Google has expanded its eighth-generation Tensor Processing Unit family with two specialised systems, giving Nvidia another headache in the increasingly competitive accelerator market.
The search outfit has introduced the TPU 8i Max and TPU 8tx, extending a processor family that already includes the TPU 8t and TPU 8i. The new designs focus on memory bandwidth, processing density, and hardware security, while Google’s decision to use AMD processors in one system adds an interesting twist.
The TPU 8i Max is the more muscular of the two. Google has taken a pair of its existing TPU 8i chips and combined them into a single package, effectively doubling the available computing resources without requiring twice the rack space.
The company claims the arrangement delivers twice the processing power, twice the on-chip SRAM and twice the high-bandwidth memory throughput of a standard TPU 8i. It targets enormous AI workloads where moving data can become more expensive than processing it.
According to Tom’s Hardware, the resulting package could theoretically deliver approximately 20.2 petaflops of FP4 performance and 17.2TB/s of memory bandwidth, based on doubling the published specifications of the original TPU 8i. Google has yet to publish complete specifications for the new system, so those figures remain estimates.
The other newcomer, TPU 8tx, takes a different approach. It uses the same underlying silicon as Google’s training-focused TPU 8t but concentrates on protecting sensitive information while it is being processed.
Google says the system uses hardware isolation to protect data and model parameters while computations are underway. The technology is aimed at governments and regulated industries that might otherwise be reluctant to hand sensitive information to shared computing infrastructure.
More interestingly, Google has dumped its Arm-based Axion host processors in favour of fifth-generation AMD EPYC chips for the TPU 8tx. The move gives AMD a place inside Google’s specialised accelerator systems, although the company has given no indication that the arrangement extends to the rest of the family.
The TPU 8tx uses Google’s Boardfly interconnect, which was developed for low-latency inference workloads. Google claims the configuration can reduce latency by up to 50 per cent compared with the TPU 8t, with almost no performance penalty from its additional security features.
This means customers could theoretically train smaller models, fine-tune existing ones and run inference on the same system while keeping sensitive information protected. Google is pitching that flexibility to organisations that would otherwise need separate computing environments for different stages of development.
The two additions bring Google’s eighth-generation TPU range to four systems, despite the company introducing the first pair only in April. Google argues that building different systems around existing silicon allows it to respond to changing requirements much faster than designing an entirely new processor for every workload.
The outfit has been trying to make its custom hardware more attractive to developers accustomed to Nvidia’s CUDA ecosystem. Its recent TorchTPU software allows PyTorch applications and Hugging Face models to run on Google’s accelerators, while support for inference frameworks including vLLM and SGLang reduces the amount of specialised programming required.
That software work could prove just as important as the hardware. Nvidia has spent years making its CUDA platform difficult for developers to abandon, and simply producing a powerful chip does little good if customers have to rewrite their applications to use it.
Google claims its TPU operation has attracted nearly 100 customers, extending beyond the company’s internal Gemini workloads. Its willingness to supply accelerator systems directly to customer data centres suggests it wants a larger share of the market currently dominated by Nvidia.
The timing is interesting given the turmoil surrounding Nvidia following Jensen Huang’s departure. Google has spent more than a decade developing its own accelerator technology, and the latest additions give customers more options as Nvidia’s leadership and future strategy come under scrutiny with shedloads of scrute.
Google has not disclosed pricing, complete technical specifications or shipping dates for either system. Until independent performance testing becomes available, its claims of improved efficiency and processing density remain promises from a company with a considerable interest in persuading customers to buy its hardware.
