Nvidia AI Chip Releases Move To ‘One-Year Rhythm’ With H100 Successors For 2024, 2025

The newly disclosed road map shows that Nvidia plans to move to a ‘one-year rhythm’ for new AI chips and release successors to the powerful and popular H100, the L40S universal accelerator and the forthcoming Grace Hopper Superchip—the latter of which combines a GPU and CPU—in each of the following two years.

Nvidia has revealed in a new road map that it plans to release successors to its powerful and popular H100 GPU for AI training and inference in 2024 and 2025, reflecting a move to a new yearly release cadence from its previous strategy of launching AI chips every two years.

The road map, included in an investor presentation on Nvidia’s website, also showed that the AI chip behemoth plans to release successors to the L40S universal accelerator and the forthcoming Grace Hopper Superchip—the latter of which combines a GPU and CPU—in each of the following two years.

[Related: Nvidia CEO Explains How AI Chips Could Save Future Data Centers Lots Of Money]

In addition, the chip designer plans to introduce a new category of chips, an offshoot of the Grace Hopper Superchip with increased computing capabilities for AI training and inference, according to the road map, which was disclosed by Nvidia in private investor meetings, a spokesperson told CRN.

It was published online as Nvidia seeks to defend its dominant position in the AI computing space in the face of Intel, AMD, a slew of chip startups and cloud service providers such as Amazon Web Services trying to capture market share amid a boom in demand for chips driven by generative AI workloads.

H200 And B100 IN 2024, Unnamed GPU In 2025

In the road map, Nvidia said it’s moving to a “one-year rhythm” for new AI chips, starting in 2024 with the H200, which is a direct successor to this year’s H100, the most powerful GPU in Nvidia’s current lineup. The H200 will use the same Hopper architecture that powers the H100.

Nvidia classified the H200, its predecessors and its successors as designed for AI training and inference workloads running on systems powered by x86 CPUs from Intel or AMD.

The successor to H200 is the B100, which is also expected in 2024. Multiple rumors have indicated that the “B” stands for a next-generation GPU architecture called Blackwell.

Following the B100 is an unnamed 100-series GPU coming in 2025 that carries a placeholder name, the “X100,” according to an Nvidia spokesperson in response to a CRN inquiry.

Successors To Grace Hopper, L40S In 2024 And Beyond

Grace Hopper, also known as GH200, is the first in a new series of chips from Nvidia that combine an Arm-based CPU and a GPU for high-performance AI inference workloads in the data center.

In the road map, Nvidia said GH200 is targeted for AI inference and a 2024 launch window. The company has previously said that GH200, which will use its Grace CPU and a Hopper GPU, will be available in systems starting in the second quarter of next year.

2024 is also when Nvidia plans to release GH200’s successor, the GB200, which will incorporate a next-gen GPU architecture reportedly called Blackwell.

A successor with the “GX200” placeholder name will follow in 2025.

On the opposite side of the road map is the L40S, a universal accelerator designed for x86 data center systems used by enterprise customers.

Based on Nvidia’s Ada Lovelace architecture, the L40S is built for AI inference and training as well as demanding graphics and simulation applications, providing faster performance than Nvidia’s A100 from 2021.

While Nvidia previously systems with the L40S will launch this fall, the road map marks the accelerator as a 2024 product. Later in 2024, Nvidia is expected to release a successor to L40S called the B40, which indicates that the company will use the next-gen architecture reportedly called Blackwell.

Following the B40 in 2025 is an unnamed GPU with the “X40” placeholder name.

New Chip Category For Arm-Based AI Training And Inference

The road map includes a new category of chips that appear to be an upgrade over Nvidia’s Superchips that combine a CPU and GPU to power AI training and inference workloads in Arm-based systems.

The chips combine a CPU and GPU but carry the suffix “NVL,” which is the same suffix Nvidia uses for the H100 SVL product that combines two H100 PCIe cards.

Nvidia has not provided any further details on these NVL chips, including which form factor they will have, but traditionally chips with a CPU, including the GH200, fit into a motherboard socket.

The new series will debut in 2024 with the GH200NVL, which suggests a Grace Hopper Superchip with higher performance due to the focus on both inference and training.

A successor will follow later that year with the GB200NVL, which will make use of the next-gen GPU architecture reportedly called Blackwell.

The GB200NVL will then be followed by an unnamed chip with the NVL suffix in 2025. That chip carries the “GX200NVL” placeholder name.

Yearly Cadence Set For InfiniBand, Ethernet Products

Nvidia’s portfolio of InfiniBand and Ethernet networking products for AI infrastructure will also move to a yearly release cadence, according to the road map.

This will start with the company’s 400 Gb/s InfiniBand and Ethernet products in 2024, which carry the brand names Quantum and Spectrum-X respectively.

Nvidia then expects to move to 800 Gb/s Quantum and Spectrum-X products later in 2024, followed by 800 Gb/s products the next year.