Hardware
Chips, accelerators and the physical substrate AI runs on.
-
The A100 becomes the unit of AI compute
NVIDIA announces the Ampere A100, built for both training and inference.
-
Cerebras builds a chip the size of a wafer
Cerebras reveals the Wafer Scale Engine, a single chip cut from an entire silicon wafer.
-
TSMC ships EUV chips in high volume
TSMC announces its N7+ node is delivering EUV-lithography chips to customers at volume.
-
AWS builds its own inference chip
Amazon announces Inferentia, a custom chip built specifically for running trained models.
-
NVIDIA’s Volta adds dedicated tensor cores
NVIDIA launches the V100 with tensor cores, silicon dedicated to matrix multiplication.
-
Apple ships a neural engine in every iPhone
Apple’s A11 Bionic debuts a dedicated Neural Engine for on-device machine learning.
-
Google reveals its custom TPU
Google announces the Tensor Processing Unit, a chip built for its own neural networks.
-
JEDEC standardises High Bandwidth Memory
JEDEC publishes JESD235, the first standard for stacked high bandwidth memory.
-
Tesla P100 brings HBM2 to the datacentre
NVIDIA reveals the Pascal-based Tesla P100, its first datacentre GPU built on HBM2.
-
Stanford trains deep networks on GPUs
Stanford researchers show graphics chips train deep belief networks far faster than CPUs.