Skip to content

Commit ddb15d1

Browse files
AseelOmerAseelOmer
committed
fixed links
1 parent ef1fa8f commit ddb15d1

1 file changed

Lines changed: 98 additions & 149 deletions

File tree

‎commercial_models/models.md‎

Lines changed: 98 additions & 149 deletions
Original file line numberDiff line numberDiff line change
@@ -1,3 +1,7 @@
1+
<!-- markdownlint-disable MD013 -->
2+
<!-- markdownlint-disable MD025 -->
3+
<!-- Disabled MD025 because multiple top-level headings (#) are needed for each model section -->
4+
15
# Model 1: GPT-4 (OpenAI)
26

37
## Model Name and Provider
@@ -14,252 +18,197 @@ Cloud infrastructure uses global data centers; regions are not public.
1418

1519
### Estimated Energy (Inference)
1620

17-
Published or estimated per-query energy values vary between studies.
21+
Published or estimated per-query energy values vary between studies.
1822
Representative numbers include:
1923

20-
**Epoch AI (2024):** ≈ 0.3 Wh (0.0003 kWh) per ChatGPT/GPT-4 query.
21-
24+
**Epoch AI (2024):** ≈ 0.3 Wh (0.0003 kWh) per ChatGPT/GPT-4 query.
2225
Source: [Epoch AI – How Much Energy Does ChatGPT Use?][epoch-ai].
2326

24-
Other analysts estimate ≈ 0.3 – 1.8 Wh (0.0003 – 0.0018 kWh)
27+
Other analysts estimate ≈ 0.3 – 1.8 Wh (0.0003 – 0.0018 kWh)
2528
depending on prompt length, token output, and GPU hardware.
2629

27-
Sources: “The Carbon Footprint of ChatGPT,” media analyses.
28-
2930
**Caveat:** OpenAI does not publish per-query energy data.
3031
All estimates depend on assumptions such as:
3132

32-
* Hardware type (GPU vs TPU)
33-
* Power Usage Effectiveness (PUE)
34-
* Data center region and carbon intensity
35-
* Prompt and token length
33+
- Hardware type (GPU vs TPU)
34+
- Power Usage Effectiveness (PUE)
35+
- Data center region and carbon intensity
36+
- Prompt and token length
3637

3738
### Training Energy (GPT-4)
3839

39-
Some analyses extrapolate GPT-4’s training energy from model size and
40-
compute budget:
41-
42-
≈ 51 – 62 GWh (51 772 500 – 62 318 750 kWh) for full-scale training.
40+
Some analyses extrapolate GPT-4’s training energy from model size and compute budget:
4341

42+
≈ 51 – 62 GWh (51 772 500 – 62 318 750 kWh) for full-scale training.
4443
Source: [The Carbon Footprint of ChatGPT][sustainability-numbers].
4544

4645
These are indirect estimates, not official OpenAI disclosures.
4746

48-
### Model Size (GPT-4)
49-
50-
Estimated model size: **≈ 1.8 trillion parameters** (widely reported
51-
estimate; OpenAI has not publicly confirmed exact parameter count).
47+
### Model Size
5248

49+
Estimated model size: **≈ 1.8 trillion parameters**
50+
(widely reported estimate; OpenAI has not publicly confirmed exact parameter count).
5351
Source: SemiAnalysis and other architecture analyses.
5452

55-
### Water Usage (GPT-4)
53+
### Water Usage
5654

5755
Official data are unavailable, but media analyses suggest:
5856

59-
A single ChatGPT query may indirectly consume ≈ 0.5 L of water,
60-
depending on data-center cooling.
61-
62-
Generating a 100-word email may use ≈ 0.14 kWh energy and 0.52 L water.
57+
- A single ChatGPT query may indirectly consume ≈ 0.5 L of water.
58+
- Generating a 100-word email may use ≈ 0.14 kWh energy and 0.52 L water.
6359

