Multi-core fantasy world

On the highway, the "No Speeding" sign is everywhere.

There are a large number of patrol cars on the Maryland Highway 140. You can of course speed, but you will soon receive a ticket.
However, there are some restrictions on how difficult it is for people to break. The first thing that jumped into the head was the speed of light. However, multiprocessors should also be considered one of them. Amdahl's Law has told us the maximum acceleration formula:

Where f represents the proportion of problems that cannot be processed in parallel and n is the number of processors. In a system, only about 50% of problems can be executed in parallel. Even if you use more CPUs, the execution time you can get is still only half that of the processor you added.

The Gustafson's Law (Gustafson's Law) considers Andaral's law to be too conservative and points out that sometimes the problem is that the parallel processing of scalars is faster than the sequential processing. Google's Pagerank algorithm is an example. However, I doubt very much that in most embedded systems, Gustafson's law is not very useful.

However, I think that in many cases, Anders and Gustafson will certainly be optimistic, especially in the operation of symmetrical multi-core processors. Each of these processors has two or more identical cores, each with its own L1 cache. These cores also share L2 and common memory bus. Without L1, these cores may not even work. However, the capacity of the cache for these cores is very small, usually only 32KB. Looking at L2, the situation may be even worse, thus pushing up the demand for performance improvement of main memory. Just imagine that as many as 10 processing programs are in a waiting situation. If you add more CPUs that need memory resources at this time, bus competition will become more serious. However, since these conditions are highly uncertain and some of them are very specific, it is difficult to establish precise models for them.

To this end, researchers at the Sandia National Labs in the United States have provided some interesting data showing that under the influence of traditional parallelism, the advantages of multi-cores will rapidly decrease. In dual-core to quad-core processors, part of the execution time will be significantly slower. When it comes to the eight cores, the situation will be doubled and there will be no signs of improvement. The more cores, the slower the system. A 64-core solution shows that system performance is reduced by half the order of magnitude compared to quad-core systems.

Many major semiconductor suppliers are developing multi-core processors. In some applications, multi-core can indeed bring significant advantages in terms of speed and power consumption. But I still think that the advantages of multi-core are somewhat exaggerated. Memory bandwidth is a huge limitation. The use of asymmetric multiprocessing techniques is usually a good solution, and of course, it must also depend on the nature of the problem to be solved.

Recently, Venray Technology proposed a new processor technology that seems to solve the problem of memory bandwidth. Unlike the traditional practice of adding DRAM to the CPU, the company is doing the opposite and adding CPUs to DRAM. The company has developed an ultra-compact (only 20K transistor) processor that can be tightly integrated with memory.

Combining four micro cores with 64 MB DRAM, the total number of CPU transistors is only 0.01% of the memory. Venray's company website only provides some technical information, but this idea is quite noticeable.