Vega Rumors

razor1 said:
12nm is already ready for production, no one but nV can use it cause its a specific node for them, nV probably helped make it, not sure though or it was tailored to their needs.
Click to expand...
Sauce?

https://techreport.com/news/31582/report-tsmc-set-to-fabricate-volta-and-centriq-on-12-nm-process
TSMC was vague AF.
Anand has 2018 1H HVM start.
Nvidia has said 2018 vague. Which means pretty obviously 2H in 'tech company book of announcement-speak excuses'. They'd say Q1-1H otherwise. 16nm was a rumour so it's definitely not start 2018... 16nm was the only other way.


Osjur said:
So its around 30-35% faster than Fury X according to 3dmark... good job AMD :facepalm:
Click to expand...
TBH if it's pulling 1.3GHz it's about where we expected. Throw in another 10-15% for shitty drivers + optimisation (this is par for the course AMD every damn time they have something remotely new) and you're probably closer to the real figure at about 1.40-1.5x when the dust settles. Water cooled might go 1.6x and a bit perhaps...
 
N4CR said:
Sauce?

https://techreport.com/news/31582/report-tsmc-set-to-fabricate-volta-and-centriq-on-12-nm-process
TSMC was vague AF.
Anand has 2018 1H HVM start.
Nvidia has said 2018 vague. Which means pretty obviously 2H in 'tech company book of announcement-speak excuses'. They'd say Q1-1H otherwise. 16nm was a rumour so it's definitely not start 2018... 16nm was the only other way.



TBH if it's pulling 1.3GHz it's about where we expected. Throw in another 10-15% for shitty drivers + optimisation (this is par for the course AMD every damn time they have something remotely new) and you're probably closer to the real figure at about 1.40-1.5x when the dust settles. Water cooled might go 1.6x and a bit perhaps...
Click to expand...


v100 is on the 12nm process ;) And that is a mother of a chip, its already in production. There is no myth or guess about it :), that article is a bit old. It was talked about when nV did their Volta V100 presentations.

It just depends on when nV wants to release its performance bracket cards. I think they will stick to their 1.5 years per gen, and that lands around early 2018 late 2017.

Comparing the die size and transistor amounts of current 16nm chips and V100 with 12nm, the transistor destiny didn't change so pretty clear its a modified 16nm node which was also hinted at by nV and TSMC.

The process name is 12nmFF NV or something like that, it has a NV at the end.
 
Last edited:
Chris_B said:
Its more possible the drivers will be better tuned for gaming, its obvious this card seems to be trying to be a jack of all trades. Not looking to upgrade anytime soon anyway, would just like to see some competition.
Click to expand...

Vega looks to be a Fury card built on 14nm, AMD has been tuning their platform for years. They won't be pulling better tuned drivers out of a hat, though they will likely claim too.
 
Well the main thing that will give performance for the gaming versions over the FE will be the activation or deactivation of ECC, which may give up to 10% increase in performance. Not much......
 
Ocellaris said:
Vega looks to be a Fury card built on 14nm, AMD has been tuning their platform for years. They won't be pulling better tuned drivers out of a hat, though they will likely claim too.
Click to expand...

I guess we'll have to sit on our thumbs and wait for the news to trickle in.

I do remember people saying there was no possible way for them to fix the PCI-E power draw issue for the reference RX480 with drivers so.... maybe?
 
Ieldra said:
Anarchist4000 I struggle to see how primitive shaders will effectively improve its performance. If you offload rasterization to the shader array as a compute shader you are simply taking away resources from the pixel and compute shader portions of the render. So this will only make sense insofar as the load is balance to provide the optimum framerate which will necessarily be lower than that of the competition in this case (no geometry bottlenecks)
Click to expand...
Rasterization is already in the shader array. Normally it's limited to one wave on the first CU of each shader engine. Probably running continuously without the cadence or designed around it. All the interpolation leaning on LDS hardware. Unless you only render triangles, even the Fiji rate was adequate with async. Primitive shaders will likely kick in with tessellation and running on more than 4 of 256 SIMDs likely increasing throughput. Everything on Vega looks programmable with all the Tier3 features, so devs technically are free to do whatever they want without bothering with the fixed geometry. That's what most of the SIGGRAPH papers were proposing as a faster, more flexible, option.

