In terms of architecture, the Eten 710 adopts ARM. ARM's Neoverse processor IP core, released in 2018 for data center scenarios, and the subsequent launch of ARM-based servers by Amazon and Nvidia, demonstrate the feasibility of ARM in the server field. The ARM instruction set is more efficient than x86, with faster instruction execution speed, simpler decoding circuitry, and relatively lower cost and power consumption. In fact, based on the ARM architecture, Pingtouge has also adopted multi-core interconnect technology, chip-to-chip interconnect technology, and made special optimizations to on-chip interconnects, employing a new flow control algorithm to reduce system backpressure, effectively improving system efficiency and scalability, and effectively translating single-core high performance into overall system high performance. In addition, the Eten 710 single-chip supports 128 cores, a significant advantage compared to the Kunpeng 920 and Amazon AWS Graviton's 64 cores. DDR5 memory and PCIe 5.0 are also industry-leading. It's worth noting that AMD and Intel have announced the application of PCIe 5.0 in their next-generation processors, which will greatly improve the transmission rate for the Eten 710, thereby enhancing performance.
Combined with Alibaba's expertise in system optimization and hardware/software co-design, the performance of the Eten 710 can be fully unleashed, serving the cloud computing market with lower costs and higher computing power.
Apple M1 Max -- 57 billion transistors
Almost simultaneously with the release of the Eternity 710, Apple released the M1 Max chip, which also uses a 5nm process and has 57 billion transistors. The CPU still has eight large and two small cores, totaling 10, and the Neural Engine still has 16 cores.
The number of GPU graphics cores has doubled again to 32, and performance has also doubled or nearly doubled across the board: 4096 execution units, up to 98304 concurrent threads, 10.4 TFlops of floating-point computing power, 327 billion texture fills per second, and 164 billion pixel fills per second. This makes it the most powerful chip Apple has ever designed.
NVIDIA A100 -- Tensor Core GPU
The process is TSMC's 7nm N7, and the A100 is powered by NVIDIA's GA100 GPU based on the Ampere architecture, which includes 54.2 billion transistors and has a chip size of 826 square millimeters.
AMD--EPYC
AMD's EPYC series server processors have long surpassed the 10 billion transistor mark.
Cambrian Origin 370
The Cambricon M370 boasts impressive performance data. It utilizes a 7nm process and integrates 39 billion transistors, delivering a maximum computing power of 256 TOPS (INT8), which is twice the computing power of Cambricon's second-generation product, the M370.
The Siyuan 370 also achieved "three firsts" and "one first".
1. China's first cloud-based AI chip supporting LPDDR5 memory
2. Cambricon's first AI chip using chiplet technology
3. Cambricon's first cloud chip supporting mainstream domestic and international encryption standards
4. The brand-new inference acceleration engine MagicMind is the industry's first inference engine based on MLIR graph compilation technology to achieve commercial deployment capability.
Trillion-dollar behemoth
Simply discussing the number of transistors is not very meaningful; the specific application scenario is more important. For example, Samsung once manufactured a very large flash memory chip—eUFS—with 2 trillion transistors. Cerebras also released its first-generation WSE chip in 2019, which had 400,000 cores and 1.2 trillion transistors, using TSMC's 16nm process.
This year, Cerebras announced the Wafer Scale Engine 2, boasting 2.6 trillion transistors and 85,000 AI-optimized cores. It's also the largest chip in terms of area, measuring 46,255 square millimeters—simply and directly creating a chip from an entire 300mm silicon wafer!
The second-generation Wafer Scale Engine was specifically designed for supercomputing missions. Its size, equivalent to that of a thin and light laptop, makes it unsuitable for use in portable electronic devices or laptops.
From a performance perspective, the more transistors there are, the stronger the computing power. However, at the chip level, it is still affected by other factors, such as architecture and communication latency between multi-die components.
Of course, this comparison is purely based on the number of transistors, not the size of the chip. We're using size, not area, because with the increasing application of 3D stacking technology, the number of transistors is expected to rise significantly. Once 3nm process technology is in mass production, we believe our mobile and desktop SoCs will easily surpass the trillion-unit mark.
