1/33
Looks like no tags are added yet.
Single thread
CPU design is about making a _ _ run as fast as possible
Latency
pipeline stalls and memory accesses are expensive in terms of _
Large cache memories
use of _ _ _ to avoid memory misses
Computation Throughput
GPU design is about maximizing _ _
Extracting ILP
GPUs avoid much of the complex CPU pipeline logic for _ _
Parallelism
each thread executes on a relatively simple core with performance obtained through _
Memory and Pipeline Latencies
computations hides _ _ _ _
Bandwidth optimized memory systems
wide and fast _ _ _ _
Single Instruction, Multiple Threads
meaning of SIMT
SIMT
takes advantage of data-level parallelism (1)
Multithreading
SIMT, can be considered a constrained form of _
Operating on scalar register
SIMT, many threads each with their own state (1)
With their own local memory
SIMT, many threads each with their own state (2)
Single Instruction, Multiple Data
meaning of SIMD
SIMD
takes advantage of data-level parallelism (2)
Vector Processing
SIMD, can be considered a constrained form of _ _
One thread
SIMD; _ _ operating on vector registers
One processing element
a single thread runs on a simple processing element
One processing element
short pipeline
Thread state
the execution context consists of the _ _
Fetch and Decode
but _ _ _ are costly
Multiple Processing element
SIMT execution of threads
Multiple processing element
cost of fetch and decode spread across all threads in a work group
Cheap unit
one per thread for _ _, multiple functional units for parallelism (1)
Fewer expensive
_ _ units than threads, multiple functional units for parallelism (2)
Stalls
however, _ are costly
Minimizing Stalls
include support for multiple work groups
Independent
minimizing stalls, each is _ of all others
Scheduler
a _ chooses work groups to run
Ready Work Groups
minimizing stalls, scheduler, maintains a list of _ _ _
Shader Core
minimizing stalls is sometimes called a
Scaling out
multiple instances of each shader core provided together
Scaling out
massively increases parallelism
Scaling out
memory hierarchy provided, too