When Nvidia unveiled its B200 chip for AI applications in March this year, it clearly had an absolute beast on its hands. Its features are staggering: 208 billion transistors, next-generation “Blackwell” architecture, a peak performance of 20 petaflops in FP4 operations when accompanied by liquid cooling, and the ability to work side-by-side with a memory map of up to 192 GB of VRAM. It also achieves a bandwidth of 8 TB/s.
Interestingly, the first units of the B200 chip have been slow to arrive. Nvidia has admitted that the performance of its manufacturing processes was lower than expected, which forced its engineers to redesign some layers of the chip to fix a problem that has delayed the delivery of the first units. “We executed a change to the Blackwell GPU mask to improve production yield,” Nvidia said, according to Tom’s Hardware.
Jensen Huang Has Shown His Face
The changes made by the company’s technicians to the mask have worked, and mass production of the chip is already underway. As such, Nvidia has presumably solved the problems that hampered the performance of its manufacturing processes. It has published the first performance metrics of the B200 chip on its blog. Based on its specifications, it should be very powerful. These initial figures indicate the B200 chip will undoubtedly be the most powerful AI GPU on the market when it arrives. However, it’s prudent to approach Nvidia’s numbers with a degree of skepticism.
Huang reiterated that Nvidia was responsible for the problem. However, thanks to its collaboration with TSMC, it has addressed the issue.
Interestingly, the issues that arose during the manufacturing of the B200 chips have led some media outlets to suggest that TSMC is to blame and that its relationship with Nvidia has deteriorated. Nothing could be further from the truth. Nvidia CEO Jensen Huang has stated that the company he leads is responsible for the problem. He added that, thanks to the cooperation of TSMC, Nvidia has managed to resolve the issue and is ready to start large-scale production of this advanced chip for AI.
“We had a design flaw in Blackwell. It was functional, but the design flaw caused the yield to be low. It was 100% Nvidia’s fault,” Huang said, according to Reuters. The delay in production had allegedly caused tensions between Nvidia and TSMC, but Huang dismissed that as “fake news.”
Clearly, Nvidia’s CEO takes full responsibility for the problem with the B200 GPU and wants to ensure TSMC isn’t adversely affected by this mishap at all costs. Producing this chip hasn’t been easy for Nvidia, but the company expects to deliver the first units of this GPU before the end of the year.
Image | Nvidia edited by Xataka En
Related | AMD Launches New GPUs for AI. The Question Is Whether They’ll Be Able to Surpass Nvidia’s Offerings
View 0 comments