Sunway TaihuLight (Wuxi)

Sunway TaihuLight was a supercomputer deployed in Wuxi, China composed of 40,960 compute nodes,¹ each with a SW26010 processor.

It debuted at the #1 position of the Top500 in June 2016 at 90.01 PF $R_{\text{max}}$ in a run that lasted 3 hours, 47 minutes across all 40,960 processors.

System overview

The system can be summarized as follows:²

Metric	Value
Peak performance	125 PF FP64
HPL performance	93 PF FP64
CPU frequency	1.45 GHz
CPU peak performance	3.06 TF FP64
Total memory	1,310.72 GB DDR3-2133
Injection bandwidth	128 Gb/s
Bisection bandwidth	560 Tb/s
Storage capacity	20 PB
Storage bandwidth	288 GB/s
Power (HPL)	15.371 MW

TaihuLight contained 40,960 compute nodes, each with a single SW26010 processor. Two nodes are packaged into a single “node card,” and four node cards are mounted into a single “node board” (which I would call a blade).

These blades are packaged into 32-blade supernodes:

These supernodes are analogous to subracks or node chassis in Cray parlance.

These supernodes are then packaged up into four-supernode cabinets:

Thus, a single cabinet has 1,024 processors ($2\text{ processors}\times 4\text{ cards}\times 32\text{ boards}\times 4\text{ supernodes}$). The entire system has 40,960 nodes which is ten cabinets. They are laid out unusually:³

There are four rows of ten cabinets and four rounded end-caps whose function is unclear.

Here is a photo of the system:⁴

Upgrade

I think it was upgraded to New Sunway in 2021. This would’ve required increasing the node count by 2.6x and replacing all SW26010 processors with SW26010-Pro processors. This is a significant facility upgrade, so the systems may be independent.

Footnotes

1. High performance computing of DGDFT for tens of thousands of atoms using millions of cores on Sunway TaihuLight - ScienceDirect ↩

2. 072001.pdf ↩

3. sunway-taihulight-dongarra-report-v17 ↩

4. Sunway-TaihuLight outperforms Tianhe-2 as world’s fastest supercomputer ↩