Nvidia unveils new AI-focused processors and development tools

Nvidia, at the Graphics Technology Conference (GTC) 2021, unveiled a range of software and hardware innovations designed to develop and deploy artificial intelligence to drive business.

On the opening day of the conference, Nvidia chief executive Jen-Hsun Huang (Jensen Huang) presented the first ARM-based Grace CPU, intended for supercomputers and “giant AI.”

The CPU supports the modern LPDDR5x memory subsystem, enabling twice the bandwidth and ten times the energy efficiency compared with DDR4.

The fourth-generation NVLink interconnect will deliver a 900 GB/s link between the Grace CPU and Nvidia GPUs.

Nvidia says Grace-based systems will train a natural-language processing model with 1 trillion parameters ten times faster than comparable x86-based systems.

The first customers for the CPU include the Swiss National Supercomputing Centre and the US Department of Energy’s Los Alamos National Laboratory.

The company expects to begin shipments in early 2023.

Nvidia also announced the BlueField-3 data processing unit — an accelerator designed to isolate infrastructure services from applications running on x86 CPUs.

BlueField-3 has 16 Arm A78 cores and accelerates network traffic to 400 Gbps. The company said that delivering similar services from a data-centre would require up to 300 CPU cores.

The chip features a tenfold increase in compute power over the previous generation, and a fourfold acceleration for cryptography tasks.

BlueField-3 is the first DPU to support PCIe Gen 5. The new generation of DPUs is fully backwards compatible with BlueField-2. BlueField-3 is set to go on sale in the first quarter of 2022.

Nvidia demonstrated the cuQuantum SDK to accelerate simulations of quantum circuits on GPUs.

Researchers at the California Institute of Technology reportedly set a world record for simulating Google’s Sycamore quantum processor and increased per-GPU performance ninefold.

“I hope cuQuantum will do for quantum computing what cuDNN did for deep learning,” Huang said.

The company also unveiled a new wave of Nvidia servers certified to ship with enterprise GPUs A30 and A10.

A30 is the lighter version of Nvidia’s A100. It supports a broad range of AI tasks, training and traditional compute workloads. The GPU can support AI use cases such as recommendation systems, speech processing and computer vision.

The graphics processor with Tensor Cores A10 supports accelerated graphics, rendering, AI and classical compute workloads.

A10 and A30 are built on Nvidia’s Ampere architecture, equipped with 24 GB of memory and PCIe Gen 4.

Nvidia also opened access to the Jarvis conversational AI framework to users of the NGC platform.

Jarvis models offer high-accuracy automatic speech recognition, real-time translation into several languages and new text-to-speech capabilities for conversational AI agents.

Using GPU acceleration, the end-to-end speech pipeline (listening, understanding and formulating an answer) can run in under 100 milliseconds.

Nvidia announced the next generation Drive platform for autonomous vehicles, called Atlan. This is the first automotive processor with 1,000 TOPS performance, representing a fourfold increase over the previous Orin generation.

The SoC Atlan features a next-generation GPU architecture, new Arm cores, new deep-learning and computer-vision accelerators. It also includes a BlueField DPU with full programmability required to prevent data leaks and cyberattacks.

Nvidia expects to begin delivering Atlan to automakers in 2025, while not hindering sales of the previous-generation platform, which will debut in autonomous vehicles next year.

During the quarter Nvidia reported results for the first quarter of fiscal 2022. Chief Executive and Chief Financial Officer Colette Kress said revenue for the short quarter had already topped the earlier forecast of $5.3 billion. She noted that the growth was driven by heightened demand for AI services.

“We expect this to lead to increased usage of our platform by cloud providers and more purchases over the year,” added Kress.

She noted that demand far outstrips supply. Nvidia expects GPU shortages to persist “for most of this year.”

The company also raised its forecast for Q1 revenue from CMP professional-mining chips to $150 million from $50 million.

Recall that in March 2021 insiders reported that Nvidia was considering releasing CMP HX mining accelerators based on the GA100 GPU.

In February Nvidia said that would limit the hash rate of GeForce RTX 3060 GPUs, “to keep GeForce GPUs in the hands of gamers,” the company stressed.

However in March the mining-performance restrictions on GeForce RTX 3060 GPUs were bypassed with an experimental driver released by the software maker.

Subscribe to ForkLog news on Telegram: ForkLog AI — all AI news!