+++ Description = “We describe N2, a massively SPMD, bit serial, one petaBOPS compute cluster” date = “2021-11-11” draft = false menu = “main” title = “N2, a one petaBOPS cluster” toc = true type = “page” +++
Content note: This is attempting to make a comparison between two very different things which are not suitable for comparison. It is therefore a very unfair comparison and nothing in this page should be taken as an argument that Roadrunner is in anyway inferior to a few PS4s.
Introduction
N2
N2 is a 16 node PS4 Pro cluster implementing a 4,718,592 bit wide SPMD bit serial virtual register machine.
~137,500,000 documents in The Pile.
Roadrunner
Roadrunner is a 6,480 node cluster for performing float64 computation.
Comparison of N2 with Roadrunner.
| Cluster | N2 | Roadrunner |
|---|---|---|
| Completion date | 2022 | 2008 |
| Cost | ~$2000 (2020) | $100 million (2008) |
| Speed | 1.07 petaBOPS | 1.456 petaFLOPS |
| Nodes | 16 x PlayStation 4 Pro (8 Jaguar + 36 GCN4 CU, 8GB GDDR5 shared) | 6,480 x TriBlade (4 Opteron 16GB + 4 PowerXCell 16GB) |
| CPU Cores | 16x8 = 128 AMD Jaguar | 6,480 x 2 = 12,960 AMD Opteron |
| Accelerator cores | 16x36 = 576 CUs \ 16x36x64 = 36,864 stream processors | 6,480x8 = 103,680 SPE cores PowerXCell 8i |
| RAM | 16x8GB = 1 terabits | (6,480 x 16GB) = 101TB, 12.67 teraweights + 101TB CPU RAM |
| Storage | 128 terabits | 1,000,000/8 = 125 petaweights |
| Power | 16x170W = 2.72 KW | 2.35 MW |
It is not completely fair to compare a binary operation with a 64 bit floating point operation.
N2 is capable of performing sparse operations with no overhead (as long as there is sufficient data parallelism) something which Roadrunner may have somewhat more difficulty with.
That said, to be fair to Roadrunner, N2 requires approximately three binary operations to perform the equivalent of a multiplication and addition and a threshold operation on the sum of n bits requires approximately 5*log2(n) binary operations.
Also Roadrunner has far superior internal networking.
Also sometimes we actually do need 64 bits of precision.
Theoretical performance of 16 x PS4 pro
- 16 nodes
- 36 CUs
- 911MHz
- 2048 bit lanes per CP
Theoretical 10^15.03 BOPS.
PS5
- 36 CUs
- 2.23 GHz
- 1024 bit lanes per CU