## 关于超算发展与架构分类的笔记

1. SIMD (Single Instruction Multiple Data) 阵列处理机。【e.g. ILLIAC-IV阵列处理机，约1】
2. 具有流水结构的向量机。
3. MIMD (Multiple instruction multiple data) 共享存储多处理机系统。
4. MPP系统。【用小节点、非一致性存储访问结构(non-Uniform memory access, NUMA)、超节点或向量加超标量的方法】
5. 集群系统。【早期大多同构，当今逐渐转向异构方式，主要由CPU+GPU或CPU+MIC (many integrated core)组成】

## 衡量计算性能的单位——FLOPS

$$FLOPS = sockets \times \frac{cores}{socket} \times \frac{cycles}{second} \times \frac{FLOPs}{cycle}$$

FLOPS is the abbreviation of  “floating point operations per second”. Therefore, sometimes it also is written as FLOP/s.

The “sockets” is the number of network sockets, which are the endpoint to sending or receiving message from others in the network.

And the “cores / socket” is the number of cores per socket.

The “cycles / second”, as we all know, is clock rate.

The “FLOPs / cycle” is the times of floating point operations per clock rate.

$$1 MFLOPS = 1e6 FLOPS$$,

$$1 GFLOPS = 1e9 FLOPS$$,

$$1 TFLOPS = 1e12 FLOPS$$,

$$1 PFLOPS = 1e15 FLOPS$$,

$$1 EFLOPS = 1e18 FLOPS$$.

https://en.wikipedia.org/wiki/FLOPS

https://en.wikipedia.org/wiki/Clock_rate