Due to technical issues, Nvidia has delayed shipments of Blackwell GPUs by a few months, but the company could push an alternative Blackwell GPU to various market segments.

Semiconductor research firm SemiAnalysis said that Nvidia plans to introduce a new Blackwell GPU called B200A, which will be a lower-end alternative to the B200 GPU, which has been delayed.

“The B200A will be used to fulfill demand for lower-end and mid-range AI systems,” SemiAnalysis said in a report.

The B200A GPU will include up to 144GB of HBM3E memory and draw up to 1000 watts of power.

It will be based on the well-established older CoWoS-S packaging technology, while the delayed B200 is based on the newer CoWoS-L packaging technology.

The B200A GPU will be used in servers such as the MGX GB200A NVL36, which supports up to 36 GPUs. That may attract hyperscaler customers looking to build smaller AI models.

Smaller companies can’t afford larger AI systems, and the MGX model with GB200A will be more affordable. Large LLMs with billions and trillions of parameters will remain the domain of large companies with established data centers.

The GB200A servers will likely ship in large volumes, which requires error-free packaging and production techniques. Nvidia is already dealing with a shortage of GPUs, with many large customers still waiting for Hopper GPUs.

SemiAnalysis said the B200A will be based on a die called B102, “which will also be used in the China version of Blackwell, called B20.”

Nvidia has already announced the Blackwell Ultra for 2025 to succeed the Blackwell GPUs. There will be a B200A Ultra version of the GPU, which may include various upgrades.

The B200A may come out a lot quicker as it may not be affected by packaging constraints that are delaying the release of the B200. The CoWoS-S, which was used for Hopper, is already well established.

The B200 may have been issues relating to bridge die issues, which need to be redesigned, SemiAnalysis claimed. The extreme density of the chip may have also created issues.

The B200 delay raises questions about the timeline for the newer GPUs, including Blackwell Ultra and the Rubin and Rubin Ultra GPU platforms, due in 2026 and 2027, respectively.

Microsoft recently announced Blackwell GPUs will be available in Azure in early 2025. Google and Amazon have announced the availability of Blackwell systems in the cloud but have not announced launch dates, which may now be in 2025.

Nvidia’s aggressive roadmap of a new GPU a year likely created design issues. But such issues may go away when CoWoS-L becomes mainstream.

Nvidia hasn’t officially commented on the delay of the Blackwell B200 GPUs, which will go into HGX, MGX, and DGX servers. However, more details may emerge on August 28, when the company announces earnings. Nvidia has seen its stock tumble after reports of delays.

Nvidia has a dominant market share in GPUs, but AMD is making up ground with its MI300X GPU, which is scheduled to come out later this year.