Has the CPU become the real bottleneck of the artificial intelligence revolution? π
A new study by Georgia Tech in collaboration with Intel shows us that the bottleneck created by the CPU may be even greater than the problems of photonics.
They present a crazy statistic: π
This CPU shortage situation only worsens as AI chips become more powerful and better.
How did they test this? π
The researchers took real-world workloads, such as internet search agents, research assistants, and coding agents, and tracked each step of task execution to understand where the system bottlenecks.
The results show a significant difference between system types: π
In a system using cheaper graphics cards (GPU), the CPU was responsible for 38% of the system's total waiting time.
But when the exact same tasks were run on the powerful H200 chips, the CPU's share of delays jumped to 65%. π
What's the logic? β
Here's the explanation: π‘
This happens because a faster graphics card simply finishes its calculations much earlier, so the CPU remains the 'weak link' that needs to handle all other agent operations.
This is precisely what makes its waiting time appear even greater compared to the GPU's fast performance. β³
The most interesting part emerging from the conclusions of this study? π
Every time a company upgrades its GPU, it essentially has to go back and buy more powerful CPUs to balance the system.
And precisely because of this, demand for CPUs is only expected to continue growing as technology advances.
This is likely what's behind the report from Intel's CFO, who noted that unmet demand for CPUs is already reaching amounts starting with B (billions).
We are monitoring. π