Google unveiled two new in-house AI accelerators at its annual Cloud Next conference in Las Vegas on Wednesday: one designed to speed up training and another aimed at driving down model serving costs. The Chocolate Factory boasts its eighth-gen tensor processing units are as much as 2.8x faster in training and deliver 80 percent higher performance per dollar for LLM inference compared with last year's Ironwood TPUs. To achieve this, Google has dual-tracked its accelerator development, building the TPU 8t for training and TPU 8i for inference. While these chips are built on similar foundations, each is specifically aimed at eliminating bottlenecks in its respective workload. Google isn't the first to go down this road. Early in its AI chip development, Amazon Web Services recognized the need for inference- and training-optimized accelerators. Nvidia has also dabbled with this kind of specialization, though not to the same extent.…