SAP and Nvidia expand partnership to aid customers with gen AI

SAP and Nvidia announced an expanded partnership today with an eye to deliver the accelerated computing that customers need in order to adopt large language models (LLMs) and generative AI at scale.

Under the partnership, SAP is integrating Nvidia’s generative AI foundry service, including the newly announced Nvidia NIM inference microservices, into SAP Datasphere, SAP Business Technology Platform (BTP), RISE with SAP, and SAP’s enterprise applications portfolio.

“We are embedding AI in our enterprise applications, and we’ve designed it in such a way that customers in the cloud can consume it easily, as-a-service, out-of-the-box,” said Philipp Herzig, SAP chief artificial intelligence officer (CAIO).

Herzig notes that SAP has a large ecosystem of partners and various LLM providers, with new LLMs popping up seemingly every day.

“We wanted to design it in a way that customers don’t have to care about complexity,” he said. “At the end of the day, customers want the best experience, the best performance, or the lowest price in order to consume LLMs within their workflows.”

SAP is using Nvidia’s services — including Nvidia DGX Cloud AI supercomputing, Nvidia AI Enterprise software, and Nvidia AI Foundation models — to add new capabilities to the generative AI hub in SAP AI Core and SAP Datasphere.

Bringing RAG to Joule

Nvidia NeMo Retriever, a semantic-retrieval microservice unveiled last November that helps gen AI applications provide more accurate responses via retrieval-augmented generation (RAG), will bolster SAP’s Joule copilot.

RAG optimizes LLMs by giving them the ability to reference authoritative knowledge bases outside their training data.  

“There are tons of documents that are not residing in an SAP system,” Herzig said. “Those might be your HR policy, your travel policy, compliance documents, your legal documents, that might be in a SharePoint or on a portal.”

Joule already has the power to answer simple questions, like, “How many vacation days do I have left?” That’s just a matter of the employee record. But Nvidia’s microservice empowers Joule to go a step further by giving it access to the HR policy and compare that with the employee record, too.

New features for data scientists, developers

The partnership is also exploring more than 20 gen AI use cases aimed at helping customers simplify digital transformation, including automating ERP with intelligent invoice matching in SAP S/4HANA Cloud, improving HR use cases via SAP SuccessFactors, and using gen AI insights from SAP Signavio to process business recommendations and optimize customer support processes.

Meanwhile, SAP is leveraging NVIDIA’s accelerated computing platforms and NVIDIA AI Enterprise data science software, including Nvidia Rapids, Rapids cuDF, and cuML, to make it easier for data scientists to access data and enhance ML workload performance in Datasphere. For developers, Nvidia AI foundry services will help them create domain-specific language code and fine-tune LLMs to write code in SAP’s Advanced Business Application Programming (ABAP) programming language.

“Lots of legacy code has been written in our programming language called ABAP, and it turns out that usually large language models are not good with ABAP,” Herzig said.

He explained that while out-of-the-box LLMs can produce ABAP code that would’ve been acceptable in the 1990s, it doesn’t mesh with the modern design principles for ABAP Cloud.

“One thing that doesn’t really work with RAG is code generation,” added Kari Ann Briski, Nvidia’s VP of generative AI software product management. “You can’t retrieve code and then copy it and write that. Generative code really has to come from knowing the language and being able to create it to give a natural prompt.”

Finally, the new Nvidia NIM inference microservices, also announced today, will help customers optimize inference performance across their SAP infrastructure.

“It’s basically microservices pre-built by Nvidia for a variety of models: open models from the ecosystem, proprietary models, some models from Nvidia, models from small startups to big clouds,” said Manuvir Das, VP of enterprise computing at Nvidia. “We take the model and we package it together with the engines that are optimized for these models to run as efficiently as possible across a range of Nvidia GPUs that you can find in laptops or workstations and data centers or clouds.”