Etched vs NVIDIA: Inside the $800M Faster, Cheaper “Transformer-Only” Chip #transformers #ai #etched

6 Просмотры
Издатель
Two Harvard dropouts just raised $800 million at a $5 billion valuation to do the “impossible”: take on NVIDIA.

Gavin Uberti and Robert Wachen founded Etched on a radical conviction: the Transformer architecture has already won, and AI hardware should finally admit it. Instead of another general‑purpose GPU, Etched built Sohu - a Transformer‑only ASIC that burns the model’s computation pattern directly into silicon.

In this video, we break down:

Why NVIDIA’s H100 is a “Swiss Army knife” and why that flexibility is now a generality tax
How only 3.3% of H100 transistors are doing the core matrix math LLMs actually use
What Transformers, Inference, and FLOPs really mean in practical, data‑center terms
How Low Voltage Inference (LVI) lets Sohu run trillion‑parameter models at 80%+ of peak FLOPs without thermal throttling
How Cluster Scale Memory (CSM) slashes chip‑to‑chip latency and treats a rack like a single shared memory pool
Why an 8‑chip Sohu server delivering 500,000+ tokens/sec on Llama‑70B could displace ~160 NVIDIA H100s
The binary risk: if the world moves beyond Transformers, Sohu becomes e‑waste
#AI #ArtificialIntelligence #MachineLearning #DeepLearning #Transformers #LLM #AIChips #Semiconductors #Nvidia #Etched #Sohu #GPUs #ASIC #TechNews #TechExplained #TechAnalysis #FutureOfAI #AIRevolution #AIEconomy #DataCenters #CloudComputing #Inference #AIModels #GenerativeAI #ChatGPT #Claude #GeminiAI #SoraAI #Startup #VentureCapital #SiliconValley #HardwareEngineering #ChipDesign #TSMC #HighPerformanceComputing #HPC #AIInfrastructure #AIEthics #Automation #Superintelligence #EdgeComputing #AIInnovation #BigTech #AIStartup #AIResearch #TechYouTube #EducationalVideo #ExplainerVideo #Innovation #howtoai #entrepreneur #siliconvalley #science #peterthiel #andrejkarpathy #h100 #news #inference #flop
Категория
AI для YouTube & Shorts
Выполните вход или зарегистрируйтесь для отправки комментария.
Комментариев нет.