6460
Source: [The Verge – Sam Altman on ChatGPT Energy and Water Use][verge-gpt].
6561

66-
### PUE and CI Context (GPT-4)
62+
### PUE and CI Context
6763

6864
Studies multiply compute energy by:
6965

70-
* **PUE** – Power Usage Effectiveness (total facility power / IT power)
71-
* **CI** – Carbon Intensity (kg CO₂e / kWh electricity)
66+
- **PUE** – Power Usage Effectiveness (total facility power / IT power)
67+
- **CI** – Carbon Intensity (kg CO₂e / kWh electricity)
7268

7369
Example assumptions:
7470

75-
* **PUE:** ≈ 1.1 – 1.3 for Azure hyperscale centers
76-
* **CI:** ≈ 0.3 – 0.4 kg CO₂e / kWh (depending on region)
71+
- **PUE:** ≈ 1.1 – 1.3 for Azure hyperscale centers
72+
- **CI:** ≈ 0.3 – 0.4 kg CO₂e / kWh (depending on region)
7773

7874
---
7975

80-
## Model 2: Claude Haiku (Anthropic)
76+
# Model 2: Claude 3 Haiku (Anthropic)
8177

82-
### Model Name & Provider
78+
## Model Name & Provider
8379

8480
**Claude 3 Haiku**, developed by **Anthropic**.
8581

8682
### Model Description
8783

8884
Part of Anthropic’s Claude 3 family (Haiku, Sonnet, Opus).
89-
Released March 2024. Smallest and fastest model for low-latency,
90-
energy-efficient inference in chat, summarization, and automation.
85+
Released March 2024. Smallest and fastest model for low-latency, energy-efficient inference in chat, summarization, and automation.
9186

9287
Source: [Anthropic Blog – Claude 3 Technical Overview][anthropic-blog].
9388

94-
### Model Size / Architecture
95-
96-
Estimated model size: **≈ 7 billion parameters** (Haiku variant,
97-
optimized for efficiency and low-latency inference).
89+
### Model Size and Architecture
9890

91+
Estimated model size: **≈ 7 billion parameters** (Haiku variant, optimized for efficiency and low-latency inference).
9992
Source: public model reports and community discussions.
10093

10194
### Hosting & Deployment
10295

103-
Hosted via Anthropic API and **Amazon Bedrock (AWS)**.
96+
Hosted via Anthropic API and **Amazon Bedrock (AWS)**.
10497
These centers maintain **PUE ≈ 1.2**.
10598

106-
Sources: [AWS Bedrock Claude Integration], [AWS Sustainability Report 2024][aws-report].
99+
Sources: [AWS Bedrock Claude Integration][aws-bedrock], [AWS Sustainability Report 2024][aws-report].
107100

108101
### Estimated Energy
109102

110-
Anthropic does not publish per-query energy data.
111-
Independent analysts estimate ≈ 0.05 – 0.1 Wh (0.00005 – 0.0001 kWh)
112-
per query based on token count and GPU efficiency.
103+
Anthropic does not publish per-query energy data.
104+
Independent analysts estimate ≈ 0.05 – 0.1 Wh (0.00005 – 0.0001 kWh) per query based on token count and GPU efficiency.
113105

114-
Claude 3 Haiku is ≈ 5× faster and more efficient than larger Claude 3
115-
models.
106+
Claude 3 Haiku is ≈ 5× faster and more efficient than larger Claude 3 models.
116107

117-
Sources: [Epoch AI – Energy Use of AI Models]Sources:
118-
[epoch-ai-training], [Anthropic Claude 3 Announcement].
108+
Sources: [Epoch AI – AI Training Compute and Energy Scaling][epoch-ai-training], [Anthropic Claude 3 Announcement][anthropic-announcement].
119109

120110
### Training Energy
121111

122-
Claude 3 models are trained on GPU clusters (NVIDIA A100/H100) primarily
123-
hosted on AWS infrastructure.
124-
For models in the 10–30B parameter range, training energy is typically
125-
3,000–10,000 MWh.
112+
Claude 3 models are trained on GPU clusters (NVIDIA A100/H100) primarily hosted on AWS infrastructure.
113+
For models in the 10 – 30 B parameter range, training energy is typically **3 000 – 10 000 MWh**.
126114

