Why it matters: The market has been eagerly awaiting Nvidia’s next generation of Blackwell-based GPUs, but it may have to wait even longer than anticipated because there are already signs of a forthcoming shortage. It is doubtful, though, that supply issues will make a dent in demand.
They haven’t even made their first appearance yet, but it looks like it could be hard to get hold of Nvidia’s much-anticipated next-generation Blackwell-based GPU products when they do arrive.
That was the message that Nvidia CFO Colette Kress delivered to the market at the company’s closely scrutinized earnings report this week.
“We expect our next-generation products to be supply constrained as demand far exceeds supply,” Kress said.
This probably should not come as a surprise to observers as Team Green is already war-gaming which clients get its current generation of AI chips. To stave off customers ordering more than they need, or worse, decamping to rival products such as AMD’s forthcoming AI chips, CEO Jensen Huang addressed the issue of fairness and how the company allocates its highly coveted products during the earnings call. “We allocate fairly. We do the best we can to allocate fairly, and to avoid allocating unnecessarily,” he said.
Even without Kress’ warning, shortages of Nvidia’s next-generation B100 products were pretty much a foregone conclusion. For starters, customers have almost surely pre-ordered some of these products. Also, while much is not known about Blackwell, it has been frequently rumored that it will be Nvidia’s first architecture to adopt multi-chiplet designs, which can complicate packaging. On the other hand, as Tom’s Hardware notes, a multi-chiplet design could simplify production on the silicon level as it is easier to maximize yields of smaller chips.
Another potential hiccup is the question of whether Nvidia’s suppliers can keep up with the demand. Nvidia’s AI GPUs are all fabbed at the Taiwan Semiconductor Manufacturing Co., which has its own production limits. Then there are the rumors that Nvidia is mulling an earlier release of the Blackwell lineup, pushing it up to the fourth quarter of 2024, a source has told Moore’s Law is Dead.
With demand so intense, the market is eager to see the wraps removed from Blackwell, supply issues or not. There has been a wealth of speculative information leaked on Blackwell’s expected performance, with some rumors purporting that the increase in performance won’t be as impressive as the jump from RTX 3000 to RTX 4000. A leaker on the Chiphell forum posted what it claimed to be stats for the RTX 5090: a 50% increase in scale (which presumably refers to cores), a 52% increase in memory bandwidth, 78% increase in L2 cache, 15% increase in frequency, and 1.7x performance uplift.
We could find out sooner than expected.
Next month Nvidia is hosting its GPU Technology Conference that will feature a keynote with Huang talking about AI. Tweaktown speculates it would be the perfect venue to unveil the Blackwell GPU architecture.
