CUDA - GB/s for PCI-E vs Gbps for memory clock speed for GPUs -


i'm astounded @ pci-e 3.0 speed: ~16 gb/s (gigabytes per second) against top gamma gpus memory clock speed (a titan x lists ~10 gbps i.e. giga-bits per second)

so have

pci-e 3.0    16 gb/s titan x      ~1 gb/s

so question is: why there abysmal difference? in regard cuda pci-e biggest hurdle @ harnessing gpu power seems contrary: pci-e can handle way bigger loads gpu memory can (even pci-e ~8 gb/s)

is reasoning flawed somehow?

the 10 gigabit/s memory bandwidth value titan x per-pin. 384 bit wide memory interface amounts total theoretical peak memory bandwidth of 480 gb/sec. practically achievable values 80% of theoretical peak bandwidth, both memory , pcie well.


-

Comments

Post a Comment

Popular posts from this blog

c++ - CPP, 'X' button listener -

i'm writing cpp program, , want execute function when 'x'(shut down) button pressed. for example, have infinite loop prints same thing everytime iterates, until 'x' button pressed, , want execute function before process ends. is there kind of listener matter? or other solution maybe (like delaying process shut down few seconds or something)? thanks! if want execute before process ends, can register std::atexit handler. function called before process exits.
Read more

shared memory - gstreamer shmsrc and shmsink with h264 data -

i trying share h264 encoded data gstreamer 2 processes(both based on gstreamer).after research way found use shm plugin. trying do gstreamer--->h264 encoder--->shmsink shmrc--->process1 shmrc--->process2 i able raw data videotestsrc , webcam working. h264 encoded data doesn't. test pipeline gst-launch-1.0 videotestsrc ! video/x-raw,width=640,height=480,format=yuy2 ! x264enc ! shmsink socket-path=/tmp/foo sync=true wait-for- connection=false shm-size=10000000 gst-launch-1.0 shmsrc socket-path=/tmp/foo ! avdec_h264 ! video/x- raw,width=640,height=480,framerate=25/1,format=yuy2 ! autovideosink have tried shm plugins h264 encoded data, please help iam not aware of capabilities of sink used in "autovideosink", per knowledge either need use videoconvert if format supported sink (like kmssink or ximagesink) different provided source (in case yuy2) or use videoparse if camera format supported sink. may check using gst-inspect-1...
Read more

.net - Bulk insert via Dapper is slower than inserting rows one-by-one -

i'm using dapper insert data realtime feed sql server, care performance. recently, noticed strange. out of box, if give dapper collection , insert query, fires insert statements each element. tests show can insert 1800 objects 12 fields in 1 second way (counting connection.execute(...) running time. now, didn't find batch insert functionality in dapper , implemented own (constructing parameter list , sql query). after that, found out can insert 1 batch in 3 seconds (which limited 1000 rows) (again, counting connection.execute(...) calls. so, makes batching 6 times slower sending each row in separate query. can explain me? thought people use batch operations speed process. insert time 1 second @ most. use sql server 2012 standard on local network. table i'm inserting has clustered index on primary key (which bigint field), no non-clustered indexes , triggers. can post code, there's nothing special i'm not sure why using dapper execute extension ...
Read more