AI Training-Dedicated ASIC Complete Guide
AI training ASICs (Application-Specific Integrated Circuits) are custom chips purpose-built for AI training, distinct from general-purpose GPUs. They trade flexibility for higher energy efficiency and better per-unit compute cost.
Mainstream AI Training ASIC Comparison
| Model | Vendor | Process | Compute (BF16) | Memory | Interconnect | Availability |
|---|
| TPU 8t (Trillium 2 Training) | Google | 3nm | ~3,500 TFLOPS | 216GB HBM | 3D Torus + Axion CPU | Google Cloud |
| TPU 8i (Trillium 2 Inference) | Google | 3nm | ~5,500 TFLOPS | 288GB HBM | 3D Torus | Google Cloud |
| Google TPU v7 (Ironwood) | Google | 5nm | 2,307 TFLOPS | 192GB HBM | 3D Torus, 9,216 Pod | Google Cloud |
| Google TPU v6e (Trillium) | Google | 5nm | 918 TFLOPS | 32GB HBM | 2D Torus, 256 Pod | Google Cloud |
| Google TPU v5p | Google | 5nm | 459 TFLOPS | 95GB HBM | 3D Torus, 8,960 Pod | Google Cloud |
| AWS Trainium 3 (Trn3) | Amazon | 3nm | 1,300 TFLOPS | 144GB HBM | NeuronLink-v4, 144 UltraServer | AWS Cloud (2025-12 GA) |
| AWS Trainium 2 | Amazon | 4nm | 667 TFLOPS | 96GB HBM | NeuronLink, 64 UltraServer | AWS Cloud |
| AWS Trainium 1 | Amazon | 7nm | 191 TFLOPS | 32GB HBM | NeuronLink, 16 cluster | AWS Cloud |
| Intel Gaudi 3 | Intel | 5nm | 1,835 TFLOPS | 128GB HBM2e | 24× 200GbE | Commercial |
| Intel Gaudi 2 | Intel | 7nm | 432 TFLOPS | 96GB HBM2e | 24× 100GbE | Commercial |
Google TPU Series Evolution
| Gen | Name | Compute (BF16) | HBM | Interconnect | Primary Use |
|---|
| v4 | — | 275 TFLOPS | 32GB | 3D Torus | Training |
| v5p | — | 459 TFLOPS | 95GB | 3D Torus | Training |
| v5e | — | 197 TFLOPS | 16GB | 2D Torus | Inference |
| v6e | Trillium | 918 TFLOPS | 32GB | 2D Torus | Training/inference |
| v7 | Ironwood | 2,307 TFLOPS | 192GB | 3D Torus | Inference-first |
| 8t | Trillium 2 Training | ~3,500 TFLOPS | 216GB | 3D Torus + Axion CPU | Training-dedicated |
| 8i | Trillium 2 Inference | ~5,500 TFLOPS | 288GB | 3D Torus | Inference-dedicated |
Selection Guide
By Cloud Provider
- Google Cloud: TPU v5p / v6e / v7 Ironwood / TPU 8t (training) + 8i (inference) split (2026-04)
- AWS: Trainium 3 (2025-12 GA, 3nm) / Trainium 2
- On-premises / private cloud: Intel Gaudi 3 (open standard Ethernet)
By Scale
- Ultra-large scale (trillion parameters): TPU 8t (216GB) + Cerebras WSE-3 / WSE-4
- Large scale (10B+ parameters): TPU v6e, Gaudi 3, Trainium 2/3
- Medium scale (1B+ parameters): TPU v5e, Gaudi 2, Trainium 1
- 400B+ model training: Trn3 UltraServer (144 chips, 52 PFLOPS FP8)
Key Advantages vs GPU
- Energy efficiency: 2-3× advantage in performance per watt
- Per-unit compute cost: 30-50% advantage
- Interconnect scale: 8,000+ chip Pods
- Custom architecture: Avoids the waste of GPU general-purpose overhead
Key Disadvantages
- Software ecosystem maturity: CUDA still dominates
- Vendor lock-in: TPU is Google Cloud only
- Model migration cost: Requires re-optimization
Detailed Product Pages