AWS Trainium3
Eight NeuronCore-v4 cores and 144 GiB of HBM per chip. Its low-precision and memory upgrades are much larger than its dense BF16 compute increase over Trainium2.
HBM
Published peak, not measured application throughput
144 GiB is about 154.62 decimal GB. NeuronLink bandwidth is separate from EFA networking and does not establish single-stream token latency. A logical NeuronCore is not an additional physical chip.
When this is a sensible choice
Start here if…
Consider it when a supported workload benefits from more memory, faster collectives or MXFP8/MXFP4 kernels. Verify the compiled model and actual chip layout before scaling training or serving.
Choose another configuration if…
Do not project the low-precision peak onto ordinary BF16 training. AWS lists 671 TFLOPS dense BF16, close to Trainium2’s 667; workload speedup needs measurement.
Specifications with their boundaries attached
- Architecture
- NeuronCore-v4
- Memory
- 144 GiB HBM
- Memory bandwidth
- 4.9 TB/s per accelerator
- Peak compute
- 671 TFLOPS BF16/FP16/TF32 dense; 2,517 MXFP8/MXFP4 TFLOPS
- Scale-up interconnect
- NeuronLink-v4: 2.56 TB/s per device, as specified by AWS
- Host attachment
- AWS-managed host; verify the selected EC2 configuration
- Power
- Not published in checked per-chip reference
- Partitioning
- Physical and logical NeuronCore allocation; not NVIDIA MIG
- Catalogue status
- Documented product
Compute figures are theoretical peaks at the stated precision. Dense and structured-sparse rates must not be mixed. Bandwidth labelled bidirectional combines both directions. See the source documents.
Three bandwidths, three different jobs
NeuronLink-v4: 2.56 TB/s per device, as specified by AWS
InfiniBand, RoCE, EFA or provider-specific transport
Where it appears in provider documentation
| Provider / machine | Network scope | What changes the decision |
|---|---|---|
AWS Trn3 architecture ↗ | NeuronLink-v4 is device-to-device; scale-out networking is a separate specification | The Neuron reference documents the chip and software architecture. It does not establish regional stock, price or an instance-level network allowance. |
Model fit and software support
These publisher or serving-engine documents mention this hardware family. They have not been reproduced on our machines.
No model-specific recipe in our reviewed set certifies this exact hardware. The memory calculator can narrow candidates, but it cannot establish software support. Read the model register.
What is the memory floor?
Start with total parameters, then add the memory the workload needs. This arithmetic does not certify a serving configuration. All output sizes below are decimal GB.
Override device capacity with the memory exposed by your allocation, particularly for cloud B300 and partitioned devices. The starting 32 GB budget and 15% reserve are editable teaching assumptions. They are not measurements for the selected model. Mixed-precision tensors, quantization scales, vision encoders and draft models can increase the weight payload.
Raw weights only
320B × 8 bits ÷ 8
Budget per device
154.62 GB × (1 − 15%)
Arithmetic lower bound
round up ((weights + 32) ÷ budget)
This assumes perfectly balanced sharding. A result of three does not prove that a three-device parallel layout is supported. Check layer/expert divisibility, actual allocatable memory, precision kernels and the fabric before renting.
320B is the model card’s total; 18B active is not its storage size. The vLLM recipe reports about 306 GiB for the native FP8 checkpoint. Hopper requires BF16 KV for this model; the documented ROCm path targets gfx950, not every Instinct GPU. Read the model source ↗
For full training, also budget gradients, optimizer states, activations and communication buffers. The inference weight estimate above is insufficient.
Tokens per second, TTFT and TPOT are not certified for this hardware in our reference. Use the benchmark checklist to compare an exact model, software revision and workload.
Nearby memory capacities, different tradeoffs
These are comparison candidates selected by memory capacity, not performance rankings or drop-in replacements.
AWS Trainium2
AWS’s own training accelerator, also used for supported inference. It runs the Neuron stack rather than a CUDA binary.
AWS Inferentia2
An inference-focused AWS chip with two NeuronCore-v2 cores. Inf2 instances can contain up to twelve chips.
Sources and review scope
Manufacturer and cloud documentation checked 2026-09-12. We reviewed specifications and the stated product boundaries; we did not run training or serving benchmarks on this device. Cloud memory and availability can differ by configuration.
- AWS Neuron Trainium3 architecture ↗ · checked 2026-09-12
Cite this reference
AI Infra Interviews. AWS Trainium3: specifications and workload fit. Reviewed . https://aiinfrainterviews.com/hardware/trainium3
Include your access date when citing a changing specification. Link to the specification section for hardware figures or the explanation for a sizing or workload decision.
Research method and limits
We compile manufacturer specifications, cloud documentation, model cards and serving recipes. The reference preserves source units and distinguishes individual devices from nodes and racks. Conflicting figures and unknown fields remain labelled.
Our contribution is the comparison, unit reconciliation, worked arithmetic and workload explanation. Published peaks are vendor specifications. Calculator results are estimates under the displayed assumptions. Neither is a measurement from our own accelerator lab.
For a manufacturer’s specification, consult the original source documents. Cite our page when using its analysis, and retain primary-source attribution for underlying figures. This curated reference does not establish market share, live availability or a universal performance ranking.
Found a discrepancy? Send a correction with the page URL, exact variant, disputed figure and supporting primary document.