Promoting like hotcakes: The extraordinary demand for Blackwell GPUs illustrates the necessity for sturdy, energy-efficient processors as firms race to implement extra refined AI fashions and purposes. The approaching months will likely be crucial to Nvidia as the corporate works to ramp up manufacturing and meet the overwhelming requests for its newest product.
Nvidia’s newest Blackwell GPUs are experiencing unprecedented demand, with the corporate reporting that it has bought out of those next-gen processors. Nvidia CEO Jensen Huang revealed the information throughout an traders assembly hosted by Morgan Stanley. Morgan Stanley Analyst Joe Moore notes that Nvidia executives disclosed that their Blackwell GPU merchandise have a 12-month backlog, echoing an analogous state of affairs with Hopper GPUs a number of quarters in the past.
The overwhelming demand for Blackwell GPUs comes from Nvidia’s conventional clients, together with main tech giants like AWS, CoreWeave, Google, Meta, Microsoft, and Oracle. These firms have bought each Blackwell GPU Nvidia and its manufacturing companion TSMC can produce for the subsequent 4 quarters. The acute demand signifies that Nvidia’s already appreciable footprint within the AI processor market ought to proceed to develop subsequent yr, whilst competitors from rivals similar to AMD, Intel, and varied cloud service suppliers seize their share.
“Our view continues to be that Nvidia is prone to acquire share of AI processors in 2025, as the most important customers of customized silicon are seeing very steep ramps with Nvidia options subsequent yr,” Moore mentioned in a consumer word.
Nvidia unveiled the Blackwell GPU platform in March. It consists of the B200 GPU and GB200 Grace “tremendous chip.” These processors can deal with the demanding workloads of enormous language mannequin (LLM) inference whereas considerably lowering power consumption, a rising concern within the business.
Nvidia has overcome preliminary packaging points with its B100 and B200 GPUs, permitting the corporate to ramp up manufacturing. Each GPUs make the most of TSMC’s superior CoWoS-L packaging know-how. Nevertheless, questions stay about whether or not TSMC has enough CoWoS-L capability to satisfy the skyrocketing demand.
One other potential bottleneck within the manufacturing course of is the provision of HBM3E reminiscence, which is essential for high-performance GPUs like Blackwell. Tom’s {Hardware} identified that Nvidia is but to qualify Samsung’s HBM3E reminiscence for its Blackwell GPUs, including one other layer of complexity to the provision chain.
Additionally see: Nvidia Blackwell server set up showcase at Sizzling Chips 2024
In August, Nvidia acknowledged that its Blackwell-based merchandise had been experiencing low yields, necessitating a re-spin of some layers of the B200 processor to enhance manufacturing effectivity. Regardless of these challenges, Nvidia stays assured in its skill to ramp up Blackwell manufacturing in This fall 2024. It expects to ship a number of billion {dollars} price of Blackwell GPUs within the final quarter of this yr.