A.I.
Rycode

Jeffrey 1 min read

Hardware · Chips

Instinct

Instinct is AMD's line of AI chips. The new chip named in the deal is a custom GPU, or graphics processing unit, based on the MI450 architecture, meaning it is built on AMD's MI450 chip design. Meta already runs the MI300 and MI350 series. The purchase is on the Meta stack.

AMD Instinct MI350-series accelerator module: a black board with the AMD logo, memory around a silver frame, and a blue, green, and orange compute die, on a black background.
An AMD Instinct MI350-series module. Image: AMD, via TechNews

MI450

On 24 February 2026 AMD said the first deployment uses a custom Instinct GPU based on the MI450 architecture, for Meta's workloads at gigawatt scale. A gigawatt is a billion watts, a common way to size data centers by the power they draw. The GPU is paired with 6th Gen AMD EPYC CPUs codenamed Venice, AMD's ROCm software for running AI code on its GPUs, and AMD Helios, a rack-scale architecture, meaning the whole server rack is designed as one machine. AMD said shipments for the first gigawatt begin in the second half of 2026. The full purchase is up to 6 gigawatts of Instinct GPUs.

AMD rendering of one Helios rack on a black background: a wide dark cabinet with AMD HELIOS printed down its side, its open front showing stacked server trays with yellow and blue cabling.
AMD rendering of the Helios rack. Image: AMD

MI455X and Helios

AMD's current top Instinct GPU is the MI455X, on its fifth-generation CDNA design: 40 PFLOPS of FP4 and 432 GB of HBM4 at 23.3 TB/s per GPU, per AMD. Helios puts 72 of them in one rack for 2.9 EFLOPS of FP4 and 31 TB of HBM4, with volume deployments expected in the second half of 2026 (AMD, Helios). A PFLOPS is a thousand trillion operations per second and an EFLOPS a thousand PFLOPS. HBM4 is the newest stacked high-bandwidth memory. The full figures are on the MI455X and Helios page.

MI300 and MI350

The same 24 February 2026 release says Meta already runs millions of AMD EPYC CPUs, AMD's server processors, and significant deployments of Instinct MI300 and MI350 series GPUs.

Sources