The green guys had a bit of a problem with the fact that the Blackwell chip design contained an error and deliveries to large customers could not be fulfilled. Someone doesn’t get bonuses!
It’s unusual, but NVIDIA CEO Jensen Huang admitted that the design flaw in the Blackwell chip was 100 percent his company’s fault. The defect was helped to be fixed by TSMC, which initially raised the issue. The whole situation seems to have caused a few gray hairs for the inks from NVIDIA, because it caused a delay in production, while big players like Meta, Google or Microsoft are completely surprised by Blackwell’s graphics. Jensen opined that the Blackwell chips were functional, but a design flaw caused low yield. And the plan was that Blackwell GPUs would be shipped to customers already in the second quarter of this year.
The problem was caused by how the chiplets are connected in the B100 and B200 GPUs. These use TSMC’s CoWoS-L technology. This technology further relies on the RDL (Redistribution layer) interpose (essentially a communication interface with silicon bridges). And here was the glitch, as there was a mismatch in the thermal expansion properties between the various components, causing the system to warp and fail (ugh). So NVIDIA fixed the chips and increased the yield. Usually, such repairs are quite lengthy, but in cooperation with TSMC, the errors were fixed very quickly, and Blackwell is now back on track. Mass production should already start and deliveries will begin in early 2025.
And interestingly, Google ordered 400,000 GB200 chips for more than $10 billion. Meta also has an order for 10 billion. Well, if they can, right?
Source: pctuning.cz
