anandtech.com
>Once we enable INT8, the $2500 Titan RTX is no less than 3 times faster than a pair of $10k Xeons 8280s.
OH NO NO NO NO NO NO NO NO NO NO NO
AHAHAHAHAHAHAHAHAHAHAHAHAHAHA
THIS KILLS THE INTEL
Once we enable INT8, the $2500 Titan RTX is no less than 3 times faster than a pair of $10k Xeons 8280s
Robert Hernandez
Caleb Scott
DELID DIS
Angel Roberts
First it was single socket Epyc, now even graphics cards obliterate Xeons. Intel just can't get a break.
Elijah Cook
>int 8
>prone to massive amounts of compacted round off errors
Pretty useless magic trick
Adam Perry
YOU CAN'T USE GPU EVERYWHERE MOST OF THE TIME THE CPU NEEDS TO DO JOB
Juan Phillips
>Doing Inference on CPU
why would you do that?
I hate Intel and all, but that's just a stupid comparison. It's like testing how a game runs on a GPU vs CPU software renderer.
Adrian Cooper
>GPU is better at a GPU workload than a CPU!
All of those intel CPU +Nvidia GPU supercomputer clusters are surely going to be thrown out upon this information being brought to light.
Evan Moore
>game
>CPU software renderer
Today you don't even have the option to do that.
They just don't make games like they used to
Andrew Reed
yeah let's ignore it's almost twice as fast in fp32
Christian Parker
>comparing GPU to CPU
>giving a shit about AI
Cameron Evans
>Titan RTX
Why aren't these poorfags using the 5k Quadro RTX 8000
Christopher Gray
Graphics cards have always been faster than CPUs are certain tasks wtf are you smoking user
Logan Rogers
>with 1 billion times more round off errors
quality > quantity
Logan Smith
Boy, aren't you retarded.
Dominic Myers
brainlet here
is this the fabled avx512 intel bet the house on?
John Wright
VNNI is good for RNNs that can't be parallelized all that much. Obviously a highly parallel CNN does better on a highly parallel GPU architecture. On client platforms (icelake) it makes sense for everything but training as most people don't want to hook up an eGPU to their ultrabook to run some basic inference app.
Ian Howard
You can have latency-sensitive but not very throughput-sensitive inference workloads, which do better on CPUs. Facebook has a lot of these for example
Eli Butler
3.5 = 4
Angel Rodriguez
What happened to the Xeon phi?
Jordan Barnes
itpeernetwork.intel.com
>One step we’re taking is to replace one of the future Intel® Xeon Phi™ processors (code name Knights Hill) with a new platform and new microarchitecture specifically designed for exascale.
Christian Kelly
>now in EOL
Joshua Lopez
Kevin Evans
It is also a lot more economical. However, I'm not so sure it is actually more efficient for large models.