127-
Sources: [Epoch AI – AI Training Compute & Energy Scaling],
128-
[Anthropic Responsible Scaling Policy].
115+
Sources: [Epoch AI – AI Training Compute and Energy Scaling][epoch-ai-training], [Anthropic Responsible Scaling Policy][anthropic-policy].
129116

130-
### Water Usage
117+
### Water Usage of claude
131118

132-
Anthropic has not published specific water consumption figures for the
133-
Claude 3 family.
134-
As it relies on AWS data centers, cooling water use is managed under AWS
135-
sustainability strategy.
136-
AWS data centers in cooler regions use air cooling to reduce water
137-
footprint, while others recycle water on-site.
119+
Anthropic has not published specific water consumption figures for the Claude 3 family.
120+
As it relies on AWS data centers, cooling water use is managed under AWS sustainability strategy.
121+
AWS data centers in cooler regions use air cooling to reduce water footprint, while others recycle water on-site.
138122

139-
Sources: [AWS Water Stewardship Report][aws-water],
140-
[Anthropic Sustainability Commitments].
123+
Sources: [AWS Water Stewardship Report][aws-water], [Anthropic Sustainability Commitments][anthropic-sustainability].
141124

142-
### PUE and CI Context
125+
### PUE & CI Context
143126

144-
AWS’s average PUE: ~1.2 (accounts for cooling and power delivery losses).
145-
Carbon intensity (CI): ~0–0.2 kg CO₂e/kWh, depending on regional renewable
146-
mix.
147-
AWS aims for 100% renewable energy by 2025, lowering emissions over time.
127+
AWS’s average **PUE ≈ 1.2** (accounts for cooling and power delivery losses).
128+
Carbon intensity (CI): ≈ 0 – 0.2 kg CO₂e / kWh, depending on regional renewable mix.
129+
AWS aims for 100 % renewable energy by 2025.
148130

149-
Sources: [AWS Global Infrastructure Efficiency Data],
150-
[Anthropic Responsible Scaling Policy][anthropic-policy].
131+
Sources: [AWS Global Infrastructure Efficiency Data][aws-efficiency], [Anthropic Responsible Scaling Policy][anthropic-policy].
151132

152133
---
153134

154-
## Model 3: Gemini Nano (Google)
135+
# Model 3: Gemini Nano (Google)
155136

156-
### Provider
137+
## Model Name / Provider
157138

158139
**Gemini Nano**, developed by **Google DeepMind**.
159-
Smallest member of Gemini family (Nano, Pro, Ultra).
140+
Smallest member of the Gemini family (Nano, Pro, Ultra).
160141

161-
### Hosting
142+
### Hosting / Deployment
162143

163144
Runs on-device via **Android AICore** (subsystem introduced 2023).
164-
Designed for mobile hardware like Pixel 8 Pro and Pixel 9.
145+
Designed for mobile hardware such as Pixel 8 Pro and Pixel 9.
165146
Reduces energy use by eliminating cloud compute and network load.
166147

167-
Sources: [Google AI Blog – Introducing Gemini][google-blog],
168-
[Android Developers – Gemini Nano Overview][android-dev],
169-
[The Verge – Gemini Nano on Pixel 8 Pro][verge-gemini].
170-
171-
### Estimated Model Size / Architecture
148+
Sources: [Google AI Blog – Introducing Gemini][google-blog], [Android Developers – Gemini Nano Overview][android-dev], [The Verge – Gemini Nano on Pixel 8 Pro][verge-gemini].
172149

173-
Gemini Nano variants (device-optimized):
150+
### Model Size / Architecture
174151

175-
* **Nano-1:** ≈ 1.8 billion parameters
176-
* **Nano-2:** (larger device variant) ≈ 3.25 billion parameters
152+
Gemini Nano variants (device-optimized):
177153

178-
These use quantized weights tuned for on-device inference.
154+
- **Nano-1:** ≈ 1.8 billion parameters
155+
- **Nano-2:** ≈ 3.25 billion parameters
179156

