- If you are a video game player,
you would have
shopped for graphics cards, like
Nvidia GeForce:
- A graphics card
(see picture above) is:
- Graphics card = a computer built using a GPU processor (instead of using a CPU processor)
- Traditional computer:
(what you have hearn in this course)
- The "traditional" CPU:
Explanation:
- The CPU fetches
an instruction from
its RAM memory
- CPU determines
(= decodes)
what instruction it has
fetched
- The ALU performs the
requested computation using
the values in the
registers/cache memory
- The ALU may also update a register/cache memory/RAM memory when it executes the instruction
- The CPU fetches
an instruction from
its RAM memory
-
Example execution: executing
the instruction
add R0, R1, R2
(R0 = R1 + R2)
- The "traditional" CPU:
- "Graphics card" computer:
- Architecture
(= internals) of
a graphics card "computer":
Explanation:
- The GPU fetches
an instruction from
its "device" memory
(= memory in the
graphics card)
- GPU determines
(= decodes)
what instruction it has
fetched
and
passes this
(one single) instruction
to
all the ALUs
- There are 1024 ALU units or more in one graphics card !!!
- Each ALU performs the
(same) requested computation using
the values in
its own
registers/cache memory
- The ALU may also update one of its own register when it executes the instruction
- The GPU fetches
an instruction from
its "device" memory
(= memory in the
graphics card)
-
Example execution: executing
the instruction
add R0, R1, R2
(R0 = R1 + R2)
- The GPU fetches the
instruction "add R0, R1, R2"
from the "device" memory:
- And sends the
instruction "add R0, R1, R2" to
all its ALU units for
execution:
Notice that:
- The GPU will
execute
the same instruction using
many different values
(because each ALU has its
own set of
registers !!!)
- So a GPU is a Single Instruction Multiple Data (SIMD) computer !!!
- The GPU will
execute
the same instruction using
many different values
(because each ALU has its
own set of
registers !!!)
- The GPU fetches the
instruction "add R0, R1, R2"
from the "device" memory:
- Architecture
(= internals) of
a graphics card "computer":
- In the above example,
I showed that
2 different processors in a
GPU can
perform
different computations:
- Because each processor uses its own registers/local memory values
- The number of
processors inside a
GPU (graphics card) is now
over 1000 !!!
An older model (2010) had 512 processors:
- In other words:
- A GPU processor
can execute
thousands of
instructions
simultaneously !!!!
(Each instruction will operate on different input values !!!)
- A GPU processor
can execute
thousands of
instructions
simultaneously !!!!
- I have only give you a
brief intro on
the current architecture of
the GPU
- I skipped the
evolution of
the GPU architecture
- I only presented enough information above to help you understand how to write programs for a graphics card "computer" (which is what we will discuss next)
- I skipped the
evolution of
the GPU architecture
- If you like to know more, here is a good article on the historical development of the GPU: click here