A look into what makes supercomputers so fast and a historical comparison between different supercomputer architectures all the way to the K computer.

1. Its all about the FLOPS

2. • Computationally expensive problems
• Simulation
• Weather
• Nuclear Explosions and Decay
• Petroleum Exploration
• Chemical Reactions
• Cryptographic Analysis

3. • 15,000 operations per
second
• 9800 vacuum tubes
• It cost 2.5 Million
Dollars
• 16 decimal digits
• Clock: 1ms
• RA CRT mem: 3600
words
• Magnetic tape: 8 units,
• 4-track, 510
char/inch, 71,500
char/sec.

4. • williams tube
• A CRT tube that can
hold 512 to 1024 bits
of data
• stored 2000 words
with 4 bits per 16
decimal digits
representation

5. • 1 MegaFLOP
• 1 CPU 10 PP
(Peripheral Registers)
• $8 Million
• 60 bit words with 18
bit address
• Instructions:
• 256k total memory
Add, Shift, Multiply
• Not Pipelined
• Divide, Branch, Incre
ment

7. • 64 bit words
• 12ns clock time
• Speed: 72 MegaFLOPS • Instructions 32 bits with 128
total instruction codes
• Memory: 1,048,576 words
- 4 words read per cycle
• Power: 115 KiloWatts
• Cost: $8.8 Million
• No wire more than 4 ft long
• Freon Refrigeration

9. • Mem: 256 Million x 64 bit
words
• Instruction: 32 bit address
• 128 Instructions total
• Clock: 4.1nanosecs
• Speed: 1.2 GigaFLOPS
• CPU: 8 CPUS
• Power: 150 - 200 kW
• Cost: $17 Million
• Vector Registers

11. • 104 Cabinets, 76 computers, 8 switches, 20 disks
• PFS- Parallel File System allow for more than 1GB/S of
data on system busses
• Mem: 1.2 Terabytes
• Speed: 1.3 TeraFLOPS
• CPU: 10,000 Intel Pentium Pros at 200 mhz each
• Power: 850 kW
• Cost: $50 Million

13. • Mem: 1.4 Petabytes
• Speed: 10 PetaFLOPS
• CPU: 705,024 cores
• with SPARC64 2.0 8 cpu cores
• Power: 12,659 kW
• Cost: $1.25 Billion

14. Speed Comparison
1 MegaFLOP Playstation 2
= Or
1964 CDC 6600 3 iPhones

15. Speed Comparison (cont...)
>
2011 IPAD 2 1985 Cray II
1.5 GigaFLOPS 1.2 GigaFLOPS

16. Speed Comparison (final)
≈
2011 Intel Knights Corner 1996 Intel ASCI RED
50 core processor - x86 1.3 TeraFLOPS
Using MIC (Multi Integrated Core)
1.5 TeraFLOPS

17. • Massively parallel
• Hundreds of thousands of nodes
• K will have 800,000 nodes

19. • Multithreaded programming by hand is hard.
• Peta-scale multithreaded programming by hand is
impossible.
• OpenMP
• OpenMPI

20. • With such a large
number of
nodes, communication
between the nodes is
a major aspect of
supercomputer
design.
• Communication
Networks
• Grid Networks
• N-dimensional Torus
Networks

21. Commodity hardware
networked together
with special software.