文字轉視訊革命需要龐大的 GPU 庫

AI 生成影片的出現引發了 GPU 需求的激增，估計需要 72 萬個高階 Nvidia H100 GPU 來支援 TikTok 和 YouTube 等平台上的創作者社群。此類模型的訓練過程需要龐大的運算能力，甚至超過了 GPT-4 和靜態影像生成。隨著採用率的增長，推理需求將進一步超過訓練，需要大量的硬體投資才能使文字到影片生成成為主流。

Text-to-Video Revolution Hinges on a Vast GPU Arsenal: Millions Required

文字轉影片的革命取決於龐大的 GPU 庫：需要數百萬

The groundbreaking capabilities of text-to-video generators, exemplified by OpenAI's Sora, have sparked a surge of excitement and investment in Artificial Intelligence (AI). However, the realization of mainstream AI-generated video hinges on the availability of an unprecedented number of specialized processing units known as Graphics Processing Units (GPUs).

以 OpenAI 的 Sora 為代表的文本轉視頻生成器的突破性功能引發了人們對人工智慧 (AI) 的熱情和投資。然而，主流人工智慧產生視訊的實現取決於數量空前的專用處理單元（稱為圖形處理單元（GPU））的可用性。

Estimates by Factorial Funds reveal that a staggering 720,000 high-end Nvidia H100 GPUs would be necessary to cater to the vast creator communities of platforms like TikTok and YouTube. The gargantuan scale of this computational demand is evident in Sora's training process, which requires up to 10,500 powerful GPUs for a month and can generate only about 5 minutes of video per hour per GPU during inference.

Factorial Funds 的估計顯示，為了滿足 TikTok 和 YouTube 等平台龐大的創作者社群的需求，需要 72 萬個高階 Nvidia H100 GPU。這種運算需求的巨大規模在 Sora 的訓練過程中顯而易見，一個月需要多達 10,500 個強大的 GPU，並且在推理過程中每個 GPU 每小時只能產生約 5 分鐘的影片。

As illustrated by the data, training text-to-video models demands significantly more computing power than training massive language models like GPT4 or even generating static images. Moreover, as AI models like Sora gain wider adoption, the computation required for inference - the process of using the trained model to generate new videos - will eclipse the computation required for initial training.

數據表明，訓練文字轉視頻模型比訓練 GPT4 等大規模語言模型甚至生成靜態圖像需要更多的計算能力。此外，隨著像 Sora 這樣的人工智慧模型得到更廣泛的採用，推理所需的計算（使用經過訓練的模型產生新影片的過程）將超過初始訓練所需的計算。

To put this into perspective, Nvidia shipped approximately 550,000 H100 GPUs in 2023. Statistics from Statista indicate that the twelve largest customers utilizing Nvidia's H100 GPUs collectively possess 650,000 of these cards, with tech giants Meta and Microsoft accounting for 300,000 of them.

從這個角度來看，Nvidia 在2023 年出貨了大約55 萬張H100 GPU。Statista 的統計數據顯示，使用Nvidia H100 GPU 的12 大客戶總共擁有65 萬張此類卡，其中科技巨頭Meta 和微軟佔據了30 萬張。

Assuming a cost of $30,000 per card, the astronomical sum of $21.6 billion would be required to procure the GPUs necessary to make Sora's vision of AI-generated text-to-video mainstream. This figure nearly matches the entire market capitalization of AI tokens at present.

假設每張卡的成本為 30,000 美元，則需要 216 億美元的天文數字來採購必要的 GPU，以使 Sora 的人工智慧生成文字到影片的願景成為主流。這個數字幾乎與目前AI代幣的總市值相當。

The procurement of such a vast quantity of GPUs presents a formidable challenge due to supply constraints and escalating demand for AI-driven applications. Notably, Nvidia is not the sole player in the GPU market. Rival chipmaker AMD offers competitive products, with investors generously rewarding the company for its innovations, propelling its stock from the $2 range in 2012 to over $175 today.

由於供應限制和人工智慧驅動應用程式的需求不斷上升，採購如此大量的 GPU 是一項艱鉅的挑戰。值得注意的是，Nvidia 並不是 GPU 市場的唯一參與者。競爭對手晶片製造商 AMD 提供具有競爭力的產品，投資者慷慨地獎勵該公司的創新，推動其股價從 2012 年的 2 美元區間升至如今的 175 美元以上。

Alternative options exist for outsourcing computing power to GPU farms, such as Render (RNDR) and Akash Network (AKT). However, the majority of GPUs on these networks consist of retail-grade gaming GPUs, significantly less potent than Nvidia's server-grade H100 or AMD's comparable offerings.

存在將運算能力外包給 GPU 場的替代選擇，例如渲染 (RNDR) 和 Akash 網路 (AKT)。然而，這些網路上的大多數 GPU 都是零售級遊戲 GPU，其效能明顯低於 Nvidia 的伺服器級 H100 或 AMD 的同類產品。

Despite these challenges, the tantalizing promise of text-to-video generation, as envisioned by Sora and other protocols, will necessitate a herculean hardware investment. While the potential for revolutionizing the creative workflow in industries like Hollywood is undeniable, it is crucial to temper expectations regarding the imminent mainstream adoption of AI-generated video.

儘管存在這些挑戰，但正如 Sora 和其他協議所設想的那樣，文字到影片生成的誘人前景將需要巨大的硬體投資。雖然徹底改變好萊塢等行業創意工作流程的潛力是不可否認的，但降低對人工智慧生成影片即將成為主流的預期至關重要。

At present, the industry faces a pressing need for a substantial increase in chip production to meet the burgeoning demand for AI-driven applications. Until this supply gap is addressed, the widespread availability of AI-generated video will remain an elusive goal.

目前，業界迫切需要大幅增加晶片產量，以滿足人工智慧驅動應用不斷增長的需求。在解決這一供應缺口之前，人工智慧產生影片的廣泛可用性仍將是一個難以實現的目標。