Red Hat pledges day-zero support for Nvidia’s newest GPUs – News

IBM Corp. subsidiary Red Hat is moving more aggressively than usual to ensure its software stacks are ready the moment new generations of Nvidia Corp.’s artificial intelligence hardware reach the market, a strategy executives said is driven by surging demand for larger and more capable AI model architectures.

The company disclosed this week that Red Hat Enterprise Linux will support Nvidia’s new Vera Rubin platform the day it ships later this year. The deal marks a tighter collaboration process between the two firms. The Vera Rubin graphics processing unit anchors Nvidia’s next rack-scale architecture built around its Grace CPU-based design for training and running multi-trillion-parameter AI systems. It sports 336 billion transistors, delivering 50 petaflops of performance and training speeds that are 250% faster than the predecessor Blackwell GPU, according to Nvidia.

Red Hat’s move is a response to an industrywide acceleration in the pace of hardware development that enterprises are struggling to keep up with, Red Hat Chief Technology Officer Chris Wright said in a briefing.

“You see Nvidia’s cycle for hardware delivery speeding up to be an annual cycle, whereas, in the past, you’d see new GPUs or accelerators coming out every couple or even three years,” he said.

As a result, he said, Red Hat and Nvidia are working together earlier in the development process. “We’ve had a long engineering partnership,” Wright said. “This is doubling down on that and making sure we’re engaged from the earliest points to co-engineer a solution with Red Hat stack complemented by Nvidia software and hardware.”

Running the gamut

Red Hat said its Vera Rubin support spans RHEL for hardware enablement, OpenShift for cluster orchestration and Red Hat’s AI software layer for inference, model deployment and agent workloads. Validated Nvidia GPU OpenRM drivers and the CUDA toolkit will be accessible directly via RHEL repositories, and RHEL will support Nvidia’s Confidential Computing framework for the entire AI lifecycle, providing cryptographic proof of workload protection.

Nvidia is accelerating hardware development to accommodate an explosion in model sizes and new reasoning agents, said Justin Boitano, the firm’s vice president of enterprise AI products. Today’s frontier open models have reached one trillion parameters, and Nvidia expects 10-trillion parameter models to be in use by the end of next year, he said.

New classes of “reasoning models,” which not only predict the next token in a sequence but also generate intermediate reasoning steps that are useful in solving complex tasks, have further amplified demand. “Those reasoning models powering agents are driving another 5x token increase per year,” Boitano said. That combination has resulted in compute demand that is “just insane at this point.”

Annual refresh

The same dynamics are pushing Nvidia toward annual refresh cycles on standardized rack-scale blueprints. “We co-engineer these rack-scale architectures, which are really six different processors, to make sure that we’re driving data and models as efficiently as possible,” Boitano said.

Early adopters are expected to include hyperscale and AI-focused “neocloud” providers, large financial institutions and public-sector organizations building private AI infrastructure. Wright said Vera Rubin will appeal to “cloud providers that are building their solutions for their end customers,” as well as “important large-scale enterprises that are building their own internal infrastructure.”

A key enterprise feature of Rubin is rack-scale confidential computing, which extends hardware-level encryption protections from CPUs to GPUs and across interconnects. Nvidia uses a hardware-based approach that extends a CPU’s trusted execution environment to the GPU, creating a unified, secure domain for AI workloads.

Boitano said security is becoming increasingly important because “a lot of model providers don’t want to deliver their model weights into a data center that they don’t operate.” Confidential computing “will open the market for these frontier models running in your data center.”

Image: News/Microsoft Designer