Hardware
Chips, accelerators and the physical substrate AI runs on.
-
AMD Instinct MI455X
Moves to CDNA 5 on 2nm and HBM4, with four times the MI355X’s peak 4-bit compute and nearly three times its bandwidth.
-
AMD Instinct MI430X
An MI400 Series GPU that keeps full hardware FP64 for supercomputers alongside 4-bit AI compute.
-
AMD Instinct MI350X
CDNA 4 on 3nm with new MXFP4 and MXFP6 formats and 288 GB of HBM3E.
-
AMD Instinct MI355X
Liquid-cooled CDNA 4 part at higher clocks than the MI350X, with MXFP4 and MXFP6 and 8 TB/s.
-
AMD Instinct MI350P
Puts MI350 silicon on an air-cooled PCIe card for existing enterprise servers, at half the MI355X’s compute and memory.
-
AMD Instinct MI300X
AMD’s first CDNA 3 GPU, adding FP8 and sparsity and 192 GB of HBM3 for large language model work.
-
AMD Instinct MI300A
Merges CPU and GPU on one package with shared HBM3.
-
AMD Instinct MI325X
Same CDNA 3 compute as the MI300X with HBM3E, raising memory to 256 GB and bandwidth to 6 TB/s.
-
NVIDIA Jetson AGX Thor (T5000)
Moves Jetson to a Blackwell GPU with up to 7.5x the AI compute of Jetson Orin.
-
Broadcom Tomahawk 6
Broadcom’s 102.4 Tbps Ethernet switch, the networking silicon that ties large AI clusters together.
