### In 2022, the first battle (strong attack):

1. Belligerents: domestic high-performance computing chip enterprises, (indirect losses of the international three giants)

2. Location of War: local governments and key industries (information innovation, Internet, smart car, industry, security and other fields)

3. Domestic high-performance CPU, accelerator card and complete machine products will be released together. Focusing on the market with the strongest demand for domestic substitution, domestic CPU / GPU / NPU companies will launch the first large-scale confrontation (the volume is healthier).

4. Even at the technical and ecological levels, domestic computing hardware still lags behind AMD, Intel and NV. However, the "strong attack" is ready to go, which is a back water battle for innovative enterprises.

5. Results: after this battle, domestic computing hardware companies will complete the first urban layout and industry layout. It will also become an important basis for the capital market to bet again.

### From 2023 to 2025, the second campaign (enclosure):

1. Belligerents: International three giants, domestic high-performance computing chip enterprises, basic software, complete machine and system integrators, etc.

2. Battle location: local governments, vertical fields (internal volume of stock market, development opportunities of revolutionary change technology in the industry)

3. According to the historical development law of computing technology, after the outbreak of computing hardware, it is followed by the outbreak of software industry. It should be emphasized that the so-called ecology of the computing technology industry takes basic software technology as the core.

4. Who will take the lead in completing the ecological layout and industrial chain coordination, then who will take the lead in entering the iterative closed loop. Hardware landing - > application implementation - > Hardware iteration - > big data + big computing drives revolutionary changes in the industry - > the outbreak of high-performance computing in the industry will be the significance of the decisive battle in the next five years.

5. For the industry, this is a backwater battle. Will the technology blockade become more violent?

6. Results: after this battle, the head enterprises appeared and successfully listed. The remnant forces carried out guerrilla warfare to find a place to live.

### From 2025, the great Showdown:

1. Belligerents: domestic high-performance computing chip enterprises challenge the international big three

2. Location of battle: meta universe

3. Talk another day ~~~

### Neglected basic software technology and talent problems of computing power

1. The basic software stack of computing power is the middle layer between computing hardware and algorithm. If you take building construction as an example, the basic software stack of force calculation is mortar and cement, and its importance is self-evident.
2. In the past, we happily purchased Intel, AMD and NV products to make products. We almost ignored the value and significance of math library (such as Intel MKL, AMD acml, Huawei KML...), and the optimization of GPU math library is more complex (such as NV cublas, cufft, etc.) , China's technical reserves and talent reserves in this direction are also very rare. Different computing architectures need to build and customize optimized mathematics libraries, which is impossible to be compatible. This operation can not be copied, and the efficiency of using open source libraries is low. There seems to be no other open source projects in the field of computing libraries in China except openblas.
3. Heterogeneous computing hardware is an indisputable trend, and a unified and open source heterogeneous computing engine is still in the exploratory stage. The long-term risk of CUDA API compatibility is very high, and it is always slow by 1 or 2 beats. Even if CUDA API is compatible, code implementation still needs great efforts! Is oneapi, ROCM and openclsycl reliable? Will a unified heterogeneous computing engine appear in the future A vision of code being implemented everywhere?
4. The iterative speed of the algorithm is too fast. Behind the "Ai + physical model" are more complex and differentiated computing requirements, which pose major challenges to the computing hardware and software stack!

### Are you ready for the arms race?

What team will risc-v send to Huawei, Feiteng, haiguang, Longxin, pingtouge, etc?

Intel, AMD and NV

Cambrian, Suiyuan, Kunlun, Muxi, Bi Ren, Xindong technology, Moore thread, Jing Jiawei, Tiantian Zhixin, Denglin technology, etc

GraphCore

---

![](././../../images_dir/1632828056/1.png)