DeepL strengthens its research and neural network training with latest NVIDIA technology
Global AI communications company DeepL is expanding its AI translation and communication capabilities with the installation of an NVIDIA DGX SuperPOD.
DeepL has added a DGX SuperPOD with 68 NVIDIA DGX H100 systems which are supplied by DELTA Computer Products. This is a significant step forward for DeepL, as it is the first commercial deployment of this scale within Europe. Along with other DeepL servers, the new DGX SuperPOD is installed at a green data center in Sweden, which is powered by renewable energy.
CIO INFLUENCE News: Rockwell Automation Signs Agreement To Acquire Autonomous Robotics Leader Clearpath Robotics
Through LINPACK benchmarking, DeepL and NVIDIA reviewed the installation to assess the cluster speed. The benchmark findings will be submitted to the November 2023 edition of the Top 500 fastest commercially available computer systems list. The benchmark yielded a performance of 21.85 PFlop/s—which would place the cluster at #26 in the world on the current Top 500 list (June 2023 edition) and #8 in Europe.
The NVIDIA DGX SuperPOD installation illustrates DeepL’s goal to push technology forward by training large language models (LLMs) to produce the world’s most advanced AI communication tools. The individual servers work in concert to form a supercomputer that enables optimal training of the advanced neural networks that power DeepL’s products.
Each of the DGX systems contains eight NVIDIA H100 Tensor Core GPUs, connected via NVIDIA NVLink interconnects for ultra-fast GPU-to-GPU communication within the server. Between each server, the NVIDIA Quantum-2 InfiniBand networking platform with in-network computing technology ensures that every GPU can communicate with each other in the cluster at maximum network speed. This enables scalable performance and ensures that the compute cluster can be used efficiently in a variety of configurations, including for training large models and running multiple training workloads.
“Research has been at the core of DeepL’s mission from the very beginning—and has allowed us to advance our state-of-the-art neural networks that achieve such high-quality translations,” says DeepL’s CEO and founder Jarek Kutylowski. “While we need team expertise and ingenuity to build a great product, the availability of powerful compute resources for training models has been important to us since the buildout of our first clusters in Iceland. The new NVIDIA DGX SuperPOD cluster is a game changer, and its speed is crucial to enabling us to build larger model sizes. We are excited to combine our existing training capabilities with the new cluster and look forward to the improvements and features it will bring our users.”
Since 2017, DeepL’s AI communication tools have changed the way businesses around the world communicate. Its translator is trained to detect linguistic nuance and context, and even to decipher industry-specific terminologies—and has thus enabled thousands of businesses to take their products and services to a global audience.
CIO INFLUENCE News: Palo Alto Networks Cements Position in Taiwan With New Local Cloud Infrastructure Investment
The NVIDIA DGX SuperPOD enables DeepL and its research team to significantly reduce their time-to-market for large-scale language models. “Enterprises require purpose-built, high-performance infrastructure optimized to accelerate LLM and generative AI workloads efficiently and at scale,” says Charlie Boyle, vice president of DGX systems at NVIDIA. “NVIDIA DGX SuperPOD provides the compute performance, networking and software delivered as a turnkey data center solution required for DeepL to build and train the modern AI applications their customers demand.”
CIO INFLUENCE News: Scala Data Centers Joins Datacloud USA 2023 as Sustainability Partner
[To share your insights with us, please write to sghosh@martechseries.com]
