需要指出训练,尤其是pre-training是data center scale的。 如果是fine-tuning,比如LoRA,bottleneck也主要是throughput,通常也需要多张GPU进行训练。 Unified memory主要的优势是在inference,也就是locally可以就跑一些较大的模型,例如llama-13B。
@doggielovelilyc12 күн бұрын
llama-13b叫做什么优势,我现在本地跑 llama-90B...
@初一-o9sАй бұрын
雖然我不是很懂,但從顯卡切入的觀點很有趣,也引發其他觀眾在下面分享觀點,覺得收穫良多! 我想參加!
@彼得森Ай бұрын
大家認為 15 萬台幣($5,000)的Macbook Pro,可以取代 139 萬台幣($45,000)的 Nvidia H100,做一些輕量級的 AI 訓練嗎?請懂的朋友一定留言告訴我。
@YSLinYSАй бұрын
硬體規格字面上的數字可以取代|應用上會碰到Jensen Huang設下的軟體護城河cuDNN, tensor-rt, pyTorch, tensorflow core~FYI. m2max 96G and i7 RTX3090 user(this is why that i still need a win PC). BTW 感謝分析apple notebook影片,始終是看您影片進行產品判斷依據。比a爹的內容多很多有用的資訊..
VRAM is good but not everything for AI... there is also Terra Operations Per Seconds (TOPS) where key and as far as I know M4max only has 38 TOPS and a 4080 RTX has 780 TOPS. If you want to play with AI wait for Nvidia's project DIGITS.