Design flaws will delay the launch of Nvidia’s next-generation Blackwell GPU architecture by three months or more, according to a report in The Information.
Blackwell is the successor to Hopper, the current generation on the market. It was supposed to ship in the fourth quarter of this year, but The Information reports that Nvidia informed Microsoft, a major customer of its GPUs, that it would not be able to ship product for at least three more months, pushing the launch into the first quarter of next year.
And if Microsoft is impacted, then so are other major customers like Meta (Facebook), Google, and Oracle. And if they are impacted, enterprises can forget about getting their own orders in, because Nvidia is going to prioritize its hyperscale cloud customers.
According to a report from semiconductor research firm SemiAnalysis, the problem lies in the manufacturing process, and it points to a problem in the packaging of the chip. “The core issue impacting shipments is directly related to Nvidia’s design of the Blackwell architecture. The supply of the original Blackwell package is limited due to packaging issues at TSMC and with Nvidia’s design,” writes SemiAnalysis.
The SemiAnalysis report delves deeply into the intricacies of the situation – much beyond my technical grasp. It is evident that Blackwell is an extremely sophisticated piece of silicon, and numerous factors could lead to complications, which indeed occurred. The responsibility to resolve these issues lies equally with TSMC and Nvidia.
An Nvidia representative provided this statement: “As we’ve stated before, Hopper demand is very strong, broad Blackwell sampling has started, and production is on track to ramp in 2H. Beyond that, we don’t comment on rumors.”
However, one analyst mentioned to me that if there is truly a delay in product shipment and if this delay could have a significant impact on earnings, Nvidia will be compelled to disclose it.
Given Nvidia’s preference for hyperscalers that place multi-billion dollar orders, these customers will likely experience the impact first. Enterprise adoption moves at a slower rate, and they are probably only now starting to implement Hopper-based systems. Therefore, if you have an order for some Hopper-based systems, there isn’t much cause for concern at this stage.
Earlier this year, Nvidia announced it would introduce a new microarchitecture every year instead of every two years, as it had been doing. At the time, I felt that this was full of risk and could potentially blow up in Nvidia’s face because it left no room for error. It would require flawless execution, and nobody’s perfect, not even Nvidia.
And now it seems it has happened. One frustrating part for Nvidia is that this problem is outside of its control. It’s a manufacturing issue with TSMC. Yes, the two companies work together on the packaging of the product, but the primary responsibility lies with TSMC.
This should give AMD pause, because it, too, has announced plans to have new GPU architecture every year for its increasingly popular Instinct line. If anyone knows the consequences of a misstep, it’s AMD. It was cruising along just fine in 2006 with a strongly competitive product against Intel before the Barcelona disaster – a new architecture that tried to do too much at once and ended up shipping late and underperforming. AMD lost all of its momentum and spent the next decade wandering the wilderness getting trounced by Intel before the Zen resurgence. AMD doesn’t need that again.
