AWS Trainium3
Amazon's first 3 nm AI chip: 144 GB of HBM3E at 4.9 TB/s and 2,517 MXFP8 TFLOPS per chip, with 144 chips per Trn3 UltraServer. Generally available since December 2025, with customers including Anthropic. Sources and verification dates.
Every published figure
Source per value- Owner The company that specified the chip and runs it in its own cloud, data centers or products.
- Amazon Web Services 1
- Role What the owner says the chip is for: training, inference, both, or a general-purpose server CPU.
- Training and inference 1
- Design partner The ASIC house that co-designed or physically implemented the chip (Broadcom, Marvell, MediaTek, Alchip, GUC and others), as stated by the owner, the partner, a filing or named trade-press reporting. 'In-house' when the owner states it designed the chip itself, including through an acquired team such as Annapurna Labs.
- Alchip; Marvell 4
- Partner basis Who says so: the owner, the partner, or press reporting with the outlet named. A reported partner is not confirmed by the owner.
- Reported by TrendForce (October 2025): Trainium3 developed jointly with Alchip and Marvell. AWS has not confirmed. 4
- Node Process node as the owner or foundry states it, for example 'TSMC N3' or '7 nm'.
- 3 nm 2
- Memory On-package memory type (HBM2e, HBM3, HBM3E, LPDDR5) or, for CPUs, the memory the socket drives (DDR5).
- HBM3E 3
- Memory capacity GB On-package memory per chip in GB. Blank for CPUs, whose memory sits off-package.
- 144 3
- Memory bandwidth TB/s Peak memory bandwidth per chip in TB/s, as published.
- 4.9 3
- BF16 TFLOPS Published BF16 throughput per chip or accelerator module in TFLOPS. FP16 is a separate column. See the precision note for sparsity and module scope.
- 671 1
- MXFP8 TFLOPS Published MXFP8 throughput per chip or accelerator module in TFLOPS. See the precision note for shared-format figures, sparsity and module scope.
- 2,517 1
- MXFP4 TFLOPS Published MXFP4 throughput per chip or accelerator module in TFLOPS. See the precision note for shared-format figures, sparsity and module scope.
- 2,517 1
- Precision note What the throughput figures mean: dense or sparse, which precision, per chip or per package, and any figure the owner states only in another unit.
- Per chip: 2,517 TFLOPS MXFP8 and MXFP4, 671 BF16/FP16/TF32 TFLOPS (AWS Neuron docs; dense or sparse not labelled for these). A 144-chip UltraServer delivers 362 FP8 PFLOPS (AWS figure). 1
- Largest pod chips The largest scale-up domain the owner states: chips per pod, UltraServer, supernode or rack joined by the chip-to-chip interconnect.
- 144 2
- Status Announced: public roadmap without verified deployment. Sampling: first silicon or limited internal use. Preview: customers can request early access, without a GA claim. Deployed: stated production use or general availability. Discontinued: program stopped.
- Deployed 2
- Announced Year the owner first announced the chip publicly.
- 2,024 5
- Deployment scale The largest deployment the owner or a named report states: chip counts, clusters, gigawatts or regions.
- EC2 UltraClusters 3.0 connect up to 1 million Trainium chips (AWS, December 2025). 2
- How to get it How an outside buyer can use the chip: a named cloud instance or service, a product, or 'Internal only'.
- AWS EC2 Trn3 UltraServers; Amazon Bedrock 2
Not verified in the sources reviewed: foundry, fp16, mx8, mx4, fp8, int8, fp4, power.
Sources
- AWS Neuron docs: Trainium3 architecture accessed 2026-11-11
- Amazon: Trainium3 UltraServers now available accessed 2026-11-11
- AWS What's New: Amazon EC2 Trn3 UltraServers accessed 2026-11-11
- TrendForce: AWS's Trainium3 draws strong interest, ramping early 2026 accessed 2026-11-11
- Amazon: re:Invent 2024 announcements include Trainium3 accessed 2026-11-11