🚨 Google versus Nvidia – who wins the real performance test? 🚨
Can Google really "steal" 10% of Nvidia's revenue? 🤔
Think it's that simple? ❓
Let's dive into the data that shows who is leading the AI race! 📊
The secret revealed: efficiency versus cost – Artificial Analysis' performance test 🔥
Forget the big statements for a moment and focus on what really matters to large language model (LLM) operators:
How much does it cost them to generate a token? ⛏️
Artificial Analysis put the hardware giants to a brutal test:
Google TPU v6e (Trillium), AMD MI300X, and NVIDIA H100/B200 chips were tested for efficiency, and the results show a clear picture, and the winner is not necessarily who you expected:
'Tokens per dollar' efficiency (inference core cost metric): Nvidia achieved 5 times (!) higher efficiency compared to Google TPU v6e. ✔️ Nvidia was 2 times more efficient than AMD MI300X. ✔️
The numbers speak: cost of generating one million tokens
The most critical test is the total cost per million input and output tokens, while maintaining high speed (30 output tokens per second per query). In this metric, NVIDIA H100 and B200 systems recorded a significantly lower total cost than their competitors. 🤔
Cost comparison (per million tokens) - Llama 3.3 70B model:
Based on Artificial Analysis' performance test (running vLLM, at a speed of 30 output tokens per second per query), these are the costs for generating one million tokens: ⚖️
NVIDIA H100 chip: The lowest cost is $1.06 per million tokens.
AMD MI300X chip: The cost more than doubles, reaching $2.24 per million tokens.
Google TPU v6e chip: Last place, with a cost of $5.13 per million tokens – almost 5 times more expensive than the H100.
Do we all understand now why, for now, with a strong emphasis on for now, Nvidia is still dominant? 🛡
The analysis shows that if you run a large AI model in the cloud (and pay by usage hours), Google's TPU v6e is 5 times more expensive than Nvidia's H100 for generating the same amount of information! 📊
Although Google and AMD are pushing hard, in terms of performance-per-dollar for everyday use with dominant AI models, NVIDIA still sets a bar that is hard to break. 👑
This efficiency is at the core of its competitive advantage in the cloud and inference market. 💎
Think about it: 💭
As long as Nvidia can offer hardware that generates tokens at half the price or even a quarter of the price of competitors, could its dominance strengthen? 🏎
The fundamental question remains open – can this huge cost gap close quickly, or will Nvidia remain the undisputed queen of AI⁉️