Scientists in China have developed a new chip with a feature: it is analog, meaning that it performs calculations on its own physical circuits, rather than using binary ones and zeros as in standard digital processors.
Moreover, its creators claim that the new chip is capable of outperforming top-end graphics processors (GPUs) from Nvidia and AMD by more than 1,000 times.
When solving complex communications problems, including matrix inversion problems used in massive multiple-input multiple-output (MIMO) systems (wireless technology systems), the chip achieved the accuracy of standard digital processors while consuming about 100 times less power.
By making the changes, the researchers said the device outperformed top-end GPUs such as the Nvidia H100 and AMD Vega 20 by more than 1,000 times. Both chips play an important role in training AI models; For example, Nvidia's H100 is a newer version of the A100 graphics card that OpenAI used to train ChatGPT.
The new device is built from arrays resistive RAM (RRAM) cells that store and process data by adjusting how easily electricity flows through each cell.
Unlike digital processors, which perform calculations in binary ones and zeros, an analog design processes information as continuous electrical currents passing through a network of RRAM cells. By processing data directly on its own hardware, the chip avoids the power-intensive task of transferring information between itself and an external memory source.
“With the rise of applications that use huge amounts of data, this poses challenges for digital computing, especially as scaling traditional devices becomes increasingly challenging,” the researchers said in the study. “Comparative analysis shows that our approach to analog computing can provide 1000 times higher throughput and 100 times better power efficiency than current digital processors at the same accuracy.”
Old technology, new tricks
Analog computing is nothing new, quite the contrary. Antikythera Mechanism, discovered off the coast of Greece in 1901, it is believed to have been built more than 2,000 years ago. He used interlocking gears to perform calculations.
For most modern computing historyHowever, analog technology was written off as an impractical alternative to digital processors. This is because analog systems rely on continuous physical signals—such as voltage or voltage—to process information. electric current. They are much more difficult to control precisely than the two stable states (1 and 0) that digital computers have to deal with.
Analogue systems are superior in speed and efficiency. Because they don't have to break up calculations into long lines of binary code, but rather represent them as physical operations in a chip's circuitry, analog chips can process large amounts of information simultaneously while using much less power.
This becomes especially important in data- and power-intensive applications such as artificial intelligence, where digital processors face limitations in the amount of information they can process sequentially, as well as in future 6G communications – where networks will have to process huge volumes of overlapping wireless signals in real time.
Recent advances in memory hardware could make analog computing viable again, the researchers said. The team configured the chip's RRAM cells into two circuits: one that provided a quick but rough calculation, and a second that refined and tweaked the result over subsequent iterations until it arrived at a more accurate number.
This chip configuration allowed the team to combine the speed of analog computing with the precision typically associated with digital processing. It's important to note that the chip was made using a commercial manufacturing process, meaning it could potentially be mass produced.
According to the researchers, future improvements to the chip's design could further improve its performance. Their next goal is to create larger, fully integrated chips that can solve more complex problems at higher speeds.
