NVIDIA GX200 launching in 2025
NVIDIA’s latest Investors Presentation outlines the company’s plans for annual data-center GPU updates.
A recent slide from NVIDIA’s Investor Presentation this month has drawn attention to SemiAnalysis, which provides extensive coverage of both public and nonpublic NVIDIA strategies to establish dominance in the data center market. The site delves into various aspects, including the company’s plans for incorporating HBM3e memory, updates to PCI Express standards (6.0 and 7.0), and modifications to NVIDIA’s multi-GPU interconnect technologies like NVLink. For those interested in these intricate details, we recommend subscribing to SemiAnalysis for more comprehensive insights.
Meanwhile, let’s shift our focus to NVIDIA’s newly unveiled official roadmap, which spans multiple years. NVIDIA has been relatively reserved about sharing roadmaps lately, possibly due to the close alignment between gaming and data-center GPU architectures, the challenges in using the most advanced process nodes and the intensifying competition in the data center market. Consequently, NVIDIA is cautious about specifying dates, primarily due to the challenges in synchronizing its plans with foundries and the readiness of software deployment.
On a broader scale, it’s worth noting that NVIDIA aims to introduce new data center GPU architectures on an annual basis. The updated roadmap reveals that the forthcoming 2024 Hopper GH200 GPU (which has been announced recently) will be succeeded by the Blackwell GB200 processor between 2024 and 2025, eventually followed by the GX200 GPU a year later. These are the codenames for the graphics processors, but the actual products will use names like H200, B100, and X100.
The “B” designation represents Blackwell, a next-generation architecture intended to span both data-center and gaming products. This roadmap confirms the existence of the GH200 for data center purposes, dispelling earlier rumors that suggested the possibility of GB100/GB102. Although all options should still be considered, NVIDIA has now officially confirmed the GB200 as the expected launch.
Furthermore, there’s a mysterious architecture referred to as “X,” which could either be named after a scientist or simply serve as a placeholder. Few scientists have last names that start with an “X,” but it’s premature to speculate at this point. What’s crucial is that the X architecture is slated for launch in 2025, accompanied by a new data center product known as “X100.”
This roadmap implies that the Blackwell data-center product is expected to be launched sometime between late 2024 and early 2025, potentially allowing more time for the introduction of the X architecture. Whether the “X” designation will also be applied to gaming products remains uncertain. However, what has been confirmed is that Blackwell will make its debut as part of the GeForce RTX 50 series.
|NVIDIA GPU Architectures|
GA100 (8192 CUDA)
GH100 (18432 CUDA)
GA102 (10752 CUDA)
AD102 (18432 CUDA)
GA103 (7680 CUDA)
AD103 (10752 CUDA)
GA104 (6144 CUDA)
AD104 (7680 CUDA)
GA106 (3840 CUDA)
AD106 (4608 CUDA)
GA107 (2560 CUDA)
AD107 (3072 CUDA)