AI accelerator chip: Evolution
Timeline
The milestones of the product, each with a source
| Year | Event |
|---|---|
| 1989 | Intel announces the 80170 ETANN, an analogue neural-network chip with 64 neurons and 10,240 synapses; no sale is recorded in the pages opened [25][26]. |
| 2015 | Google's tensor processing unit (TPU) is deployed in its own data centres [27]. |
| 2016 | NVIDIA's Tesla P100, with HBM2 memory stacks in the same package (CoWoS), is bought in bulk by hyperscalers, and the first DGX-1 ships [52][53]. |
| 2017 | NVIDIA announces the Tesla V100, with 21.1 billion transistors, 640 tensor cores, 16 GB of memory on the module and a 300 W design power; shipping was set for the third quarter [28]. |
| 2024 | NVIDIA announces Blackwell with 208 billion transistors on two reticle-limit dies [14]. |
| 2025 | Google's seventh-generation TPU, Ironwood, arrives with 192 GB of HBM3e memory and liquid cooling [39][40]; Gartner counts more than USD 200 billion of AI processor sales in the year [36]. |
| 2026 | NVIDIA reports USD 193.7 billion of data-centre revenue for its fiscal year [5]; India approves Semicon 2.0 and has onboarded over 38,000 GPUs under the IndiaAI Mission [24][23]. |
as of 2026-11-16
Lineage
The products and technologies it evolved from and into
Evolved from- GPU Graphics processing unit The data-centre GPU grew from the graphics chip: the Tesla V100 of 2017 added 640 tensor cores to a GPU die [28].
- Integrated circuit (chip) An early neural-network chip, Intel's ETANN of 1989, was an integrated circuit with 64 analogue neurons [25].
as of 2026-11-16
Recipe changes
How what it is made of changed over time, and what drove each change
| Change | From | To | Driver | Source |
|---|---|---|---|---|
| Memory on the package | 16 GB of memory on the Tesla V100 module, with an HBM2 interface (2017) | 80 GB of HBM3 on the H100 SXM and 192 GB of HBM3e on Google's Ironwood | Large models need more weights held close to the arithmetic units [28][4][39] | [28][4][39] |
| Number of logic dies | One die of 815 mm2 (V100) | Two reticle-limit dies joined by a 10 TB/s link (Blackwell) | A single die cannot grow past the reticle limit of 858 mm2 [28][14][43] | [28][14][43] |
| Power per package | 300 W (V100, 2017) | 700 W (H100 SXM) and 750 W (AMD MI300X) | More transistors and memory run faster, and cooling must remove the heat [28][4][51] | [28][4][51] |
| Cooling | Air cooling in the DGX H100 | Liquid cooling in Google's Ironwood | Not stated in the pages opened; Google says Ironwood gives 2x the performance per watt of its previous chip [40] | [21][40] |
as of 2026-11-16
What could improve it next
Directions for improvement, today's value and the limit
| Direction | Metric | Today | Limit | Unit | Source |
|---|---|---|---|---|---|
| Less embodied carbon per unit of compute | Embodied carbon intensity of Google TPUs | 38 | – | gCO2e per exaflop (TPU v6e; 114 for TPU v4i) | [3] |
| Stack chips in 3D instead of spreading them out in 2.5D | Rise in embodied emissions when compute units and HBM capacity are doubled | 141 | 89 | % increase (today 2.5D; limit shown is 3D) | [11] |
| Bigger interposers and more HBM stacks per package | Interposer size | two reticle-limit dies per GPU | 14 | reticle sizes (TSMC 2029 roadmap) | [43][14] |
| More work per watt | Performance per watt of Google's Ironwood against its predecessor | 2 | – | times | [40] |
as of 2026-11-16
Sources
- Life-cycle emissions of AI hardware: a cradle-to-grave approach and generational trends Schneider et al. (Google), arXiv preprint 2502.01671, 2025
- NVIDIA H100 Tensor Core GPU: specifications NVIDIA, 2026
- NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2026 NVIDIA investor relations, 2026
- 3D Integration Reduces Embodied Emissions Impact Of GPU Scaling (imec and VUB) Semiconductor Engineering, 2026
- NVIDIA Blackwell Platform Arrives to Power a New Era of Computing NVIDIA investor relations, 2024
- Introduction to NVIDIA DGX H100/H200 Systems (last updated 17 October 2026) NVIDIA, 2026
- IndiaAI Mission expands AI ecosystem with affordable compute and startup support Digital India, Government of India, 2026
- Cabinet approval of Semicon 2.0 (Prime Minister's Office page) Prime Minister's Office, Government of India, 2026
- Neural Network Chip Joins the Collection Computer History Museum, 2024
- John C. Dvorak on Intel's First Neural Network Chip The Chip Letter (Substack), 2024
- In-Datacenter Performance Analysis of a Tensor Processing Unit (Jouppi et al.) Google Research, 2017
- NVIDIA Announces Tesla V100 with Volta GPU at GTC 2017 PC Perspective, 2017
- By decade's end, AI will drive more than half of all chip sales (reporting Gartner) The Next Platform, 2026
- Google's Ironwood TPUs represent a bigger threat than Nvidia would have you believe The Register, 2025
- Google offers 7th-gen Ironwood TPUs for AI, with AI-inspired comparisons The Register, 2025
- TSMC's next-gen CoWoS roadmap: over 14-reticle packages and a 48x leap in compute power expected by 2029 Tom's Hardware, 2026
- AMD unveils Instinct MI300X GPU and MI300A APU, claims up to 1.6X lead over Nvidia's competing GPUs Tom's Hardware, 2023
- NVIDIA Tesla P100: technical overview NVIDIA, 2016
- Deep Learning Drives Nvidia's Tesla Business To New Highs The Next Platform, 2016