Up until this point, not much has been said about Navi architecture, other than vague descriptions citing “scalability” and “next generation memory”. Other than that, we’ve been told that Navi will be built on a 7nm FinFET process when it launches next year.
Today, however, a new report is suggesting that Navi will be the first GPU from AMD with circuity that is dedicated to artificial intelligence (AI) operations. According to Fudzilla, Raja Koduri’s Radeon Technologies Group wanted to enable such functionality in Vega, but just didn’t have the time/resources to make that leap.
However, the smaller 7nm process along with a more favorable development schedule, AI underpinnings should be ready for primetime with Navi. If you recall, NVIDIA was able to implement full-fledged AI circuity into its Tesla V100 (Volta GV100) GPU. Tesla V100 has 640 dedicated Tensor cores to tackle heavy AI workloads, which can deliver 120 TFLOPS of deep learning performance.
This performance uplift with regards to AI operations is exactly what AMD is after with Navi. NVIDIA says that the Tensor cores in Volta V100 offer a 12x performance advantage over Pascal GPUs that lack the necessary hardware, and instead rely on CUDA cores.
AMD’s Navi GPUs will be built using GlobalFoundries’ 7nm node using 7LP technology. “Thanks to additional improvements at both the transistor and process levels, the 7LP technology is exceeding initial performance targets and expected to deliver greater than 40 percent more processing power and twice the area scaling than the previous 14nm FinFET technology,” said GlobalFoundries back in mid-June.