AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

[Deleted member 13524]

Moving on, it seems we were mistaken in the interpretation of that "128 32bit ops per clock" slide.
It's "ops", not "madds", so ALU count on the "NCU" is the same. Each ALU can do two operations in a cycle (multiply + add), so that hasn't changed.

Considering the countdown actually said "Countdown to Vega Architecture Preview", I don't see what's the fuss is about - they delivered what they promised.

They promised a preview but didn't deliver a preview.

First and foremost, today’s detail release is a teaser, not a deep dive, or even a preview. AMD is only releasing a few details about Vega, and those are being kept at a high level.

 

[Silent_Buddha]

HPC is the "Total pool of memory" that was called VRAM in the Past says here:
http://techreport.com/review/31224/the-curtain-comes-up-on-amd-vega-architecture

Er, wait, what?

Though that performance might not sound so impressive, it's worth noting that all of the demo system's vents (including the graphics card's exhaust) were taped up, and it's quite likely the chip was sweating to death in its own waste heat.

That's fairly impressive if there was no place for the system to vent heat generated by the GPU.

Regards,
SB

 

[Anarchist4000]

In the end, AMD speaks 128-ops 32-bit per cycle (= 64 FMA FP32 as on all GCN GPUs), 256 ops 16-bit per cycle (= 128 FMA FP16) and 512 ops 8-bit per cycle Cycle (= 64 DP4A). No information regarding the FP64 throughput that is just announced as configurable.
Link

So far it looks like they made a nearly Fiji sized die with roughly the same CU configuration. We could be looking at a bunch of scalars and SIMDs with fused mul+mul ALUs along with other instructions. FP16 and FP32 flops may realistically be comparable for typical workloads. This might actually be a GP100 competitor requiring a good deal of compiler optimization to really see the benefit.

 

[Silent_Buddha]

HBC, but yeah. Let's rename a few more things just to confuse the hell out of people.

It AMD's defense, it appears that the role of what traditionally would be called VRAM on their cards is going to be asked to potentially operate as a cache for off card memory (L3 cache? L4 cache?), not just local video memory. In that roll, HBC is probably a more descriptive moniker for how on card memory could be used. At least that's how I understand it. Now, the question is, has the memory controller been changed to facilitate that roll? Or is it just a handy marketing name for how AMD envisages it being used in certain scenarios.

Regards,
SB

 

[MDolenc]

Can anyone explain better how the primitive shader is supposed to work (and interact with current APIs) ?

That's going to be a fun one to guess around yes. Especially as the slide has Vertex Shader => Geometry Shader replaced by Primitive Shader and ignoring that there are tessellation stages (including programmable hull and domain shaders) in between.

 

[Kaotik]

Moving on, it seems we were mistaken in the interpretation of that "128 32bit ops per clock" slide.
It's "ops", not "madds", so ALU count on the "NCU" is the same. Each ALU can do two operations in a cycle (multiply + add), so that hasn't changed.





They promised a preview but didn't deliver a preview.

I guess that comes down to what one considers "a preview"

 

[Anarchist4000]

https://www.youtube.com/user/amd/videos
Bunch of videos AMD recently uploaded.

• "Removing resource management from the developer with Vega" Link

That's going to be a fun one to guess around yes. Especially as the slide has Vertex Shader => Geometry Shader replaced by Primitive Shader and ignoring that there are tessellation stages (including programmable hull and domain shaders) in between.

Just use a shader to implement all of it. Same thing that occurred with DX9 replacing fixed function vertex/pixel functions with programmable.

 

[Deleted member 13524]

So far it looks like they made a nearly Fiji sized die with roughly the same CU configuration. We could be looking at a bunch of scalars and SIMDs with fused mul+mul ALUs along with other instructions. FP16 and FP32 flops may realistically be comparable for typical workloads. This might actually be a GP100 competitor requiring a good deal of compiler optimization to really see the benefit.

Regarding size:

Has anyone ever made measurements on Raja Koduri's fingers?
Unless the perspective is tricking me, the chip looks enormous from the picture, definitely larger than Hawaii and maybe larger than even Fiji.

Two HBM2 stack double-confirmed, by the way.

 

[seahawk]

It looks more like a competitor to GP100 than anything else.

 

[Deleted member 13524]

More pics here:

http://videocardz.com/65477/amd-vega-gpu-pictured-features-two-hbm2-stacks

Someone has already done the comparison using the HBM2 chips:

Looks like a tad bigger than GP102 but definitely smaller than Fiji or GP100.

 

[Anarchist4000]

maybe larger than even Fiji.

Fiji was right around 600mm2. One of the articles ballparked it ~520mm2. So it's definitely a big one.

 

[Rootax]

So, they talk about memory a lot, but by the size of it, can they put 4 hbm2 stack ? Seems limited to 2. Does that mean 8gb max ? Or HBM2 stack can have bigger capacity per stack ?

EDIT : sorry, the answer is bigger stacks are coming. I just wonder when.

"Sur base des premiers modules disponibles qui sont de type 4 Go (4-Hi), Vega 10 sera ainsi associé à 8 Go de HBM2, mais pourra passer à 16 Go quand les modules 8-Hi seront disponibles."

 

[DavidGraham]

It looks more like a competitor to GP100 than anything else.

Performance wise, it doesn't look so. So far at least. TechReport had this to say about it's Doom 4K performance:

AMD did have an early piece of Vega silicon running in its demo room for the press to play with. In the Argent D'Nur level of Doom's arcade mode, that chip was producing anywhere between 60 and 70 FPS at 4K on Ultra settings. Surprisingly, the demo attendant let me turn on Doom's revealing "nightmare" performance metrics, and I saw a maximum frame time of about 24.8 ms after large explosions. I also noted that the chip had 8GB of memory on board, though of course we couldn't say what type of memory was being used.