Just consider Sebbbi's 100% compute game that is now out we discussed a while back. Not quite years off as some suggested. It doesn't even use triangles as I understand. Just Ray marching and voxels. As they gave him a Vega I'm expecting an article soon enough. Or they're hiring him and it was an interview.

N4CR said:
They have already announced navi will use a new memory type. That said I'm wondering if instead they mean SSG and will keep HBM especially if they do MCM - they might go GDDR for the mid-low end cards though like Polaris is.
Click to expand...
Probably HBM3, which is supposedly in a low cost variant. Best guess is stacked memory like HBM2, but without the interposer. Integrated into their MCMs just like a Ryzen die. Vega, and even Fiji, already do the SSG thing. NVDIMMs, like SSG ideally uses, work well for reading and density, but would likely require a huge cache on the chip.

Ocellaris said:
Alternatively, AMD could have sent around review cards to avoid the whole mess. And AMD further screwed themselves by including a Gaming Mode, yet people will still claim RX Vega will have better performance when it launches.
Click to expand...
A card for game devs sort of needs a gaming mode. Even if performance is limited. That way devs can at least start experimenting with all the Tier3 features and patching current games for the RX release. Or devs could just wait until after the release to start patching their games with hardware they haven't seen.

Armenius said:
As far as I know all NVIDIA cards have supported feature level 12_1 since Maxwell.
Click to expand...
Features have different tiers. 12_1 was crafted by limiting 12_0 to what Nvidia supported and making 12_1 features AMD lacked. So AMD has better support for 12_0, which is what most DX12 uses.

https://en.m.wikipedia.org/wiki/Feature_levels_in_Direct3D

Nice chart down a bit there. Vega, according to a senior AMD guy, now matches WARP12. That being the "perfect" software reference. So even on 12_0 Nvidia lacks a lot of the abilities AMD possesses. Then async behavior can be problematic even though they technically meet the requirements. Just take Sebbbi as an example there. Ported to DX12 PC and AMD worked, Titan went backwards. Until he disabled async and it gained under DX12.
 
  • Like
Reactions: N4CR
like this
Don't know who/what Sebbbi is, but probably has async shaders confused with async compute.
 
Anarchist4000 said:
Rasterization is already in the shader array. Normally it's limited to one wave on the first CU of each shader engine. Probably running continuously without the cadence or designed around it. All the interpolation leaning on LDS hardware. Unless you only render triangles, even the Fiji rate was adequate with async. Primitive shaders will likely kick in with tessellation and running on more than 4 of 256 SIMDs likely increasing throughput. Everything on Vega looks programmable with all the Tier3 features, so devs technically are free to do whatever they want without bothering with the fixed geometry. That's what most of the SIGGRAPH papers were proposing as a faster, more flexible, option.
Click to expand...

Doesn't work that way, learn some graphics programming before drawing conclusions like that. You still have to factor in cache retention and that is where the problems will occur.

Tier levels have nothing to do with programmablity, the programmibility is done by the DX version not the tiers!

Now Tier levels stipulate the amount of certain features that can be done that is it. As I stated before we aren't even using 50% of tiled resource limits of Tier II right now. And CRV's aren't really being used much at all, so Tier I isn't even being stress.

Just consider Sebbbi's 100% compute game that is now out we discussed a while back. Not quite years off as some suggested. It doesn't even use triangles as I understand. Just Ray marching and voxels. As they gave him a Vega I'm expecting an article soon enough. Or they're hiring him and it was an interview.
Click to expand...

And no they aren't out yet, Sebbbi was working on a 100% compute application that was not even in beta or alpha at the time 6 months ago. I pmed him 6 months ago when we had this discussion and he stated the same. Stop with the BS. At that time he stated he was no where near ready for release let for testing purposes let a lone a full game on something like that.

