A New Kind of AI Chip Is Shaking Things Up
The AI hardware race has been dominated by Nvidia for years, especially with its powerful and expensive AI focused GPUs. But that might be about to change. A Chinese startup and Google are both pushing a different type of processor for AI tasks, and their success could eventually benefit PC gamers.
These new chips are called tensor processing units or TPUs. They are a type of ASIC, which stands for application specific integrated circuit. Unlike general purpose GPUs that can handle a wide mix of workloads, ASICs are designed from the ground up to do a narrow set of jobs extremely well.
We have seen this story before. In the early days of bitcoin mining, people used normal GPUs. Then purpose built mining ASICs appeared, which were dramatically faster and more efficient, and they pushed GPUs out of that market. Now a similar transition might be starting in the AI world.
The Chinese TPU That Beats Nvidia A100 on Efficiency
A Chinese startup called Zhonghao Xinying claims to have created a new AI chip named Ghana that can outclass Nvidia’s A100 AI GPU from 2020 in key ways. According to the company, the Ghana chip delivers around 1.5 times the performance of the A100 while cutting power usage by a massive 75 percent.
What makes this especially interesting is that the Ghana chip is reportedly manufactured using older Chinese chip making technology. China’s domestic manufacturing processes are generally behind the cutting edge nodes used by TSMC, Intel, and Samsung. So on paper this chip should not be able to compete with modern high end GPUs. Yet, by being highly specialized for AI workloads like tensor operations, it can still achieve impressive performance per watt.
The raw performance crown still belongs to Nvidia’s latest Blackwell based AI GPUs, not the A100. However, AI companies care deeply about two things:
- How much performance they get per dollar.
- How much performance they get per watt of power.
If a specialized ASIC like Ghana is much cheaper to produce and far more power efficient, it becomes very attractive for huge AI clusters where electricity and hardware costs are enormous.
There is a catch. Transitioning away from Nvidia is not simple. The AI ecosystem is heavily invested in Nvidia hardware, CUDA, and its software stack. Models, tools, and workflows are all tuned for Nvidia GPUs. Moving to a new platform means rewriting and optimizing code, retraining people, and dealing with compatibility issues. That is painful in the short term.
However, with Nvidia reportedly charging around 45,000 to 50,000 dollars for each high end B200 GPU, many companies are motivated to look for alternatives that reduce long term costs.
Google’s TPUs and the Push Beyond Nvidia
On the other side of the world, Google has been building its own TPUs since 2017 for internal use in its data centers. These chips are also ASIC style processors aimed at AI workloads like training and running large models.
Now, Google is reportedly considering selling its TPUs directly rather than just renting access through its cloud services. According to reports, Google is in talks with several major customers, including a potential multibillion dollar deal with Meta. Google is said to be targeting roughly 10 percent of Nvidia’s AI revenue by getting more companies to adopt its TPU platform.
Google’s TPUs are more narrowly focused and are likely more power efficient than many of Nvidia’s general purpose AI GPUs. If big players like Meta start using them at scale, that could be a serious signal that the market is ready to diversify away from Nvidia’s hardware monopoly.
Still, just like with the Chinese Ghana chip, adoption is not trivial. Companies need tools, libraries, and support. Google will have to convince customers that switching or adding TPUs to their stack is worth the effort compared to just continuing to buy Nvidia GPUs.
Why PC Gamers Should Care
At first glance this might sound like data center industry drama that has nothing to do with gaming rigs. But there is an important connection that could affect the price of your next graphics card.
Right now, the same cutting edge manufacturing capacity that produces high end gaming GPUs is also heavily used to build AI GPUs. The huge demand for AI hardware has pushed up prices and strained supply. Gaming GPUs end up costing more because they share the same advanced silicon processes and production lines as data center AI GPUs.
If AI workloads start moving from general purpose GPUs to specialized ASICs and TPUs, it could change that equation. Here is what could happen over time:
- Less demand for the very latest GPU class silicon from data centers.
- More manufacturing capacity available for gaming focused GPUs.
- Reduced pricing pressure on consumer graphics cards.
In other words, if ASIC based AI chips grab a significant slice of the AI market, the supply and demand dynamics for GPU production could shift in favor of gamers. It is not guaranteed, and it will not be overnight, but it is a realistic outcome.
There is another possible side effect. If Nvidia faces stronger competition in the AI space from Google, Chinese startups, and others, it might start to value its gaming business more again. For years, GPU designs and marketing have increasingly been driven by AI and data center needs rather than pure gaming performance and value. A more balanced revenue mix could give Nvidia fresh incentives to push harder on gaming features, pricing, and availability.
For now, these AI ASICs are mostly a story about big tech companies fighting over the future of artificial intelligence. But if they succeed, PC gamers might finally see some relief on GPU prices and a renewed focus on gaming as a first class priority.
Original article and image: https://www.pcgamer.com/hardware/graphics-cards/could-a-new-generation-of-dedicated-ai-chips-burst-nvidias-bubble-and-do-for-ai-gpus-what-asics-did-for-crypto-mining/
