Read this - pretty much the best explanation I've found. From google translate: 作者：秦楚戈 链接：https://www.zhihu.com/question/47628487/answer/107866295 https://www.

Read this - pretty much the best explanation I've found. From google translate: 作者：秦楚戈 链接：https://www.zhihu.com/question/47628487/answer/107866295 https://www.zhihu.com/question/47628487/answer/107866295 来源：知乎 著作权归作者所有。商业转载请联系作者获得授权，非商业转载请注明出处。 Mentioned above by employing CPU + Accelerator way for heterogeneous computing, only use a CPU was homogenous computing. But Shen Wei 26010 seemed more special, if you define isomorphic calculated using the calculation method to calculate the same type instruction set and architecture units system, then, since the light divinity Taihu Lake using only Shen Wei 26010, and computing core and management of the core instruction set are the same, may be considered homogenous computing. But in fact, divinity Taihu double precision floating point peak light up 125PFlops, stable performance 93PFlops, is indeed high-performance accelerators was made - essentially, Shen Wei is a 26010 CPU and accelerators combined - Shin 260 Granville 26010 core is divided into two kinds, one is the management of the core, play a similar CPU function, the other is the core operations, play a similar role in the accelerator, which makes Shen Wei 26010 single chip to complete the Intel E5 + PHI or Power + Tesla two products function. And with respect to Intel E5 + PHI, or Power + Tesla, Shen Wei 26010 can be shared memory, which avoids the Intel E5 + explicit copy PHI, or Power + Tesla must face, thereby reducing the pressure on the memory, and reduces the performance loss. Presumably, too, Shen Wei 26010 cache and memory are obviously too small, because memory access model may be very simple - it is tantamount to abandonment complex memory management model existing cpu and memory scheduling the task entirely to the developers, only CPU supports a simple memory access model, the hardware does not cache the conformance requirements of hardware (Intel KNL the cache consistency referred hardware), will work to synchronize software. This unusual design makes Shen Wei 26010 at the same time have high performance and low power consumption to make up their own short board memory.