A card for game devs sort of needs a gaming mode. Even if performance is limited. That way devs can at least start experimenting with all the Tier3 features and patching current games for the RX release. Or devs could just wait until after the release to start patching their games with hardware they haven't seen.
Click to expand...

Dude there is no difference between features sets from dx12 to 12.1 at least not when teirs are concerned. DX12.1 is the addition of CRV's and ROV's that is it, The tiers are just more of what is already there in DX12. Tires and features are two different things. Tiers =! features sets period.

https://msdn.microsoft.com/en-us/library/windows/desktop/ff476876(v=vs.85).aspx?f=255&MSPPError=-2147217396

Tier III features are the same as Tier II no differences just more.

Actually quote the stuff from the source and it won't be as confusing.

First you have to support a dx version that version stipulates what tiers it needs and the tiers have their own requirements.
 
Last edited:
bildad said:
So they made a 1080 over a year after the fact that costs more than 2x as much? Should we just start calling it the Vega Failure Edition?
Click to expand...

1080 has ECC memory?

Or did you miss the part of this is a workstation card

the p4000 is wayyy below the 1080 in timespy benchmarks
 
shad0w4life said:
1080 has ECC memory?

Or did you miss the part of this is a workstation card

the p4000 is wayyy below the 1080 in timespy benchmarks
Click to expand...

Does the Vega FE even have ECC memory? Does HBC even need ECC to be added on or is it just baked in?
 
Armenius said:
Don't know who/what Sebbbi is, but probably has async shaders confused with async compute.
Click to expand...
Former rendering lead at Ubisoft and past member of DirectX advisory board. Wrote at least one SIGGRAPH paper and knows what he's doing as a AAA dev. Also a guy AMD hand delivered a Vega too.
 
razor1 said:
Doesn't work that way, learn some graphics programming before drawing conclusions like that. You still have to factor in cache retention and that is where the problems will occur.
Click to expand...
So you're suggesting the drivers and way it's been so for years is wrong? I'd suggest you learn how GPUs actually work. Even an entry level college course should suffice. The actual rasterization step isn't that involved. Just get enough pixels to fill a wave and continue. There will be far more pixels than triangles in almost all cases.

razor1 said:
And no they aren't out yet, Sebbbi was working on a 100% compute application that was not even in beta or alpha at the time 6 months ago. I pmed him 6 months ago when we had this discussion and he stated the same. Stop with the BS. At that time he stated he was no where near ready for release let for testing purposes let a lone a full game on something like that.
Click to expand...
Claybook, which renders over DirectCompute? Guess it's only announced and not quite released yet. Gameplay video looked cool though. Regardless he should be working on a blog about it with Vega. Guess we wait and see how far off "coming soon" means. At least it looks like he moved on from consoles to PC. So what BS? Just because your facts have a habit of turning upside down in short order doesn't mean I'm full of BS.
 
Anarchist4000 said:
So you're suggesting the drivers and way it's been so for years is wrong? I'd suggest you learn how GPUs actually work. Even an entry level college course should suffice. The actual rasterization step isn't that involved. Just get enough pixels to fill a wave and continue. There will be far more pixels than triangles in almost all cases.
Click to expand...

No I'm not, but if you want to discuss about this, please read up, cause I have talked to you about this before, and you don't seem to have an inclination of understanding it.


Claybook, which renders over DirectCompute? Guess it's only announced and not quite released yet. Gameplay video looked cool though. Regardless he should be working on a blog about it with Vega. Guess we wait and see how far off "coming soon" means. At least it looks like he moved on from consoles to PC. So what BS? Just because your facts have a habit of turning upside down in short order doesn't mean I'm full of BS.
Click to expand...

IT IS NOT RELEASED, they are still working on it.

Yes its BS when you say its ready for to be used, it is no where near prime time use for AAA games yet.

i will quote you on what you stated here in red

Just consider Sebbbi's 100% compute game that is now out we discussed a while back. Not quite years off as some suggested. It doesn't even use triangles as I understand. Just Ray marching and voxels. As they gave him a Vega I'm expecting an article soon enough. Or they're hiring him and it was an interview.
Click to expand...

