AWS Trainium3

What Trainium3 is

AWS Trainium3 is the successor to Trainium2 in AWS's family of chips built for AI workloads, and it is AWS's first AI chip manufactured on a 3-nanometer process^[1][2]. The process number describes the manufacturing technology used to build the silicon: shrinking the features on a chip generally allows more transistors in the same area and less energy per operation, which is why chipmakers treat each node as a milestone.

Compute and memory per chip

Per chip, Trainium3 provides 2.52 petaflops of FP8 compute^[1][2]. FP8 is an 8-bit floating-point format; storing numbers in fewer bits reduces the memory and bandwidth needed to move them around, which matters once models get large. A petaflop is a quadrillion floating-point operations per second, so the figure describes peak throughput at that precision.

Each chip carries 144 GB of HBM3e memory and 4.9 TB/s of memory bandwidth, which AWS says is 1.5 times the capacity and 1.7 times the bandwidth of Trainium2^[1]. HBM, or high-bandwidth memory, is DRAM stacked in layers and placed close to the compute die so data can be fed to it quickly. AWS says the chip also doubles the compute performance of Trainium2^[1].

Trn3 UltraServers

AWS does not sell these chips only as individual parts; it packages them into a rack-scale unit. AWS says a Trn3 UltraServer supports up to 144 Trainium3 chips connected with NeuronSwitch-v1^[1]. Together, AWS says one such server provides up to 20.7 TB of HBM3e, 706 TB/s of memory bandwidth and 362 MXFP8 petaflops^[1].

The rack matters because training a large model spreads a single job across many chips that must exchange data constantly. The interconnect — here a switch named NeuronSwitch-v1 — determines how quickly that exchange happens and therefore how many chips can work together as one system.

Compared with Trn2 UltraServers, AWS says the Trn3 version delivers up to 4.4 times higher performance, 3.9 times higher memory bandwidth and more than 4 times better energy efficiency^[1].

Data types and workload types

AWS says Trainium3 supports the MXFP8 and MXFP4 data types and is designed for dense and expert-parallel workloads^[1]. Dense workloads run the whole model on every input, so the work scales with model size. Expert parallelism is used with mixture-of-experts models, which split a model into many separate "expert" sub-networks and route each input to only a few of them; spreading those experts across many chips makes the communication between chips a central design concern — which is where the switch and the rack-level memory figures come in.

Availability

AWS announced general availability of Amazon EC2 Trn3 UltraServers powered by Trainium3 in December 2025^[2]. Offering the hardware through Amazon EC2 means customers rent capacity in the cloud rather than buying and operating the chips themselves, and that the chip reaches users through the same service interface as AWS's other compute instances.

Where it sits in the line

The published numbers frame Trainium3 as a step up from Trainium2 along several axes at once: compute per chip doubles, memory capacity and bandwidth per chip grow, and the rack-scale system is claimed to reach several times the performance and energy efficiency of its predecessor^[1]. The memory and efficiency gains are proportionally larger than the per-chip compute gain, which reflects a practical constraint in large-model training: how fast a model can be trained depends not only on arithmetic throughput but on whether data can be kept close to the compute units and on how much power the whole system draws.