Yeah wrote it above, makes sense. Except if you find the mid range cards doing 1080p at 20fps acceptable, while they would still require a CPU much more expensive than the gpu.
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
AMD 7nm GPU News and Rumours 2018/2019
- Thread starter Beren
- Start date
Yeah wrote it above, makes sense. Except if you find the mid range cards doing 1080p at 20fps acceptable, while they would still require a CPU much more expensive than the gpu.
Because Tensor cores are useless on critical accuracy, however how they could demonstrate a protein calculation or investment calculations on screen?
It is only a problem if you don't understand mixed precision (both in general and the way the Tensor cores work with input/output) and try to use them for calculations where there is a loss of floating point precision obviously if you need high floating point precision then you lose some or all of the performance of Tensor cores. It isn't that they are inaccurate as such most/all variable types for storing numbers have their limits and application and it isn't binary that you can or can't use them for a specific task as you aren't restricted to using only one variable type - in various tests people have managed to get around 60-70% of the theoretical performance of the Tensor cores while avoiding precision error.
The problem is when people work out how the performance increase is done and then you find out that your face ID secured bank account has been cleared out by holding up an egg.
That is only a problem if the programmer doesn't know what they are doing - and if someone doesn't understand that kind of maths they really shouldn't be working on those kind of applications.
Again, I posted numerous times the benchmarks here, even when Tensor cores are used on matrix calculations, yet still AMD products are competing on even ground.
A Vega FE is just 10% slower than TV100 using it's Tensor cores. And a Vega 64 is bit less than 20%. TV100 costs £8000 by the way.
I replied to you, Bru and Gregster with the numbers and the benchmarks in four different occasions. Should I do it again?
Tensor cores are not some magical entity that give perf advantage to Nvidia just because it exists.
Is not as great at you believe it is.
have you actually linked to something showing apples to apples testing? coz the last two times I've seen you link to results they weren't comparing on an even ground.
EDIT: Might want to click through to this bit where they actually show the results with Tensor cores in use
http://blog.gpueater.com/en/2018/03/20/00006_tech_flops_benchmark_2/
All times the benchmark was common eg Cifar10 (image recon), and few others. All using Ubuntu 16.10, Python 3.5, TensorFlow 1.6.
You have to consider that the libraries are closed to each architecture. You cannot use ROCm on Nvidia cards, as you cannot use CUDA9 or cuDNN7 on AMD cards.
But the benchmarks are on even ground, using Tensor cores.
I am not going down the route trying to say that ROCm 1.9.1 is even faster than ROCm 1.3 used on the above benchmarks, because I do not have a TV100 to run the tests.
Had I won some millions on the lottery, then could have done so
David Wang From AMD Confirms That There Will Eventually Be an Answer to DirectX Raytracing
Nevertheless, this executive seems to be that "AMD will definitely respond to DXR"
Nevertheless, this executive seems to be that "AMD will definitely respond to DXR"
OK all those defending AMD on this cherry picked benchmarking slide that shows them with comparable performance to NVidia' V100. have a watch of the next horizon presentation at the 59 minute mark and then the 110 minute mark.
"MI60 includes special machine learning operations to accelerate machine learning and inferencing"
Thats what David Wang states and then he shows the benchmark slide at the 110 minute mark.
next he states
"for machine learning workload we are also achieving comparable performance."
Yes it does say in the end notes that they used FP32 batch size 256.
I particularly like the line "Performance may vary based on use of latest drivers and optimisations"
Oh you mean like if you actually used the functions of the V100 ie: the tensor cores. Surely they could be special machine learning operations to accelerate machine learning and inferencing.
It is only in the WCCFtech article does it say that
So only after AMD had been caught out did they think of mentioning face ID and Banking.
So to all those who have commented and seem to be saying its ok for this sort of thing to happen in presentations, let me ask you this and try to answer honestly.
If it was NVidia pulling a stunt like this and AMD were the ones losing out, wouldn't you be a little put out by it and maybe comment on this forum that we all like to frequent.
"MI60 includes special machine learning operations to accelerate machine learning and inferencing"
Thats what David Wang states and then he shows the benchmark slide at the 110 minute mark.
next he states
"for machine learning workload we are also achieving comparable performance."
Yes it does say in the end notes that they used FP32 batch size 256.
I particularly like the line "Performance may vary based on use of latest drivers and optimisations"
Oh you mean like if you actually used the functions of the V100 ie: the tensor cores. Surely they could be special machine learning operations to accelerate machine learning and inferencing.
It is only in the WCCFtech article does it say that
So only after AMD had been caught out did they think of mentioning face ID and Banking.
So to all those who have commented and seem to be saying its ok for this sort of thing to happen in presentations, let me ask you this and try to answer honestly.
If it was NVidia pulling a stunt like this and AMD were the ones losing out, wouldn't you be a little put out by it and maybe comment on this forum that we all like to frequent.
Sorry forgot to link the AMD next Horizon info and video page.
https://www.amd.com/en/events/next-horizon
https://www.amd.com/en/events/next-horizon
Definitely - I will add the slight caveat, this is stuff technical enough that I don't know if they have definitely fiddled with something. I would love to see proper commentary from someone using this tech in the wild who has seen both new technologies.
Until it makes its way into consoles raytracing will not become mainstream, the technology is 3-5 years away from mass adoption IMHO. If you assume that currently the GTX1060 (about 15%) is the most popular card on the Steam Hardware survey, it gives you an idea of how far away we are.
https://store.steampowered.com/hwsurvey/videocard/
Basically GTX1070 and above still less than 5% of the market on DX12.
Also DXR requires DX12 and the BF5 is still showing stuttering on DX12 despite being there since beta. This is one of the titles that was “hyped” in the Turing launch.
Yeah though one of the things about Ray Tracing, even hybrid versions if done right, they don't stop you implementing another lighting engine that is used on lower end hardware and can switch over to the RT one on suitable hardware.
Despite a lot of enthusiasm for RT amongst older developers (some of the younger ones ironically I don't think have been exposed to the potential of the tech the same and too used to shadow maps, etc.) it would really take nVidia pumping out 2080tis at like £200-300 as well as putting strong support behind developers for the tech to take off any time soon
Soldato
- Joined
- 20 Apr 2004
- Posts
- 4,383
- Location
- Oxford
Years down the line after others have refined and improved there own systems
It's not like AMD haven't brought many forward looking technologies to their GPUs, only to be lambasted for "wasting transistors" on features that are not supported till a few generations down the line. The tech is barely usable as a spot effect on current top of the line Nvidia cards due to it's massive framerate hit. DXR is really only there as a marketing selling point, and for devs to start to get their eye in.
Last edited:
Soldato
- Joined
- 20 Apr 2004
- Posts
- 4,383
- Location
- Oxford
Yeah though one of the things about Ray Tracing, even hybrid versions if done right, they don't stop you implementing another lighting engine that is used on lower end hardware and can switch over to the RT one on suitable hardware.
Despite a lot of enthusiasm for RT amongst older developers (some of the younger ones ironically I don't think have been exposed to the potential of the tech the same and too used to shadow maps, etc.) it would really take nVidia pumping out 2080tis at like £200-300 as well as putting strong support behind developers for the tech to take off any time soon
Agreed, but they need a redefined design and much smaller process as its a HUGE die after all currently . ie a 8800GTX to 8800GT type product
Soldato
- Joined
- 20 Apr 2004
- Posts
- 4,383
- Location
- Oxford
Trueform comes to mind, DXR has been held back as that windows update was rolled back so developers are unable make the patches and what not.