Even Sebbbi stated the same thing. Its an experiment that is going well. So either you are BSing what he stated, or you know more then what the programmer of the damn thing is doing. Don't make other's, Sebbbi a lair by making things up about his own product. Its a pretty cool product, but its not ready for prime time use yet, nor will it be anytime in the near future. So yeah its year or two off.

You don't know how long it takes for things like this to get adopted, there is a long phase of adoption based on tools, bug testing, stability etc. They aren't ready for that yet. Then they will need the hardware for it, starting from low end all the way up to high end. Not just high end. If I remember correctly Sebbbi stated it was getting a decent frame rate on top end cards @ the time, So yeah for next generation cards it would be something to look at. You minimalize the work being done and say "its just Ray marching and voxels", that is part of it, but there is a ton more things going on that has to be considered for a full engine integration.

He also mentioned he wasn't the only team working on such a renderer, so it will happen, but nothing imminent.
 
Last edited:
define is doing the tests on disqus: https://disqus.com/by/klaudiuszkaczmarzyk/

At least it seems to be pretty good card in SpecView. Performance is comparable to P5000.

ZdkaxGl.png


Seems that AMD ran their tests with native resolution (4k?) when they were comparing it against Titan XP.
 
razor1 said:
snip....

Comparing the die size and transistor amounts of current 16nm chips and V100 with 12nm, the transistor destiny didn't change so pretty clear its a modified 16nm node which was also hinted at by nV and TSMC.

The process name is 12nmFF NV or something like that, it has a NV at the end.
Click to expand...

hqdefault.jpg

My density has popped me to you...
 
N4CR said:
Sauce?

https://techreport.com/news/31582/report-tsmc-set-to-fabricate-volta-and-centriq-on-12-nm-process
TSMC was vague AF.
Anand has 2018 1H HVM start.
Nvidia has said 2018 vague. Which means pretty obviously 2H in 'tech company book of announcement-speak excuses'. They'd say Q1-1H otherwise. 16nm was a rumour so it's definitely not start 2018... 16nm was the only other way.



TBH if it's pulling 1.3GHz it's about where we expected. Throw in another 10-15% for shitty drivers + optimisation (this is par for the course AMD every damn time they have something remotely new) and you're probably closer to the real figure at about 1.40-1.5x when the dust settles. Water cooled might go 1.6x and a bit perhaps...
Click to expand...

Even with an optimistic 1.6x... I'll wait on gaming card price before I speak! But Volta is coming...
 
N4CR said:
TBH if it's pulling 1.3GHz it's about where we expected. Throw in another 10-15% for shitty drivers + optimisation (this is par for the course AMD every damn time they have something remotely new) and you're probably closer to the real figure at about 1.40-1.5x when the dust settles. Water cooled might go 1.6x and a bit perhaps...
Click to expand...

More like it's where we expected it to be from TF number and usual lower AMD efficiency in dx11 per TF.

I can't wait for meltdowns when actual gaming card scores barely better
 
I don't check Vega news for a day and suddenly everything goes to shit.
Now AMD employees are getting involved?
Hey guys – AMD employee speaking here! I want you guys to know that we appreciate the community speaking to each other and discussing our upcoming products. We even come by here ourselves to chat with you every now and then, but we have no control over this subreddit – we do not condone censorship and we’re 100% open to hearing what you all have to say.
Click to expand...

qWVU8Jn.gif
 
Last edited:
N4CR said:
You're getting confused with prosumer type gear and workstations.

Prosumer is like e.g. my handycam. It's got enough decent glass and addon parts that it was widely used as a B or small/hidden camera for filming programmes used in national broadcasts. I have a few videos out online used throughout the national education curriculum shot entirely on that camera.
Yet I use it also for holidays and videoing family stuff.
That is prosumer.
For a card that'd be e.g. if I was CAD or video/photo editing part time or semi professionally and wanted to play a few games on the side after a long week of BS...


