AMD and Intel Collaborate to Enhance the Role of CPUs in AI
In the world of Artificial Intelligence (AI) hardware, discussions often focus on graphics processing units (GPUs). However, the role of central processing units (CPUs) is often overlooked. Now, two major players in the semiconductor industry, AMD and Intel, are taking a bold step to bring attention back to CPUs through the launch of a new specification known as Advanced Compute Extensions (ACE).
Introducing Advanced Compute Extensions (ACE)
ACE is designed to provide a new approach to AI processing with the goal of improving efficiency on x86 architecture-based processors. Although this step is not intended to replace GPUs in large-scale AI training, ACE focuses more on smaller AI models, tasks requiring low latency, and situations where GPU usage is considered excessive.
In many cases, transferring data between the CPU and GPU can cause significant performance degradation, especially for workloads that require quick response times. By minimizing this data transfer process, ACE aims to avoid potential bottlenecks that often hinder system performance.
Technical Advantages of ACE
Technically, ACE focuses on matrix multiplication, which is at the core of many AI operations. Although CPUs have previously handled this math with AVX instructions, their efficiency was limited because those instructions were not designed for heavy matrix workloads. Here are some advantages of ACE:
- Familiar Structure: ACE maintains the existing AVX10 register structure with 512-bit input, while adding dedicated hardware for matrix operations. This means developers do not need to completely change data formats.
- Improved Performance: ACE can perform up to 16 times more operations compared to AVX10 for certain vector input sets. While this does not mean all applications will run 16 times faster, it indicates increased instruction efficiency.
- Energy Efficiency: With better instruction efficiency, ACE also leads to reduced power consumption and memory bandwidth load, making it ideal for edge computing or single-user applications.
Benefits for Developers
For developers, ACE offers much-needed consistency. This specification is designed without being tied to a specific implementation, making it easier for developers using popular frameworks like PyTorch and TensorFlow. They no longer need to worry about matching code with various AVX support differences.
Additionally, ACE highlights some weaknesses of Neural Processing Units (NPUs). Although NPUs are becoming more common, existing standards remain fragmented, and offloading workloads to NPUs can sometimes create new complications depending on the hardware used. This is where ACE provides a solution by keeping certain tasks on the CPU to ensure speed and simplicity.
Conclusion: The Evolution of CPUs in the AI World
Although the ACE specification will not dethrone GPUs as the king in large-scale AI training, this innovation signals that CPU evolution continues. With the right architecture, x86 processors can now handle a broader and more efficient range of AI workloads than before. This step shows that CPUs have an important role that should not be overlooked in the ever-growing AI ecosystem.
Source: https://inet.detik.com/consumer/d-8543884/amd-dan-intel-garap-ace-buat-tantang-dominasi-nvidia-di-ai