180-
Source: device benchmark reports and public model parameter listings.
157+
These use quantized weights tuned for on-device inference.
158+
Source: device benchmark reports and public parameter listings.
181159

182-
### Estimated Energy (Inference) gemini
160+
### Estimated Energy of gemini
183161

184162
No official values.
185-
Device benchmarks show ≈ 0.01 Wh (0.00001 kWh) per query —
186-
10 – 30× more efficient than GPT-4.
163+
Device benchmarks show ≈ 0.01 Wh (0.00001 kWh) per query — 10 – 30× more efficient than GPT-4.
187164

188-
Sources: [Google Pixel AI Benchmarks (2024)],
189-
[Epoch AI – How Much Energy Does ChatGPT Use][epoch-ai].
165+
Sources: [Google Pixel AI Benchmarks (2024)][google-pixel-ai], [Epoch AI – How Much Energy Does ChatGPT Use][epoch-ai].
190166

191-
### Training Energy Estimates
167+
### Training Energy of gemini
192168

193-
Gemini Nano was distilled from larger Gemini models trained on **TPU v5e**
194-
clusters.
195-
Training energy for Nano ≈ 200 – 1,200 MWh (≈ 1–5% of Gemini Ultra’s
196-
training compute).
169+
Gemini Nano was distilled from larger Gemini models trained on **TPU v5e** clusters.
170+
Training energy for Nano ≈ 200 – 1 200 MWh (≈ 1 – 5 % of Gemini Ultra’s training compute).
197171

198-
Sources: [Google Research – Efficient TPU Training (2024)],
199-
[Google Cloud Sustainability Report (2024)].
172+
Sources: [Google Research – Efficient TPU Training (2024)][google-tpu-paper], [Google Cloud Sustainability Report (2024)][google-cloud-sustainability].
200173

201-
### Water Usage (Nano)
174+
### Water Usage of gemini
202175

203-
Inference uses no data-center water since it runs locally on devices.
204-
Training used Google data centers with Water Usage Effectiveness (WUE)
205-
≈ 0.18 L/kWh.
176+
Inference uses no data-center water since it runs locally on devices.
177+
Training used Google data centers with Water Usage Effectiveness (WUE) ≈ 0.18 L/kWh.
206178
Google targets net-positive water impact by 2030.
207179

208-
Sources: [Google Environmental Report (2024)],
209-
[Bloomberg – Google AI Water Consumption (2024)].
180+
Sources: [Google Environmental Report (2024)][google-env-report], [Bloomberg – Google AI’s Thirst for Water][bloomberg-water].
210181

211-
### PUE & CI Context
182+
### PUE / CI Context
212183

213-
Google Data Centers report average PUE ≈ 1.10–1.12.
214-
Carbon Intensity (CI) ≈ 0.15 kg CO₂e / kWh due to 70%+ renewable energy mix.
184+
Google data centers report average **PUE ≈ 1.10 – 1.12**.
185+
Carbon Intensity (CI) ≈ 0.15 kg CO₂e / kWh due to 70 %+ renewable energy mix.
215186
On-device execution uses < 5 W of mobile power per inference.
216187

217-
Sources: [Google Data Center Efficiency Overview (2024)],
218-
[Google TPU v5e Efficiency Blog (2024)].
188+
Sources: [Google Data Center Efficiency Overview (2024)][google-efficiency], [Google TPU v5e Efficiency Blog (2024)][google-tpu-blog].
219189

