NVIDIA AI Updates – Blackwell Platform, NIM Agent Blueprints and MLPerf

With earnings recently announced, NVIDIA has been busy with several AI updates. The company recently clarified Blackwell, introduced NIM Agent Blueprints, and released the results of MLPerf performance tests. These innovations from NVIDIA have helped companies develop, deploy, and scale AI solutions.

To understand the significance of these announcements, I attended a briefing with Dave Salvator, NVIDIA’s director of accelerated computing products, and Justin Boitano, VP of enterprise AI.

Not just a GPU—Blackwell is a platform

To start the briefing, Salvator emphasized one point by sharing a slide depicting the chips that Blackwell uses. “Blackwell is a platform,” he said. “It’s very important to understand this. The GPU is just the beginning. As you see across the top row, those are photos of all the chips that go into a Blackwell system to make it do what it does, which is taking us into that next era of generative AI.”

NVIDIA said it designed Blackwell to meet the rigorous demands of modern AI applications. The latest MLPerf Inference v4.1 benchmarks show Blackwell delivering up to four times the performance of previous-generation GPUs. The company said this leap in performance came from several key innovations, including the second-generation Transformer Engine and FP4 Tensor Cores.

“MLPerf is an industry-standard AI benchmark that looks at training and inference performance for data center edge and even small devices,” Salvator said. “We’re big believers in industry standard benchmarks because these are where different companies can come together, run the same workload, and have directly comparable results. In addition, these results, of course, are vetted by all submitters.”

Bringing together multiple technologies

According to the company, Blackwell integrates multiple NVIDIA technologies, including NVLink and NVSwitch, for high-bandwidth communication between GPUs. This approach is essential for real-time, large-scale AI inference tasks. NVLink and NVSwitch enable the Blackwell system to handle the increasing demands of LLMs, such as Llama 2 70B, which require low-latency, high-throughput token generation for real-time performance.

In the MLPerf benchmarks, Salvator said Blackwell handled complex inference tasks across various AI workloads well. One example: its ability to efficiently run LLMs with billions of parameters highlights its potential in industries like finance, where real-time data analysis and decision-making are critical.

Blackwell’s superior performance ensures that enterprises can meet stringent latency requirements while simultaneously serving many users.

Understanding Blackwell as a system

Salvator underscored that Blackwell is about integrating multiple components into a high-performing, cohesive system. It includes a suite of NVIDIA chips—such as the Blackwell GPU, Grace CPU, BlueField data processing unit, and NVLink Switch—that work together to set the standard in AI and accelerated computing.

This system-level approach enables Blackwell to achieve impressive results in AI inference tasks. By optimizing the interaction between these components, NVIDIA has created a platform that not only excels in performance but also efficiency and scalability, making it a game-changer for enterprises looking to deploy AI at scale. Businesses should be able to deploy a Blackwell system to gain both performance and cost efficiency.

NIM Agent Blueprints: Accelerating enterprise AI adoption

Justin Boitano followed Salvator to discuss NVIDIA NIM Blueprints. To set up that discussion, he took a broad view. “This transition to generative AI really has the potential to usher in a wave of productivity the world’s never seen before,” he said. “Now, the first wave of generative AI was really the infusion of AI into internet-scale services driven by makers of foundational models. And we traditionally think of this as something like ChatGPT, and it was created to improve individual user productivity by writing language and writing code. But it’s expanded into how we search the internet, write email, transcribe and record meetings.”

He said that the next wave is starting now.

“It represents a bigger business process transformation that will affect how teams work across the enterprise,” he said. “It’s going to happen in enterprises to help them activate what we traditionally think of as institutional knowledge that only they have about how they run their businesses and how they engage their customers, helping them create a new form of intelligence to drive innovation faster than ever before.”

That’s where NIM Blueprints come in as a foundation for enterprises looking to get started with generative AI. They’re essentially comprehensive reference workflows tailored to specific AI use cases, such as customer service, drug discovery, and data extraction from PDFs. Each blueprint comes equipped with sample applications, reference code, customization guides, and deployment Helm charts, offering developers a head start in creating AI-driven solutions.

Some case studies

NVIDIA said that what sets NIM Blueprints apart is their ability to foster a continuous improvement cycle through a data-driven AI “flywheel.” As companies use applications and generate new data, data feeds back into the system to refine and enhance the AI models, making them more intelligent and effective over time.

In the healthcare industry, NVIDIA said that NIM Blueprints can accelerate drug discovery by leveraging generative virtual screening workflows. As a result, researchers can identify promising molecules more efficiently, reducing time and cost while increasing the likelihood of successful outcomes.

In customer service, enterprises can use NIM Blueprints to create digital human avatars that interact with customers in a more engaging and personalized way, enhancing user experience and satisfaction.

In addition, the adaptability of NIM Blueprints means that enterprises across various sectors—from retail to finance—can tailor these workflows to meet their needs. The modular design enables businesses to integrate NIM Blueprints with their existing systems, supporting a more seamless and efficient deployment of AI solutions. This flexibility is crucial for companies that want to remain competitive in an increasingly AI-driven market.

Some final thoughts

The biggest misconception about NVIDIA is that it’s a chip company. While it makes best-in-class GPUs, maintaining market leadership at the chip level at all times is impossible. NVIDIA has used its prowess in software and systems to create a moat around itself. In fact, Salvatore told me the company has almost twice as many software engineers as it does hardware, which shows the level of commitment NVIDIA has in delivering solutions as systems.

Zeus Kerravala is the founder and principal analyst with ZK Research.

Read his other Network Computing articles here.