When it comes to the upgrade of hardware such as computers and mobile phones, in recent years we have often used the term "squeezing out toothpaste" to describe the performance of new generation products. Indeed, with the rhythm of annual updates, the performance improvement of products usually remains between 10% and 20%, and breakthrough progress is rare. But every technological leap will mark the beginning of a new era in our world.
With the explosive growth of AI among individual users this year, not only have unprecedented AIGPU accelerator cards with super computing power appeared on the server side, but also on the terminal, computer CPUs have also achieved a comprehensive technological leap.
CPU is no longer just a CPU
When it comes to CPU, our inertial thinking will think that it is the computing core of our computer, and we will pay great attention to its performance. But in fact, starting from ten years ago, the "black bump" in our computers no longer simply has the function of the CPU.
As early as 2011, Intel's second-generation Core processor (codenamed SandyBridge) had integrated the GPU and related display output interface circuits with the CPU chip on a die.
In addition, various dedicated acceleration units such as video codecs and signal processors are also integrated into the CPU to form a "system-on-chip", which greatly improves the processing efficiency of the CPU and reduces dependence on peripheral chips.
However, as the chip manufacturing process is iteratively upgraded towards increasingly finer dimensions, the physical properties are getting closer and closer to the performance limits of existing materials, technology implementation is becoming more and more difficult, and chip manufacturing costs are increasing sharply. In order to be able to integrate more circuits while effectively controlling yield and production costs, Chiplet technology emerged.
Chiplet technology splits a large chip into multiple small chips (Chiplets) and then integrates them together through advanced packaging. In contrast,SoC technology manufactures the entire system on one wafer. Chiplets can use the manufacturing technologies of different manufacturers to allow each chip to be independently optimized and reassembled to perfectly combine performance and cost. It is regarded by the industry as a new generation of system-on-chip solution after SoC.
Although Intel's latest Core Ultra processor is not the first product to use Chiplet design, this upgrade marks the official entry of PC processors into a new era.
Core Ultra four cores in one
Intel's latest MeteorLake architecture Core Ultra processor mainly integrates four small chips with different process technologies, including ComputeTile, GraphicsTile, SoCTile and I/OTile.
Among them, the ComputeTile part is manufactured based on the Intel4 process technology of EUV technology, while the Graphics Tile, SoCTile and I/OTile are all manufactured by Semiconductor Manufacturing Co., Ltd. and produced using TSMC's 5nm and 6nm process technology. Finally, they are connected together through Intel's Foveros 3D packaging technology.
Compared with previous generations of single-process single-wafer manufacturing, Intel Core Ultra not only achieves heterogeneous integration in the core architecture, but also adopts multi-source wafer manufacturing for the first time. It can be said to be Intel's double breakthrough in process technology and chip design.
As the world's largest semiconductor company, Intel has always insisted on independent and controllable chip process technology. Early Moore's Law promoted the steady growth of its processor performance, and its vertically integrated closed ecosystem became the cornerstone of its dominance of the PC industry.
However, after chip manufacturing reached 10 nanometers, Intel's process technology began to significantly lag behind its competitors. Obstacles such as the physical limits of lithography machines and dielectric layer problems have significantly slowed down its iteration speed.
Professional foundries such as TSMC that focus on wafer manufacturing continue to make breakthroughs at the 7nm/5nm node. This forced Intel to start promoting strategic transformation and proposed a "five process nodes in four years" plan. That is, by promoting five process nodes of Intel7, Intel4, Intel3, Intel20A and Intel18A within four years, it will regain process leadership in 2025.
With the smooth promotion of this plan, the ComputeTile part of the Core Ultra processor uses the Intel 4 process technology. At the same time, the new small chip design, through the coordinated integration of multiple process technology advantages, has bought valuable time for Intel to return to the process leadership. The chiplet model also indicates that the semiconductor industry will present a highly collaborative and symbiotic pattern across manufacturers and borders.
Process processes and key technologies used by Intel’s previous generations of processors:
1971 4004 uses 10 micron PMOS process
In 1978, 8086 used 3 micron HMOS process
In 1985, 80386 used the 1.5 micron HMOS process and introduced 32-bit for the first time.
In 1989, 80486 used a 1 micron process and the number of integrated transistors exceeded 1 million.
In 1993 Pentium used 0.8 micron BiCMOS process
In 1997 PentiumII used 0.35 micron CMOS process
In 2000, Pentium III used 0.18 micron Coppermine process
In 2006 Core2 used 65nm process and high-K metal gate
2011 2nd generation Core, 32nm process, CPU integrated GPU
2012 3rd generation Core, 22nm process, tri-gate FinFET technology
2017 8th generation Core, 14nm++ process, Tick process node has experienced a large delay
2019 10th generation Core, a mix of 10nm and 14nm
2021 12th generation Core, Intel7 (10nm) process, CPU adopts performance hybrid architecture, large and small core design
2024 Core Ultra, Intel4 (7nm) process, separate modular design, using multi-source wafer manufacturing for the first time