Transcript Wonders of GPUs

Moore’s Law
Electronics
19 April 1965
Moore’s Original Data
Gordon Moore
Electronics
19 April 1965
Graph of Moore’s Law
11/01/2008
EADS
3
Graph of Moore’s Law – with MS
11/01/2008
EADS
4
Graph of Moore’s Law – with MS
Human
Intelligence
11/01/2008
EADS
5
Memory bottleneck
• The CPU can add two numbers in less than
one nanosecond.
– If they are both in registers
• Putting a number from memory into a register
takes about 100 nanoseconds.
• “Stall” – the CPU waits on memory.
11/01/2008
EADS
6
Intel Haswell
Size/speed
1K / 1 ns
128Kb / 5 ns
1 Mb / 20 ns
4 Gb /125 ns
1 Tb / 1 ms
11/01/2008
EADS
7
GPU architecture
• GPUs have much less space devoted to cache.
• GPUs have multiple (100-1000) cores, which
are simpler, slower processing units.
• GPU cores all perform the same instructions,
but on different data.
• Not all the cores can be active at once. When
one stalls, another one starts up.
11/01/2008
EADS
8
GPU and CPU: The Differences
ALU
ALU
ALU
ALU
Control
Cache
DRAM
DRAM
CPU
GPU
GPU
More transistors devoted to computation, instead of caching
or flow control
Suitable for data-intensive computation
High arithmetic/memory operation ratio
Intel Core i5-3470 Ivy Bridge Processor:
11/01/2008
EADS
11