For a long time, AMD's graphics card sales have been beaten by NVIDIA, and the gap between the two sides is very large, and the gap in the past two years has even become wider, and many friends don't understand why.
For example, AMD's Radeon RX 7600's gaming performance can completely catch up with NVIDIA's GeForce RTX 4060, and ** is also 30 US dollars cheaper, the overall cost performance of AMD graphics cards is obviously higher, why is the sales always at a disadvantage?
Because graphics performance is not limited to, not exclusively, gaming performance, but also includes productivity performance. For example, all kinds of first-class codecs, various industrial 3D animation rendering software, and the now popular artificial intelligence computing, these are the major shortcomings of AMD graphics cards.
Solutions for this kind of application are often integrated with software and hardware, and NVIDIA started and developed relatively early in this regard, and many of its technologies and solutions have now become de facto technical standards in many professional fields.
For example, the current field of artificial intelligence, including computing hardware, software, and third-party AI frameworks (such as TensorFlow), is based on NVIDIA's CUDA solution, which is obviously very detrimental to AMD (and other graphics card manufacturers such as Intel).
Nowadays, almost all artificial intelligence projects and development books will give priority to the use of NVIDIA graphics cards, which do not support AMD graphics cards, or are much less efficient, or even cannot be installed and run directly.
In this context, only some consumer-level users will choose to buy AMD graphics cards, mainly for gaming and entertainment, which is no problem. In addition, the vast majority of business users will give preference to Nvidia's graphics cards, even if they are much higher, which has caused a huge disparity in the sales of AMD graphics cards and Nvidia graphics cards.
At present, artificial intelligence is the general direction of future technology development, which is no longer limited to commercial-level application scenarios, and has begun to slowly penetrate into the ordinary consumer market. For example, the current mainstream mobile phone processors are equipped with NPU artificial intelligence computing units, and the rumored Windows 12 will also be reconstructed with artificial intelligence, and the future desktop processors will also develop in this direction.
This development trend is now very clear, and AMD has also begun to actively make efforts in the field of artificial intelligence in recent years, and has begun to catch up with NVIDIA. For example, AMD has added AI computing units to the last two generations of mobile processors.
However, these efforts are still far from enough, and aside from the AI computing power of AMD GPUs, it must also come up with a set of efficient, excellent, and acceptable solutions for the industry, similar to NVIDIA's CUDA computing solutions.
In this regard, the solution proposed by AMD is "ROCM", the full name in English is "Radeon Open Compute Platform", and the Chinese name is Radeon Open Source Computing Platform, which is an open source project.
Based on GPU computing, through an open programming model and standard APIs, the Xi solution enables developers to use AMD's GPU resources for efficient data processing and computing, and has a wide range of application prospects in the fields of scientific computing, deep Xi and data analysis.
However, the performance of ROCM has been tepid since its launch, but AMD has not given up and has been working hard to improve and improve, and AMD released ROCM 60, currently open source***
rocm 6.0In addition to supporting AMD's latest Instinct Mi300A Mix300X AIGPU, there are several highlights:
First, the performance of low-precision mathematical calculations and attention algorithms has been improved. 2. The new HiPSPralsParselt library can accelerate AI operations through AMD's sparse matrix core technology. 3. Added support for libraries such as Deepspeed, Onnx-RT, and Cupy.
4. Support various mainstream AI frameworks, such as TensorFlow, Jax, and PyTorch. 5. Provide additional and comprehensive technical support to developers, including pre-packaged HPC and AI ML frameworks available on the AMD Infinity Hub** with new and improved ROCM development documentation and resources.
For ROCM 60, AMD is full of confidence, claiming ROCM 60 has reached the same level as CUDA in terms of large language model training.
Objectively speaking, it is unrealistic for AMD to catch up with Nvidia in the field of artificial intelligence in the short term, after all, Nvidia has been accumulating and cultivating in this field for many years. But now AMD has begun to fire at full power, although there is still a gap, but to a certain extent, it will form greater competitive pressure on NVIDIA, and this is always conducive to the market and technological progress.
As for ROCM 6What is the actual performance of 0?What are the prospects?It is still difficult to estimate whether it will be accepted by large commercial users, and it remains to be seen, and it is hoped that AMD will be stable and far-reaching.