@ontariorob5150@plutos_eth lol plus prefill is slow as hell! This thing canโt even run Qwen 28b without it being painfully slow! Spark with CUDA for the win.
๐ wrong! This is basically repackaged old tech. Prefill is not up to par and it canโt compete with Spark and CUDA. This may be great for some things like RAG over private docs, batch inference, fine tuning, and chat style workloads where prefill is short. I donโt want similar token gen speeds that take 3 times longer to process.
Iโm extremely bullish on SpaceX over the long term. But if it IPOs at a multi-trillion-dollar valuation, I expect 6โ12 months of price discovery where Wall Street figures out what the company is actually worth. History shows that even exceptional companies like Facebook (now Meta) and Uber experienced significant post-IPO volatility before their long-term trajectories became clear. A great company doesnโt guarantee the stock wonโt need time to digest an aggressive valuation.
A great company doesn't automatically make a great stock purchase at any price. The market has to determine whether the valuation already reflects the future everyone is expecting.
@AMD@NeowinFeed Still not an efficient as Spark with CUDA. This will struggle with prefill but may be better for RAG over private docs, fine tuning, and chat style workloads where prefill is short.