Skip to main content

AWS Trainium 4

Product Overview​

AWS Trainium 4 is AWS's fifth-generation custom AI accelerator, first publicly previewed on 2025-12-02 during the opening keynote of re:Invent 2025. According to AWS, Trainium 4 delivers 6x FP4 performance, 3x FP8 performance, 4x memory bandwidth, and 2x HBM capacity versus Trainium 3, with delivery beginning in 2027.

The most significant structural change announced is the interconnect strategy. Jointly with NVIDIA at re:Invent 2025, AWS confirmed that Trainium 4 will support NVIDIA NVLink Fusion high-speed interconnect and work alongside the NVIDIA MGX rack architecture — meaning Trainium 4 can be co-located with NVIDIA GPUs within the same rack. This is widely read as AWS acknowledging that NVIDIA's networking ecosystem has become the default customer expectation, while keeping high-margin workloads on its own ASICs. This is the first product of a multigenerational NVIDIA–AWS NVLink Fusion collaboration.

Critical disclosure boundary: to date, AWS has never officially disclosed Trainium 4's process node, absolute HBM capacity in GB, absolute memory bandwidth in TB/s, or any absolute compute figure in PFLOPS. All public information consists of relative multiples versus Trainium 3. Widespread reports that it will use HBM4 are supply-chain rumors, not official AWS statements, and are not treated as fact here. Every field requiring an absolute number is marked "not disclosed" on this page rather than filled with estimates.

Trainium 4 already has substantial capacity pre-booked, and AWS is simultaneously in talks to sell physical Trainium chips directly to external enterprises.

Key Specifications​

ItemSpecification
ArchitectureNot disclosed
ProcessNot disclosed
MemoryNot disclosed
Memory TypeNot disclosed
Memory BandwidthNot disclosed
FP8Not disclosed
FP4Not disclosed
FP16Not disclosed
FP32Not disclosed
INT8Not disclosed
TDPNot disclosed
InterconnectNVIDIA NVLink Fusion (co-location with NVIDIA GPUs / MGX racks)
Release / ProductionFirst previewed 2025-12-02 (re:Invent 2025); delivery expected from 2027

⚠️ Why almost the entire table reads "not disclosed": AWS's public statements on Trainium 4 are entirely relative multiples (6x the FP4 performance of Trainium 3, 3x FP8, 4x memory bandwidth, 2x HBM capacity) with no absolute specifications whatsoever. Process node, HBM capacity in GB, absolute TB/s, absolute PFLOPS, and TDP are all missing. The "not disclosed" cells above carry no multiples, so downstream data pipelines do not mistake them for absolute metrics.

Multiplying Trainium 3's 2.52 PFLOPS by these multiples to derive figures such as "15.1 PFLOPS FP8" would be self-derived arithmetic rather than official data. This page deliberately avoids that conversion so as not to contaminate cross-chip compute comparisons.

The "HBM4" claim currently appears only in supply-chain rumors and secondary reporting; AWS has not confirmed it, so memory type is likewise marked as not disclosed.

The one known interconnect fact: Trainium 4 supports NVIDIA NVLink Fusion and can be co-located with NVIDIA GPUs / MGX racks.

Key Features​

  • 6x FP4 performance over Trainium 3: the headline metric. Trainium 3's known weakness was precisely FP4 (a significant gap versus contemporary NVIDIA Blackwell), so the 6x gain signals AWS is prioritizing FP4 low-precision inference as the primary target workload
  • 3x FP8 performance over Trainium 3: the dominant precision for training workloads, improved 3x
  • 4x memory bandwidth / 2x HBM capacity: bandwidth gains exceed capacity gains, indicating a design emphasis on mitigating the memory wall rather than simply adding capacity — particularly relevant to long-context and MoE inference
  • NVLink Fusion enters AWS custom silicon for the first time: Trainium 4 will interoperate with NVIDIA Blackwell GPUs and NVIDIA CPUs within the same MGX rack
  • From internal cost-reduction tool to external product: AWS is in talks to sell physical Trainium chips directly to external companies, beyond cloud instance rental
  • Substantial pre-booked capacity: Trainium 3 was near full allocation shortly after shipping, and Trainium 4 capacity was booked well over a year before availability

Use Cases​

  • ✅ Large-model training and inference on AWS (EC2 Trn4 instances, expected 2027)
  • ✅ FP4 low-precision inference (the primary source of the performance gain)
  • ✅ Long-context / MoE-architecture inference (design prioritizes bandwidth 4x over capacity 2x)
  • ✅ Hybrid architecture deployments (co-located with NVIDIA GPUs to reduce full-migration risk)
  • ✅ AWS AI Factories (managed AI infrastructure deployed in customers' own data centers, able to host Trainium)
  • ❌ Procurement decisions requiring concrete specifications (no absolute compute, capacity, or process data)
  • ❌ Deployment before 2027 (official guidance is delivery starting 2027)

Comparison​

Note for cross-generation comparison: Trainium 3's absolute values are fully disclosed by AWS, while Trainium 4 offers only multiples. The two cannot be placed into the same absolute compute comparison table.

Data Sources​