Nvidia has finally fixed the defect that prevented the mass production of Blackwell AI accelerators

Nvidia’s Blackwell-based compute accelerators, introduced in March of this year, were ready for mass production closer to the fall, but a flaw discovered in their design prevented the production of chips in sufficient quantities to meet demand. Now the head and founder of the company claims that the defect was eliminated, and TSMC itself helped developers in this area.

Image source: NVIDIA

«We had a flaw in the Blackwell design. The chip functioned fully, but the defect would prevent it from being produced with an acceptable level of defects. One hundred percent of the blame lay with Nvidia,” Jensen Huang unexpectedly frankly spoke about the failures of his brainchild. At the same time, he denied rumors about the emergence of contradictions between Nvidia and TSMC on this basis. He added that to create a Blackwell-based computing system, seven types of semiconductor components had to be developed from scratch and prepared for production simultaneously.

According to Huang, it was TSMC that helped Nvidia eliminate the identified defect and resume mass production of Blackwell in an incredibly short time. Deliveries of chips from this family will begin in the fourth quarter, and the company expects to earn several billion US dollars from their sales by the end of the current fiscal year. The reporting period in the Nvidia calendar ends at the end of January next year. Some partners of companies like Dell intend to ship the first Blackwell-based server systems to customers as early as next month, but they will not become widespread until next year. According to some estimates, Microsoft remains the largest customer of systems based on Blackwell accelerators.