AMD/Google technology could combine AI acceleration with conventional computing.
Google is reportedly working with AMD on the design of a next generation Tensor Processing Unit (TPU) aimed at workloads that increasingly combine artificial intelligence acceleration with conventional computing.
According to a number of industry reports, the companies are exploring a hybrid AI accelerator that could integrate CPU cores directly into the TPU processor package. The design would be intended in part for reinforcement learning and emerging agentic AI workloads, where AI accelerators need to interact more closely with general-purpose processors.
The integration of TPU and CPU resources into one package is not a new concept. AMD and later Intel integrated GPU technology with their CPUs and have done so for some time. But the GPU that is in the CPU is considerably less powerful than a discreet GPU. It’s not known if the TPU integrated with the AMD CPU package would be equivalent to a discrete TPU card.
However, it will meet much faster performance. Google uses its TPU processors now in servers with standard X86 CPU processors, but there is a considerable lag in communication between the two chips as the data has to transfer from one physical chip to another, according to the company.
With the CPU and TPU in one package, there is considerably less distance to be covered and a potential great reduction in latency. How much is unclear at this time.
That approach could be particularly useful for reinforcement learning, in which an AI system repeatedly generates actions, evaluates the results and uses that feedback to improve its behavior. Such workloads can involve frequent interaction between accelerator and CPU resources, potentially making data movement and communication overhead significant.
The concept also fits Google’s broader push toward hardware optimized for increasingly sophisticated AI systems. Google has already been developing TPUs for more than a decade TPU family is now in its eighth- generation.
Google is also reportedly developing other specialized AI hardware. A project known as “Frozen v2” is said to be targeting highly specialized AI inference and could potentially deliver substantially higher token efficiency than Google’s conventional TPUs.
For its part, AMD has been pursuing increasingly integrated AI systems for some time. The company recently announced a partnership with Cerebras to combine AMD EPYC processors and Instinct accelerators with Cerebras’ wafer-scale AI processors in a system designed to divide AI workloads between different types of processors.
For AMD, meanwhile, a role in Google’s next-generation TPU program would provide another indication that its CPU and chip-design technologies are becoming relevant to the rapidly expanding custom AI accelerator market.
One analyst approved of the proposed alliance. “I think it makes good sense. AMD has more than proven that they’re good at custom, high-performance, and affordable APUs, and that they are a trustworthy partner, and Google needs diversity,” said Jon Peddie, president of Jon Peddie Research.
Peddie’s referring to AMD’s custom video game silicon. More than a decade ago, the console makers wanted to get away from using their own specialized processors and switch to X86 CPU and DirectX GPU technology to be compatible with PC gaming.
AMD won because Intel didn’t have a GPU and NVIDIA didn’t have a CPU. For several years, the CPU/GPU combination being used by Microsoft Xbox and Sony PlayStation was the only thing keeping AMD alive. It still dominates video game consoles with its SoC technology.
