Why Nvidia Soars 10%

My estimate is that the llama38b performance is similar to the ChatGPT3.5 performance. There are still 30b and 400b in this. Open AI and Maso have 150,000 GPUs each and only barely at the level of GPT4.

If OpenAI sells about 2 trillion won as a paid service, not to mention other companies’ paid LLM services. The INT8 has a gate size of 1/20 compared to the FPU32. Nvidia is not the only company that can use HBM.

It is true that Nvidia is faster than Intel and AMD Meta Tesla, but will it be able to maintain such an overwhelming performance gap in the future? Isn’t that questionable? If the difference between the INT8 and Nvidia and Neo-Wine is only process refinement… There are so many masters in the world.

Open AI is overwhelmed by Meta’s llama38b performance… If we don’t make a meaningful gap in GPT5, doesn’t it mean that open AI performance will deteriorate and eventually GPU sales will decrease?

Compared to H100, the announced H200 performance is so good that the H100 seems to be overflowing with the thought of buying H200 rather than buying H100. Isn’t it too early to announce the H200 performance?

It’s a difficult time. In the long run, fewer people are likely to be convinced that Nvidia will beat Apple. I think it will be … a baseless estimate.

라마3 #llama38b #3b #400b #Meta #openai #H100 #H200 #B100 #LLM #LMM #NPU #INT8 #GPU #FPU32