Tim Culpan reports that Nvidia plans to start shipping GB200 servers to major cloud service providers in early December after resolving yield-killing issues with the packaging of its B100 and B200 AI GPUs. The analyst said major customers such as AWS, Meta, Microsoft and Oracle are lining up to receive Nvidia’s next-generation AI servers just a month behind schedule. However, the exact amount is unknown.
Microsoft plans to introduce first GB200-based server
As expected, Nvidia will ship NVL36 and NVL72 servers first. Due to revised shipping schedules, the GB200-NVL36 (with 36 Blackwell GPUs) and GB200-NVL72 (with 72 GPUs) server versions, originally scheduled for late October and early November, respectively, will be delivered in the first week of December. It will be shipped to the United States, Karpan said. claims. Unfortunately, it is unclear how many machines are expected to ship in December.
According to analysts, Microsoft is expected to receive one of the largest allocations of these servers. AWS, Meta, and Oracle are also among Blackwell’s alpha clients. Despite production disruptions, demand for GB200 servers remains strong.
Nvidia admitted in August that it had to produce “low-yield Blackwell material” to meet demand, hurting its profit margins. These low-yield B200 GPUs may be used in the first GB200-NVL servers and may impact their volumes.
Nvidia Chief Executive Officer Jensen Huang said last month that all necessary design changes for the Blackwell B100 and B200 GPUs have been completed. The company plans to start mass production in the fourth quarter as planned, and NVIDIA’s will start in late October. Given that TSMC’s 4NP production cycle is approximately three months, unless TSMC finds a way to expedite the production and packaging process, the sophisticated Blackwell GPU will be ready as soon as late January. It will be.
Meanwhile, Blackwell GPU module manufacturers and system integrators are already getting certified samples of reworked GPUs (it’s about time), according to Kalpan. Once the new silicone has been qualified, the appropriate product can be assembled.
How Blackwell servers are manufactured
Nvidia sells Grace CPUs and Blackwell GPUs directly to clients such as Microsoft and Oracle. However, the latter obtains the actual servers from the manufacturer, which requires the manufacturer to secure a contract with a customer that can allocate GPUs from Nvidia. Here’s how it works:
TSMC uses its CoWoS-L process to produce and package silicon, and this is rumored to be the cause of Blackwell’s production disruption. Once Grace GPUs and GPU packages are ready, they are sent to companies such as Foxconn Industrial Internet (FII) for module assembly. You can then install the module on an Ariel (1 Grace CPU, 1 Blackwell GPU) or Bianca (1 Grace CPU, 2 Blackwell GPUs) board.
These components are sent directly to the system integrator, who assembles the final server rack and ships it to the customer. Foxconn and Quanta are the primary integrators delivering the first Blackwell-based servers to Nvidia customers. Brands such as Asus, Gigabyte, Luxshare, and Wiwynn can also assemble reference NVL machines, but before they can get the components they will need to win contracts with key customers to negotiate GPU allocations with Nvidia .
Initially, all server assembly will be done in Taiwan. However, Nvidia is considering expanding production to other regions, including Mexico and Texas, in the first quarter of 2024. This will increase production capacity as demand (and supply of B100 and B200 GPUs) increases.
