Breaking through the "memory wall" allows better CPU performance and lower energy consumption

Article link:

1. Importance of CPU cache Performance
2. Introduction of cache and memory Wall
3. LRU cache algorithm error and ucompower discovery
4. Two common ideas on high-speed CPU design
5. New CPU design ideas of ucompower
6. Market Opportunities for CPU that uses the new Cache Technology

Abbreviations ):

The current memory mainly uses dram. dram for increasing transmission speeds, but the problem of high access latency remains unsolved.
DRAM latency takes the CPU time to wait for memory access to complete. This is the memory wall ". The memory wall brings about a series of problems, including limiting CPU performance and increasing CPU heating.
Ucompower proposed that the CPU computing time only accounts for less than 20% of the total time consumed by the program. A large part of the remaining 80% is caused by the "memory wall. Therefore, either increasing the CPU frequency or increasing the number of cores does not greatly help improve performance.
The storage density of SRAM is too low and the cost is too high. Therefore, DRAM memory is irreplaceable for the time being. It is currently impossible to eliminate the "Memory wall" by improving the memory technology.
Cache is currently the only effective method to resist the "memory wall.
The basic theory of caching is the principle of time locality (principle of temporal locality). Generally speaking, recently accessed projects are likely to be accessed again in the near future.
LRU is a cache replacement algorithm that reflects the timing mechanism. For a long time, almost all CPU caches use LRU and its variants.
After the experiment/analysis, ucompower found that the time scene principle is not accurate: the memory access trace distribution is extremely uneven, and the memory address usage count of the program shows a strong "Long Tail" form. Most addresses are used only once, and these addresses have no value for caching.
LRU thinks that each address will be useful in the future, so it will save a lot of memory addresses that are used only once for a long time, and valuable addresses in the CPU cache will be squeezed out.
Ucompower invented an improvement of the LRU cache replacement algorithm, called wlru. Wlru can quickly exclude memory addresses that are used only once, thus protecting valuable memory addresses in the CPU cache.
Ucompower believes that this technology helps to create a revolutionary CPU, and beat traditional chip giants such as Intel and NVIDIA.

My opinions:

1. The commercial taste of the article is heavy. The LRU algorithm cannot be said to have errors. It can only be said that the solution is not good enough.
2. No strong experimental data is provided to demonstrate the effectiveness of wlru (wlru is their lifeblood and can be understood without any help ).
3. Can a Chinese CPU beat intel with only wlru? Process/other technical details how to solve? How does one rely on ARM/MIPS authorization + OEM? (I don't know much about software)