So recommending a quadro 4000 to someone like me, is an absolute joke. When I want to fire up Doom, or anything demanding, 30fps/1060-1070 performance at best is okay for a 1k+ card eh?
This is why Pro Duo and Vega FE exist with dual driver versions. They are targeted at the original Titan market, which Nvidia quickly stopped catering to and pushed the 'buy two cards, goyim' route.

IF FP16 applications become more popular, this becomes an even more enticing offer for those who can use both capabilities of the card. It's not a purely pro card, neither is it a purely gaming card. It's prosumer. Quadro in comparison is balls deep in workstation..
Click to expand...

No I am just pointing out the illogical narrative from AMD :)
THEY are the ones pushing this product with CAD benchmarks and DeepBench, others mention it as 'Prosumer' like anarchist does.
So again in their narrative they are comparing to wrong product, if you want to compare to gaming then you use a 1080ti not Titan xp...
FP16 can only become popular on PCs for consumers once there is a reasonable supported client base....
This is a product that is so niche market it makes Nvidia's 1080 look mass market mainstream, and even then its use in games will be limited; remember how great async compute was going to be for AMD, even on Doom it only provides max 10% while rest of gains come from their Vulkan extensions and for Gears of War 4 async compute averages around 5%.
So I would not hold ones breath expecting radical gains in games that eventually support a mix of FP16 and FP32 on PCs; on B3D actual developers say it will be nice but not a game changer on PC, and they are thinking primarily for console where such benefits are more easily translated into gains just like async compute (provides more benefits on those fixed more limited and very low level platforms than it does on PCs).

The Pascal Titan model is more launched to support the DL/academic/scientific/etc community and a step beyond the previous Titan models in some ways in terms of its focus but still not a professional visualisation GPU and that is why Quadro has always existed.
But then not sure how large the market is for professional visualisation users wanting also great gaming experience; most will want the complete support/SDK/performance for the professional workflow.
Cheers
 
Last edited:
stream goes down just before he does the power measurements.....

back again.

its using ~ 300 watts in Hitman.

Damn 1080 performance at 300 watts. perf/watt is great! ;)

rx Vega needs to be around 400 bucks to make this thing fly, damn, a gtx 1080 can use 500 watt power supply......

41 frames max settings Hitman 4k.

WOW that is gtx 1070 levels shit.
 
razor1 said:
stream goes down just before he does the power measurements.....

back again.

its using ~ 300 watts in Hitman.

Damn 1080 performance at 300 watts. perf/watt is great! ;)

rx Vega needs to be around 400 bucks to make this thing fly, damn, a gtx 1080 can use 500 watt power supply......

41 frames max settings Hitman 4k.

WOW that is gtx 1070 levels shit.
Click to expand...

Damn if this is all cases, this is DOA will maybe except for mining.
 
Chimpee said:
Damn if this is all cases, this is DOA will maybe except for mining.
Click to expand...


Even for mining its DOA, can't overclock the memory. he was only able to go 20mhz more or so on the mem :/

Well the youtuber says NOT for gaming or mining.
 
razor1 said:
stream goes down just before he does the power measurements.....

back again.

its using ~ 300 watts in Hitman.

Damn 1080 performance at 300 watts. perf/watt is great! ;)

rx Vega needs to be around 400 bucks to make this thing fly, damn, a gtx 1080 can use 500 watt power supply......

41 frames max settings Hitman 4k.

WOW that is gtx 1070 levels shit.
Click to expand...

This is the FE?
 
yep!

metro 20 fps 4k max settings, yeah that is 1070 levels too if I remember correctly.

Its a bust.
 
I think we need to wait several more days to get a clearer picture of the performance/behaviour IMO with regards to Vega FE.
Still too many unknowns and uncontrolled variables for now.
Cheers
 
Oh boy, I'm totally not buying RX Vega if this is the performance. Hopefully the gaming version is cheap, I kind of doubt it though. I had a feeling this was going to happen, if I was Raja I'd be worried about my future.
 
You must log in or register to reply here.
Back
Top