Recently, Ali Damo Institute said that it has developed the world's first DRAM-based 3D bonding stacking storage and computing integrated chip.
This chip breaks through the performance bottleneck of the von Neumann architecture, and in specific AI scenarios, the performance of the chip is increased by more than 10 times, and the performance ratio is increased by up to 300 times.
As soon as this news came out, netizens were excited, and everyone said that this is the example of the domestic core's curve overtaking.
Why is this overtaking in a corner? You may have to solve with you what is the "memory and computing integrated chip".
We know that ordinary computers have two very important pieces, that is, CPU and memory, which almost determine the performance of the entire computer.
Why does the CPU and memory determine the performance of the computer? Because the data is transferred from the memory unit to the computing unit, as long as the memory and CPU performance are sufficient, then the computer will not be slow.
But what everyone doesn't know is that in the past few years, the CPU has developed rapidly, and the development speed of memory is far behind the CPU, more importantly, like the CPU and memory in the computer are separate, the transmission speed between the two also restricts the CPU performance, and at present, under all protocols, the transmission speed cannot keep up with the CPU computing speed.
Ali's integrated chip of memory and computing is to directly integrate the data storage unit and the computing unit into one, which can greatly reduce data handling and alleviate the bottleneck caused by data transportation, thereby greatly improving computing parallelism and energy efficiency.
In the specific chip, DAMO Academy will use hybrid bonding (Hybrid Bonding) 3D stacking technology to interconnect the computing chip and memory chip face-to-face with specific metal materials and processes.
Among them, the memory unit uses DRAM, and the computing unit adopts a streaming custom accelerator architecture, which enables the memory unit and the computing unit to be better combined to speed up the operation speed of the system.
Perhaps many people are not unfamiliar with this practice, because last year Apple's M1 chip also used a similar technology, putting memory and chip together to reduce the loss caused by data transmission, but the principle is slightly different from the memory and computing chip of Damo Academy.
I don't know what everyone thinks about this technology? In fact, the integrated technology of storage and computing was proposed 20 years ago, but due to the difficulty, high technical requirements, and CPU and memory technology, which were far from reaching the limit of development before, this technology has not been paid special attention to by everyone, so it has been slow progress.
Nowadays, the development of chips and memory technology is about to reach the limit of current physics, Moore's law is about to fail, unilaterally to study stronger computing chips, memory chips have become more and more difficult, so this kind of memory and computing integrated chip, to a certain extent, is really a technology of curve overtaking, what do you think?