Vega's Integrated Graphics performance
23 Sep 2023 8:27:29 UTC
Topic 230107
(moderation:
My laptops cpu is an Ryzen 5 5600H with the Integrated Graphics chip Vega 7 (7 CU, 448 shaders) with 4 GB shared RAM alongside with the dedicated GeForce GTX 1650 gpu.
I noticed that Vega's 7 performance in All-Sky Gravitational Wave search on O3 v1.04 (GW-opencl-ati) task is higher that i was expecting to see (4200 secs approximately) when i compare it to much bigger chips with ultra fast dedicated VRAMs at the same task.
Anyone else crunching systematically with AMD's integrated Vegas?
Some months ago I ran an
)
Some months ago I ran an Athlon 3000G with a Vega 3 Igpu, it ran FGRPB1G and BRP7 very smoothly, I don't remember the crunching times, but I think BRP7 was a bit over 4 hours... I actually still have this CPU laying around, maybe I put it on some used mobo to crunch again... don't want to spend money on it...
I have an elderly Ryzen 2400G
)
I have an elderly Ryzen 2400G laying around. If the MB I bought would just show up I could try it out.
I have a 5700g that I am not currently crunching on (windows daily driver). Hmmmm
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
The Vega 11 of my 3400G (@45
)
The Vega 11 of my 3400G (@45 Watt) took about 1,5 h for one BRP7-WU.
Tom M wrote:I have an
)
It looks like my 5700G is taking about 62 minutes per MeerKat task. Single task on the gpu.
And about 92 minutes on the AllSky task.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
Tom M wrote:Tom M wrote:I
)
About the same run times here. Now i switched to 2 wus simultaneously but i didnt get any real gain (almost doubled the run times).
I notice now that my GT710 (192 shaders, 2GB VRAM) needs almost 17000 sec for a Meerkat task (Binary Radio Pulsar Search (MeerKAT) v0.12 (BRP7)) and the Vega 7 of the Ryzen almost only 4000s.
I think the the horsepower difference is not that much (448 shaders vs 192) that could explain the massively different run times.
Could the veteran cpu (Intel Q6600) working with the GT710 be the reason?
HAL9000 wrote:I notice now
)
Might be the smaller video ram? Not sure what generation of video ram the gt710 runs. But all the e@h GPU tasks seem to be sensitive to the video speed and GPU memory bandwidth of the video system running the task.
https://www.techpowerup.com/gpu-specs/geforce-gt-710.c1990
But that is around twice as many shaders in conjunction with (probably) more bandwidth.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
Yes Tom, i also noticed that
)
Yes Tom, i also noticed that Vega gpu has the double frequency of the GT710 (1800MHz vs 900MHz) so the total difference starts to make sense
Without limitations my Vega
)
Without limitations my Vega 11 complets a BRP7 in slightly more then 1h.
HAL9000 wrote: [Could the
)
The MeerKat BRP7 tasks require very little cpu support. Your old Intel cpu would still be sufficient.
But cpu clock or system clock does affect the runtime since storage and memory access is slower on slow system clock hosts compared to modern high core clock hosts.
But your cpu has about the same core clock of a modern server cpu under full load so likely not any more effect that those hosts with the same gpu.
The primary reason for the long runtimes is the older gpu architecture, gpu shader count and gpu clocks.
So, if a get it right, the
)
So, if a get it right, the modern integrated GPUs are very power efficient and worth crunching with them as they do not require any extra power or peripheral circuits. If a i had to choose between a dedicated GPU and an integrated GPU (with comparable number of shaders) that would be definately the integrated.
Also i guess that the communication and data transfer between CPU and GPU that sit on the same die (like Ryzen 5 5600H) is much faster than when you have to use PCIE bus?