Parts/Accelerators/Intel Xeon Phi 31S1P
Accelerators · Knights Corner · 2013
Intel Xeon Phi 31S1P
57-core coprocessor card, three per node in the original Tianhe-2. The planned upgrade to this line was blocked by the 2015 Entity List addition.
Why this part
Intel's pitch against CUDA was that Xeon Phi ran x86, so existing code would work without a rewrite. It did work, and it worked slowly: getting real performance still required restructuring for wide vectors and many weak cores, which is the same effort CUDA demanded without the tooling. Tianhe-2 used 48,000 of these until the April 2015 Entity List addition blocked the planned upgrade, at which point the machine was rebuilt on NUDT's domestic Matrix-2000. That substitution is the single clearest natural experiment in this dataset.
Systems using this part
- Systems
- 2
- Rmax underneath
- 63.51 PFlop/s
- Units recorded
- 2,688
| System | Role | Country | Units | System Rmax | Supplied by | Rank |
|---|---|---|---|---|---|---|
| Tianhe-2 LvLiang Solution 2014 | Accelerator | China | 2,688 | 2.07 PFlop/s | Intel | #22 → |
| Tianhe-2A 2013-06 | Accelerator | China | - | 61.44 PFlop/s | Intel | #191 → |
“Units recorded” sums only the edges where a public source states a quantity. It is a floor, not a total, and should never be read as installed base.