AWS Trn2 UltraServer
The Trn2 UltraServer joins four trn2u.48xlarge instances into a single 64-chip Trainium2 machine, connected by NeuronLink-v3 at 1,024 GB/s per chip within an instance and 256 GB/s per chip between them. It carries 6,144 GiB of device memory at 185.6 TB/s, 768 vCPUs and 8 TB of host memory. AWS rates it at 83.2 PFLOPS of dense FP8, 42.8 PFLOPS across dense FP16, BF16 and TF32, and 11.6 PFLOPS of dense FP32, with a separate sparse figure of 164 PFLOPS covering FP8, FP16, BF16 and TF32 but not FP32. The sparse figure exceeds the dense FP8 one, so the two must not be conflated: dense FP16 is 42.8 PFLOPS, not 164. AWS's sparse mode runs at the chip's low-precision rate regardless of input format, making the sparse figure 1.97x dense FP8 but 3.84x dense BF16, so it is not 2:4 structured sparsity. Every figure here is AWS's own published aggregate and each one reconciles exactly against the Trainium2 chip specification. AWS publishes no power figure for the UltraServer or the chip.
We will point you at suppliers who have it. Free, and no signup.
System Details
GPU Configuration
System Specifications
Precision Performance Breakdown
| Precision | System Performance | Per GPU | Efficiency |
|---|---|---|---|
| 11.6 PFLOPS | 181 TFLOPS | — | |
| 42.8 PFLOPS | 669 TFLOPS | — | |
| 42.8 PFLOPS | 669 TFLOPS | — | |
| 42.8 PFLOPS | 669 TFLOPS | — | |
| 83.2 PFLOPS | 1,300 TFLOPS | — |
Where a vendor states a basis, the figure shown is the dense one, and any figure it publishes is listed separately. Vendors commonly headline the sparse number instead: for NVIDIA's 2:4 structured sparsity that is exactly twice the dense figure, though other vendors' sparse modes do not all follow that ratio.
Powered by AWS Trainium2
This system utilizes 64 × AWS Trainium2 GPUs, each delivering exceptional performance for AI and HPC workloads.
Per GPU TDP
—
Per GPU Memory
96 GB
Process Node
—
Architecture
NeuronCore-v3
Documentation & Resources
Flopper Spec Sheet
AWS Trn2 UltraServer specifications, generated from our database. Printable.
Vendor product page
AWS Trn2 UltraServer documentation
AWS announces Amazon EC2 Inf2 instances (Preview)
AWS • 2022-11-29
Typical Use Cases
The AWS Trn2 UltraServer runs 64× Trainium2 GPUs over NeuronLink-v3 at 1,024 GB/s per chip within each instance and 256 GB/s per chip between instances; EFAv3 networking at 3,200 Gbps, delivering 83.2 PFLOPS FP8 dense.