The other Battlefront demonstration wasn't revealing also, considering even a GTX 1080 can do a 4K60 in several levels.

http://www.legitreviews.com/nvidia-geforce-gtx-1080-founders-edition-video-card-review_181298/8

http://www.hardwarecanucks.com/foru.../72619-nvidia-geforce-gtx-1080-review-17.html

Though this could potentially be a cut down Vega. Or a full one with not so tuned drivers. Or a full chip throttling hard.

 

[Infinisearch]

And we still don't know if it supports ROV's and conservative rasterization.

 

[revan]

Performance wise, it doesn't look so. So far at least. TechReport had this to say about it's Doom 4K performance:



The other Battlefront demonstration wasn't revealing also, considering even a GTX 1080 can do a 4K60 in several levels.

http://www.legitreviews.com/nvidia-geforce-gtx-1080-founders-edition-video-card-review_181298/8

http://www.hardwarecanucks.com/foru.../72619-nvidia-geforce-gtx-1080-review-17.html

Though this could potentially be a cut down Vega. Or a full one with not so tuned drivers. Or a full chip throttling hard.

Those are not directly comparable. The VEGA was demoed on Endor Map. Techspot tested the Titan X Pascal using the "Battle for ENDOR" Scenario:
http://www.techspot.com/review/1246-ludicrous-graphics-test/
So, 51fps for GTX 1080 and 67fps for Titan X Pascal...

Lat year I also played SW "Battle for Endor" with my overclocked (@1500/8000MHz ...=25% core ovck+12% memory ovck ) 980Ti @Ultra settings 4K last year ~ 50fps average ; Without overclocking , an 980Ti will be ~40fps on this particlar map I guess; Right now Vega seems to be in-between 1080 and Titan XP...

 

[seahawk]

In SW:BF though which suits the Fury also quite well.

 

[DavidGraham]

Those are not directly comparable.

Hmm, we have:
LegitReviews testing of Battle on Tatooine: GTX 1080 doing an avg of 72fps @4K Ultra
http://www.legitreviews.com/nvidia-geforce-gtx-1080-founders-edition-video-card-review_181298/8

HardwareCanuks testing of AT-ST level: GTX 1080 doing 60 fps @4K Ultra
http://www.hardwarecanucks.com/foru.../72619-nvidia-geforce-gtx-1080-review-17.html

BabelTech testing of an unknown level, GTX 1080 doing 60fps @4K Ultra, Titan X doing 79fps @4K Ultra
http://www.babeltechreviews.com/titan-x-vs-gtx-1080-25-games-tested-4k-2k-1440p/3/

TechSpot testing an unknown level: GTX 1080 doing 51 fps @4K Ultra, Titan X doing 67fps @4K Ultra, 56 fps @6K Ultra.
http://www.techspot.com/review/1246-ludicrous-graphics-test/

And DigitalFoundry testing Endor Chase, Titan X doing a comfortably locked 60fps @4K Ultra + TAA
An OC'ed 1080 can also achieve 60fps in this level.

We have 3 levels where a GTX 1080 can achieve 4K60, Titan X can achieve more in them obviously, and we have at least one level where a 1080 can't do 4K60, more data is needed here, someone needs to test at least a 1080 in this Endor map. Though the safe assumption here indeed is that Vega is in between 1080 and Titan X in this title.

 

[Malo]

So you both came to the same conclusion, glad we got that sorted out ;-)

 

[DavidGraham]

Found a user who did exactly that, tested the Endor level with a GTX 1080 and a Core i5, he managed to do a perfect 60fps 4K @Ultra with a lot more action on screen.

EDIT: Apparently youtube has many videos where people with GTX 1080s can do 4K60 Ultra in this title.

 

[xEx]

Hmm, we have:
LegitReviews testing of Battle on Tatooine: GTX 1080 doing an avg of 72fps @4K Ultra
http://www.legitreviews.com/nvidia-geforce-gtx-1080-founders-edition-video-card-review_181298/8

HardwareCanuks testing of AT-ST level: GTX 1080 doing 60 fps @4K Ultra
http://www.hardwarecanucks.com/foru.../72619-nvidia-geforce-gtx-1080-review-17.html

BabelTech testing of an unknown level, GTX 1080 doing 60fps @4K Ultra, Titan X doing 79fps @4K Ultra
http://www.babeltechreviews.com/titan-x-vs-gtx-1080-25-games-tested-4k-2k-1440p/3/

TechSpot testing an unknown level: GTX 1080 doing 51 fps @4K Ultra, Titan X doing 67fps @4K Ultra, 56 fps @6K Ultra.
http://www.techspot.com/review/1246-ludicrous-graphics-test/
And DigitalFoundry testing Endor Chase, Titan X doing a comfortably locked 60fps @4K Ultra + TAA
An OC'ed 1080 can also achieve 60fps in this level.

We have 3 levels where a GTX 1080 can achieve 4K60, Titan X can achieve more in them obviously, and we have at least one level where a 1080 can't do 4K60, more data is needed here, someone needs to test at least a 1080 in this Endor map. Though the safe assumption here indeed is that Vega is in between 1080 and Titan X in this title.

Ur not taking into account drivers. As far as we know the drivers for vega were in alpha. And due to how different vega is from the past GCN we can be sure AMD have a lot of work to do with the drivers.