabit

Cisco

Integrated Appliances

Product Information

NVIDIA cuEquivariance and NEWS DETAIL

Product Information

Product Information

Product Information

Products

Solutions

Service Support

Contact ITZKXY