China’s AI Startup shocked the global high-tech industry when deploying Deepseek-V3 in a part of the AI model competition costs in December 2024. The open source model was developed for less than $ 6 million in just two months, exceeding the latest AI models of Openai and Meta over multiple benchmarks. A few weeks later, on January 23, the company introduced Deepseek-R1, a reasoning model positioned as a direct competition partner of O1 of O1.
Deepseek’s claim is doubtful, but this development has been challenging the long -standing belief that the construction of a sophisticated AI model requires billions of dollars and the NVIDIA GPU army. Aiming to set up a large -scale data center in the United States, we are questioning Openai, Oracle, SoftBank, and MGX for a large -scale Stargate Project of $ 500 billion in venture.
DeepSeek trained models using about 2,000 NVIDIA H800 GPUs. This is less than tens of thousands of people deployed by major players. By focusing on the efficiency of algorithms and the collaboration of open source, the company has achieved results comparable to Openai’s ChatGpt at a few minutes of cost.
The market reaction was immediate. NVIDIA’s shares suffered a historic decline, sliding nearly 17 %, and wiped out market value for about $ 600 billion. This is the largest loss in US stock market history. The shock wave has forced many people to reconsider the need for large -scale AI infrastructure expenditures and the dependence on the industry’s top hardware.
Some analysts argue that DeepSeek’s success can easily access AI development, significantly reduce costs, and reconstruct industry competition.
Can I execute AI and major language models without NVIDIA GPUS?
NVIDIA H100 Tensol Core GPU (Credit: NVIDIA)
Yes, the simple thing is NVIDIA GPU is not the only option to execute the AI model. They are widely used for training and deployment of complex machine learning models, but have other options. For example, AMD GPU provides competitive performance and is attracting attention in AI R & D.
Beyond the conventional GPU, the AI model can be executed with a custom AI accelerator designed by companies such as CPU, TPU (tensor processing unit), FPGA (field programable gate array), and even Apple, Google, Tesla. 。 Hardware choices depend on factors such as model complexity, speed requirements, power consumption, and budget restrictions.
NVIDIA GPUs are popular for optimized software acoat systems, including CUDA and CUDNN, which are seamlessly operating with frameworks such as TensorFlow and Pytorch for large -scale deep learning training. However, alternatives such as small tasks, reasoning workloads, or high -energy development can be executable options such as AMD ROCM, Intel AI Accelerator, Google TPU, etc.
If you are developing an AI model, it is worth considering the best hardware for your needs, rather than defaulting to NVIDIA, just because it is the best -known choice.
Is GPU still important for AI?
Deepseek’s breakthrough makes important questions: AI model really needs NVIDIA GPU? If not, what does this mean for companies such as NVIDIA, AMD, and Broadcom? The simple answer is that you don’t always need NVIDIA GPU or any GPU to execute the AI model. This example is a neural magic, an AI startup that enables deep learning without special hardware.
GPU.
For many years, the GPU was a reliable option to train deep learning models, but it may not be essential for all AI workloads. The bigger problem is whether or not the GPU is as ordinary as today’s CPU, and whether NVIDIA will take the risk of CISCO’s route, or not, before competition. Whether it was dominated.
There is no simple answer. The GPU is still important for high -performance AI, but the efficiency advances can change the balance, making it easier for AI to access it without a huge amount of hardware investment.
CPU vs GPU: What is the difference?
What is a CPU?
The CPU or the central processing unit is the core of the computer operation built to process tasks step by step. Designed for flexibility, ideal for execution of operating systems, managing input and output, and processing general computing workloads. Using some powerful cores, the CPU is excellent in sequential processing, but it is difficult to generate a large number of calculations at the same time. As LifeWire points out, the CPU is suitable for tasks that require accurate execution, but is not optimized for the large parallel processing required for the latest AI models.
What is a GPU?
The GPU or graphics processing unit was originally designed to render graphics, but became an important tool for AI development. Unlike the CPU, the GPU is packed with hundreds or thousands of small cores that can process multiple calculations at once. This function, which handles a large amount of data in parallel, is especially useful for machine learning and deep learning applications. Reddit discussions among AI researchers emphasize how to make the complex model of training faster and more by simultaneously processing a huge amount of matrix operations.
Why AI training depends on the GPU
Parallel processing
AI models, especially deep neural network training, include millions or billions of calculations. The GPU is built to efficiently process these workloads and spreads to many cores to speed up the process. On the other hand, the CPU needs to process each step 1 at a time, and the training time is longer. With a GPU, training is done faster, and researchers and developers can repeat the model more quickly.
Optimized AI framework
Many of today’s top machine learning libraries, including TensorFlow and Pytorch, are built to utilize GPU acceleration. These frameworks do not only benefit the hardware itself. It also comes with a built -in tool that streamlines the development of AI models. This is a major reason that researchers and companies tend to invest a large amount of GPU -based setup.
If you don’t need GPU
Printing and training
The GPU plays an important role in training AI models, but it is not always necessary to execute these models in actual applications. When the model is trained, the scenario where the speed is not particularly prioritized can make predictions on the CPU without significant deceleration. For lightweight tasks, CPUs are more cost -effective and more energy -efficient choices.
Initial stage development
If you are working on a small project or simply experimenting, you may not need a powerful GPU. Many cloud platforms provide access to the GPU, so developers can start the CPU and upgrade as needs increase. In some cases, a small model may be executed on the CPU without a major performance problem, but the execution time may be slow.
Alternative hardware
Beyond CPU and GPU, there are other options, such as tensor processing unit (TPU) and special inference chips. The TPU developed by Google is optimized for certain AI workloads and may exceed the GPU. However, most developers continue to rely on GPUs for flexibility because of their unique setup requirements.
Cost and energy efficiency
GPU is powerful, but expensive and consumes a lot of energy. Companies with severe budgets or companies with sustainability may need to compare the advantages of sticking to a high -performance CPU or hybrid approach and using GPUs. If the application does not require large -scale parallel processing, the CPU may be a more practical choice.
Hardware suitable for jobs
Whether you need a GPU depends on the nature of the AI workload.
Large AI model training -GPU is almost indispensable if you use a large dataset and complex neural network. Improving CPU speed may be important. Predictions on executed -For small models and inference tasks, CPUs often process workloads without large deceleration. Balance between budget and performance -If the cost or power consumption is concerned, the combination of CPU and GPU may be the best approach.
The correct choice depends on the needs of a specific project, the goal of performance, and the available resources. AI hardware decisions continue to evolve, but understanding what the CPU and GPU bring to the table helps your work to make the best decision. Even if you start with a CPU or go all -in with a GPU, the key to gaining the best results is to match hardware to the project request.
The following is a video about how to execute a large language model on a unique local machine.
Source link
