SAN JOSE: CHIP DESIGNER Nvidia admitted at its GPU Technology Conference (GTC) today that physically large GPUs have difficulty passing verification.
With Nvidia's Volta GPU architecture stacking DRAM on the same silicon substrate as the GPU, it will require the firm to increase the size of its chips. According to Sumit Gupta, GM of Nvidia's Tesla GPU Accelerated business unit, large chips are not hard to manufacture but rather tend to encounter difficulties passing verification.
Nvidia's decision to stack DRAM will bring both performance and capacity benefits but it will result in the firm sending larger chip designs to TSMC. However Gupta said that fabbing big chips is not a problem for Nvidia or its manufacturing partners, instead the challenge is getting those chips verified to make sure they operate within a set of tolerances.
Typically fabs such as TSMC, Globalfoundries and UMC are judged on their yield, that is, the percentage of chips that pass verification. Gupta is indeed right when he says that it isn't harder to fab larger chips, but the probability of a chip having a defect goes up as chip size - either in physical dimensions or more commonly transistors - increases.
Nvidia plans on taking a risk with Volta's DRAM stacking. The performance and energy efficiency benefits could indeed be large but it will also put considerable pressure on TSMC's manufacturing process. And given Nvidia's desire to span future architectures across all of its product lines, low chip yields due to the complexities of the underlying architecture could be very costly indeed. µ