Side-by-side, with the numbers that matter.
Two rigs, one table. Higher numbers win in cobalt; TDP and price flip.
Rig A
Alibaba Qwen2.5-72B-Instruct (Q4 quant)
Alibaba · cloud-model
Change
- 1X Neo
- Agility Robotics Digit
- Alibaba Qwen2.5-72B-Instruct (Q4 quant)
- AMD Instinct MI300X
- AMD Instinct MI325X
- AMD Instinct MI355X
- Anthropic Claude Opus 4.7
- Apple Mac Pro (M2 Ultra, 2023)
- Apple Mac Studio (M3 Ultra)
- Apple MacBook Pro 16-inch, M4 Max (2024)
- Apptronik Apollo
- Aurora
- AWS Trainium2
- Boston Dynamics Atlas
- Boston Dynamics Spot
- Cerebras Wafer-Scale Engine 3
- El Capitan
- Figure 02
- Frontier (OLCF-5)
- Google Cloud TPU v5p
- Groq Language Processing Unit
- Intel Gaudi 3 HL-325L
- MSI GeForce RTX 5090 Suprim Liquid
- NVIDIA B100
- NVIDIA B200
- NVIDIA DGX B200
- NVIDIA DGX H200
- NVIDIA DGX Spark
- NVIDIA Eos DGX SuperPOD
- NVIDIA GB200
- NVIDIA GB200 NVL72
- NVIDIA GeForce RTX 4090 Founders Edition
- NVIDIA GeForce RTX 5070 Ti
- NVIDIA GeForce RTX 5080 Founders Edition
- NVIDIA GeForce RTX 5090 Founders Edition
- NVIDIA GH200 Grace Hopper
- NVIDIA H100
- NVIDIA H200
- NVIDIA HGX B200
- NVIDIA Jetson AGX Thor
- NVIDIA RTX PRO 4500 Blackwell
- NVIDIA RTX PRO 5000 Blackwell
- NVIDIA RTX PRO 6000 Blackwell
- OpenAI GPT-5
- SambaNova SN40L Reconfigurable Dataflow Unit
- Sanctuary AI Phoenix
- Tesla Optimus
- The Stargate Project
- Unitree B2
- Unitree G1
- Unitree H1
Rig B
AWS Trainium2
AWS · cloud-model
Change
- 1X Neo
- Agility Robotics Digit
- Alibaba Qwen2.5-72B-Instruct (Q4 quant)
- AMD Instinct MI300X
- AMD Instinct MI325X
- AMD Instinct MI355X
- Anthropic Claude Opus 4.7
- Apple Mac Pro (M2 Ultra, 2023)
- Apple Mac Studio (M3 Ultra)
- Apple MacBook Pro 16-inch, M4 Max (2024)
- Apptronik Apollo
- Aurora
- AWS Trainium2
- Boston Dynamics Atlas
- Boston Dynamics Spot
- Cerebras Wafer-Scale Engine 3
- El Capitan
- Figure 02
- Frontier (OLCF-5)
- Google Cloud TPU v5p
- Groq Language Processing Unit
- Intel Gaudi 3 HL-325L
- MSI GeForce RTX 5090 Suprim Liquid
- NVIDIA B100
- NVIDIA B200
- NVIDIA DGX B200
- NVIDIA DGX H200
- NVIDIA DGX Spark
- NVIDIA Eos DGX SuperPOD
- NVIDIA GB200
- NVIDIA GB200 NVL72
- NVIDIA GeForce RTX 4090 Founders Edition
- NVIDIA GeForce RTX 5070 Ti
- NVIDIA GeForce RTX 5080 Founders Edition
- NVIDIA GeForce RTX 5090 Founders Edition
- NVIDIA GH200 Grace Hopper
- NVIDIA H100
- NVIDIA H200
- NVIDIA HGX B200
- NVIDIA Jetson AGX Thor
- NVIDIA RTX PRO 4500 Blackwell
- NVIDIA RTX PRO 5000 Blackwell
- NVIDIA RTX PRO 6000 Blackwell
- OpenAI GPT-5
- SambaNova SN40L Reconfigurable Dataflow Unit
- Sanctuary AI Phoenix
- Tesla Optimus
- The Stargate Project
- Unitree B2
- Unitree G1
- Unitree H1
0 of 31 fields comparable
⇄ swap sides| Spec | Alibaba Qwen2.5-72B-Instruct (Q4 quant) | AWS Trainium2 |
|---|---|---|
| Bf16 Tflops PER Chip | — | 667 |
| Chips PER Trn2 Instance | — | 16 |
| Chips PER Ultraserver | — | 64 |
| Context Window Tokens K | 128 | — |
| Dual Card Count | 2 | — |
| Dual Card Vram GB | 24 | — |
| FP8 Dense Petaflops PER Chip | — | 1.3 |
| FP8 Sparse Petaflops PER Chip | — | 5.2 |
| GPT Version | 5 | — |
| HBM Bandwidth TB S PER Chip | — | 2.9 |
| HBM Capacity GB PER Chip | — | 96 |
| HBM Type | — | HBM3 |
| Neuroncores PER Chip | — | 8 |
| Parameter Count B | 72 | — |
| Qwen Version | 3 | — |
| Qwen25 72B Instruct | Qwen2.5-72B-Instruct | — |
| Qwen25 Version | 2.5 | — |
| Qwen3 Largest Dense Checkpoint B | 32 | — |
| RTX 5090 Pairing Vram GB | 32 | — |
| Status Checked | 2026-09-24 | — |
| Trainium3 GA | — | 2025-12-02 |
| Trainium3 HBM Bandwidth TB S | — | 4.9 |
| Trainium3 Hbm3e GB PER Chip | — | 144 |
| Trn2 Instance FP8 Petaflops | — | 20.8 |
| Trn2 Instance HBM TB | — | 1.5 |
| Trn3 Ultraserver FP8 Pflops | — | 362 |
| Trn3 Ultraserver MAX Chips | — | 144 |
| Ultraserver FP8 Petaflops | — | 83.2 |
| Ultraserver Total HBM Bandwidth TB S | — | 185 |
| Ultraserver Total HBM TB | — | 6 |
| Vram Required GB | 44 | — |