220190
---
221191

222-
[azure-blog]:
223-
https://azure.microsoft.com/en-us/blog/introducing-gpt4-in-azure-openai-service/
224-
[epoch-ai]:
225-
https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use
226-
[sustainability-numbers]:
227-
https://www.sustainabilitybynumbers.com/p/carbon-footprint-chatgpt
228-
[verge-gpt]:
229-
https://www.theverge.com/2023/1/18/energy-water-chatgpt
230-
[anthropic-blog]:
231-
https://www.anthropic.com/blog/claude3-overview
232-
[aws-report]:
233-
https://aws.amazon.com/about-aws/sustainability/
234-
[anthropic-policy]:
235-
https://www.anthropic.com/responsible-scaling
236-
[aws-water]:
237-
https://aws.amazon.com/about-aws/sustainability/#water
238-
[google-blog]:
239-
https://blog.google/technology/ai/google-gemini-ai/
240-
[android-dev]:
241-
https://developer.android.com/ai/gemini-nano
242-
[verge-gemini]:
243-
https://www.theverge.com/2023/12/6/23990823/google-gemini-ai-models-nano-pro-ultra
244-
[AWS Bedrock Claude Integration]:
245-
https://aws.amazon.com/bedrock/
246-
[Anthropic Claude 3 Announcement]:
247-
https://www.anthropic.com/news/claude-3-models
248-
[epoch-ai-training]:
249-
https://epoch.ai/gradient-updates/ai-training-compute-energy-scaling
250-
[Anthropic Sustainability Commitments]:
251-
https://www.anthropic.com/sustainability
252-
[AWS Global Infrastructure Efficiency Data]:
253-
https://aws.amazon.com/about-aws/sustainability/
254-
[Google Pixel AI Benchmarks (2024)]:
255-
https://ai.google/discover/pixel-ai/
256-
[Google Research – Efficient TPU Training (2024)]:
257-
https://arxiv.org/abs/2408.15734
258-
[Google Cloud Sustainability Report (2024)]:
259-
https://sustainability.google/reports/environmental-report-2024/
260-
[Bloomberg – Google AI Water Consumption (2024)]:
261-
https://www.bloomberg.com/news/articles/2024-02-13/google-ai-water-consumption-analysis
262-
[Google Data Center Efficiency Overview (2024)]:
263-
https://cloud.google.com/sustainability/data-centers
264-
[Google TPU v5e Efficiency Blog (2024)]:
265-
https://cloud.google.com/blog/products/ai-machine-learning/introducing-tpu-v5e
192+
[azure-blog]: https://azure.microsoft.com/en-us/blog/introducing-gpt4-in-azure-openai-service/
193+
[epoch-ai]: https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use
194+
[sustainability-numbers]: https://www.sustainabilitybynumbers.com/p/carbon-footprint-chatgpt
195+
[verge-gpt]: https://www.theverge.com/2023/4/19/openai-ceo-sam-altman-chatgpt-energy-water-use
196+
[anthropic-blog]: https://www.anthropic.com/news/claude-3-family
197+
[aws-bedrock]: https://aws.amazon.com/bedrock/
198+
[aws-report]: https://aws.amazon.com/about-aws/sustainability/
199+
[anthropic-announcement]: https://www.anthropic.com/news/claude-3-models
200+
[epoch-ai-training]: https://epoch.ai/gradient-updates/ai-training-compute-energy-scaling
201+
[anthropic-policy]: https://www.anthropic.com/news/responsible-scaling-policy
202+
[aws-water]: https://aws.amazon.com/about-aws/sustainability/#water
203+
[anthropic-sustainability]: https://www.anthropic.com/sustainability
204+
[aws-efficiency]: https://aws.amazon.com/about-aws/sustainability/
205+
[google-blog]: https://blog.google/technology/ai/google-gemini-ai/
206+
[android-dev]: https://developer.android.com/ai/gemini-nano
207+
[verge-gemini]: https://www.theverge.com/2023/12/6/23990823/google-gemini-ai-models-nano-pro-ultra
208+
[google-pixel-ai]: https://ai.google/discover/pixel-ai/
209+
[google-tpu-paper]: https://arxiv.org/abs/2408.15734
210+
[google-cloud-sustainability]: https://sustainability.google/reports/environmental-report-2024/
211+
[google-env-report]: https://sustainability.google/reports/environmental-report-2024/
212+
[bloomberg-water]: https://www.bloomberg.com/news/articles/2023-08-09/google-ai-s-thirst-for-water-could-leave-towns-dry
213+
[google-efficiency]: https://cloud.google.com/sustainability/data-centers
214+
[google-tpu-blog]: https://cloud.google.com/blog/products/ai-machine-learning/introducing-tpu-v5e

0 commit comments

Comments
 (0)