AWS Trainium3
Overview
AWS Trainium3 is Amazon's fourth-generation AI accelerator and, in AWS's own words, "our first 3nm AWS AI chip", made generally available at re:Invent on 2 December 2025. Each chip carries eight NeuronCore-v4 delivering 2,517 TFLOPS of MXFP8 and MXFP4, 671 TFLOPS across BF16, FP16 and TF32, and 183 FP32 TFLOPS, with 144 GiB of HBM3e at 4.9 TB/s and 256 MiB of on-chip SBUF scratchpad. Device-to-device traffic runs over NeuronLink-v4 at 2.56 TB/s per chip, double Trainium2. AWS publishes a sparse figure of 2,517 TFLOPS spanning FP16, BF16 and TF32 but explicitly not FP8; that is 3.75x the dense BF16 rate and exactly equal to the dense MXFP8 rate, so it reflects the chip running sparse work at its low-precision rate rather than any conventional 2:4 doubling. AWS publishes no TDP for the chip.
Performance
Peak theoretical throughput by precision type
| Precision | Dense | 2:4 Sparse |
|---|---|---|
FP64 | No verified data available | Structured sparsity is a tensor-core feature; this vector precision has no sparse form |
FP32 32-bit floating point | 183TFLOPS | Structured sparsity is a tensor-core feature; this vector precision has no sparse form |
TF32 TensorFloat-32 | 671TFLOPS | 2,517TFLOPS |
BF16 Brain Float 16 | 671TFLOPS | 2,517TFLOPS |
FP16 16-bit floating point | 671TFLOPS | 2,517TFLOPS |
FP8 8-bit floating point | 2,517TFLOPS | The part supports 2:4 sparsity, but the vendor publishes no sparse figure for this precision |
FP6 | No verified data available | The part supports 2:4 sparsity, but the vendor publishes no sparse figure for this precision |
FP4 4-bit floating point | 2,517TFLOPS | The part supports 2:4 sparsity, but the vendor publishes no sparse figure for this precision |
INT8 | No verified data available | The part supports 2:4 sparsity, but the vendor publishes no sparse figure for this precision |
Specifications
Architecture
NeuronCore-v4
Form Factor
No verified data available
Launch Year
2025
Process Node
3nm
Memory
144 GB HBM3e
Bandwidth
4,900 GB/s
TDP
No verified data available
Max power
No verified data available
Interconnect
2.6 TB/s NeuronLink-v4
direction and scope not stated by vendor
NeuronCore-v4
8
Spec Confidence
Official
Full Specifications
| Compute Engine | |
|---|---|
| NeuronCore-v4 | 8 |
| Memory | |
| Memory | 144 GB |
| Memory Type | HBM3e |
| Bandwidth | 4.9 TB/s |
| Interface Width | No verified data available |
| On-Die SRAM | 256 MB |
| Interconnect & I/O | |
| GPU-to-GPU | NeuronLink-v4 |
| Interconnect Bandwidth | 2.6 TB/s direction and scope not stated by vendor |
| Power & Thermal | |
| TDP | No verified data available |
| Max power | No verified data available |
| Enterprise Features | |
| Sparsity | Yes |
| General | |
| Form Factor | No verified data available |
| Architecture | NeuronCore-v4 |
| Process Node | 3nm |
| Launch Year | 2025 |
Datasheet & Resources
Data Provenance
Every figure traced to a source
Primary Source
- Publisher
- AWS
- Published
- No verified data available
Data Quality
- Spec confidence
- Official
- Clock basis
- Boost
- Core precisions with figures
- 6 of 9
- Normalization
- All values in TFLOPS
Compare cloud providers offering on-demand GPU instances for AI training, inference, and HPC workloads.
Browse GPU Cloud ProvidersSystems Using This GPU
Pre-configured systems featuring the AWS Trainium3
| System | GPU Count | Peak Performance | Total Power | |
|---|---|---|---|---|
AWS Trn3 UltraServer UltraServer · rack · 2025 | 144x Trainium3 | 362 PFLOPS | No verified data available | View System |
Frequently Asked Questions
How many TFLOPS does the AWS Trainium3 have?
The AWS Trainium3 delivers 183 TFLOPS FP32, 671 TFLOPS FP16 and 2,517 TFLOPS FP8 at peak.
What is the power consumption of the AWS Trainium3?
Flopper does not currently have a verified TDP (Thermal Design Power) figure for the AWS Trainium3.
How much memory does the AWS Trainium3 have?
The AWS Trainium3 is equipped with 144 GB of memory with 4,900 GB/s of memory bandwidth.
What architecture is the AWS Trainium3 based on?
The AWS Trainium3 is based on the NeuronCore-v4 architecture, launched in 2025.
Stay Updated on GPU Releases
Get notified when new GPUs are added or specifications are updated.
No spam, unsubscribe anytime.