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

[DavidGraham]

• 4 triangles per clock

So same as FuryX?

 

[Clukos]

Thanks Rys, I really appreciate it. Going to put it into good use

Better leak some terrible mining results so that gamers may have a chance of getting one of these (or at least the consumer version) at launch *wink* *wink*

 

[CarstenS]

So same as FuryX?

From the pure spec-sheet numbers - yes.

 

[iMacmatician]

From Videocardz: "AMD preparing Radeon Pro WX 9100 with Vega GPU."

According to the data from CompuBench, WX 9100 would have 1792 cores more than WX 7100. The listing reveals that WX 9100 has 64 Compute Units (4096 Cores) and a maximum clock of 1200 MHz. Listed global memory size is definitely pointing towards 16GB configuration, so this would be sixth Vega with such memory size.

They also mention another AMD GPU on SiSoftware with 1536 SPs and a 1024-bit bus, and is associated with Kaby Lake.

 

[CarstenS]

This is btw the device ID in the driver:
AMD6863.1 = "Radeon Vega Frontier Edition"
and the PCI-ID:
"%AMD6863.1%" = ati2mtag_R7500DS, PCI\VEN_1002&DEV_6863&REV_00

all preceding cards in the driver's INF are filed as legacy such as"%AMD99A4.1%" = ati2mtag_Legacy, PCI\VEN_1002&DEV_99A4


New(ly added) antialiasing mode incoming as well?
PXAA (http://blog.metaclassofnil.com/?p=11, the GeDoSaTo guy) in the driver
[AMD_Service_Inst_R7500]
AddReg = ati2mtag_SoftwareDeviceSettings
AddReg = ati2mtag_R7500_SoftwareDeviceSettings
AddReg = ati2mtag_MultiUVD_SoftwareDeviceSettings
AddReg = ati2mtag_Mobile_SoftwareDeviceSettings
AddReg = ati2mtag_Mobile_PX_SoftwareDeviceSettings
AddReg = ati2mtag_Mobile_PXAA_SoftwareDeviceSettings
AddReg = ati2mtag_ProxySettings_DX10
AddReg = ati2mtag_ProxySettings_DX10_PXAA
AddReg = ati2mtag_PXAA
AddReg = ati2mtag_R7500_PX
AddReg = ati2mtag_DX11
AddReg = ati2mtag_PXService_2ID
DelReg = ati2mtag_RemoveDeviceSettings
ServiceType = 1 ; SERVICE_KERNEL_DRIVER
StartType = 3 ; SERVICE_DEMAND_START
ErrorControl = 0 ; SERVICE_ERROR_IGNORE
LoadOrderGroup = Video
ServiceBinary = %12%\atikmdag.sys

 

[DavidGraham]

From the pure spec-sheet numbers - yes.

So this lends credit to the theory that more triangle rate is dependent on the optimizations made through primitive shaders, which should be application specific, and not transparent.

RX Vega will be comparable to Radeon 9700 in terms of impact.

Gut feeling or behind the scene info?

New(ly added) antialiasing mode incoming as well?
PXAA

I think I remember PXAA being an old modified FXAA filter, it should have a temporal component as well called TPXAA.

From Videocardz: "AMD preparing Radeon Pro WX 9100 with Vega GPU."

So Vega pure Pro cards will be slower but come with certified drivers? Well there is one distinction alright.

 

[CarstenS]

Gut feeling or behind the scene info?

Wether or not I have one or the other or both would be subject to confidentiality agreements between two or more of the following parties: Me, my gut and AMD.

But just think about it, it only makes sense.

 

[Lightman]

Wether or not I have one or the other or both would be subject to confidentiality agreements between two or more of the following parties: Me, my gut and AMD.

But just think about it, it only makes sense.

Did we get so much secrecy before R300 launch? I don't remember any leaks for R9700 from that era and was totally blown away by that card at a time! Best I could afford was poor man's version, aka Radeon 9500Pro which unlocked to 9700Pro! Still have one R300 card in my collection

 

[Kaotik]

Did we get so much secrecy before R300 launch? I don't remember any leaks for R9700 from that era and was totally blown away by that card at a time! Best I could afford was poor man's version, aka Radeon 9500Pro which unlocked to 9700Pro! Still have one R300 card in my collection

9500 Pro's didn't mod to 9700 Pro, 9500's did (if they had the memories in "L" pattern around the chip instead of one row like 9500 Pro)

 

[entity279]

I think most of us here can't work out the Vega - 9700 resemblance. Especially the competitive lanscape seems vastly different (Carsten did mention the impact of Vega, hence the landscape is relevant). 9700s were gpus which needed smaller boards and they were less power hungry, and they were built on the n-1 node. Only the small boards will be possible with Vega (HBM).

@ nv camp, it seems indeed that consumer Volta parts won't be coming all that soon. But Pascal is a very efficient architecture regardless.


Which makes me believe that the one ace Vega could have (and which might have been indirectly pointed out) would be the APIs support? Implementing any and everything DX12 could throw at it? (bindless and all that stuff sebbi's been speaking about)

 

[Clukos]

Looks like some user at wccftech (lol) got his Vega FE: http://wccftech.com/amd-radeon-vega-frontier-edition-preview/

https://disqus.com/by/klaudiuszkaczmarzyk/comments/

So it's possible that we might get some early benchmarks

 

[_cat]

calculate with a big efficiency-boost in the about 160-175MHz lower clock, like the efficiency went up from FuryX 1050 MHz to the about 890 MHz of the Fiji Nano (from 250-275W down to 170-180W)
We don´t know where the sweetspot is, yet. The WX 9100 could be clocked below the sweetspot because of other circumstances.

 

[no-X]

entity279: I agree, that similar impact seems to be unlikely, but…

9700s were gpus which needed smaller boards

Compared to? GeForce 4600 Ti was longer, but the difference was insignificant. Custom models of GeForce 4800 Ti introduced shorter designs, which were comparable to Radeon 9700 PRO

and they were less power hungry

Radeon 9700 PRO had power connector, GeForce 4600 Ti didn't have.

and they were built on the n-1 node.

Radeon 9700 PRO was based on the same process as GeForce 4600 Ti. Later released FX brought newer process. I think current situations is similar, Vega is based on older process (14nm) than upcoming Volta (12nm).

 

[seahawk]

Wether or not I have one or the other or both would be subject to confidentiality agreements between two or more of the following parties: Me, my gut and AMD.

But just think about it, it only makes sense.

It makes sense, but only if Pascal has the role of R300 and Vega is the FX5800. Then we could say it is late, loud and low performing.

 

[Malo]

It makes sense, but only if Pascal has the role of R300 and Vega is the FX5800. Then we could say it is late, loud and low performing.

Yeah I can see how you're saying that from all the reviewers benchmarks and design testing we've seen.

 

[Clukos]

itshappening.gif

 

[CarstenS]

Looks like they forgot to change memory into HBC at one point.

edit: possible explanation, why 8 GiByte HBM2 stacks were not in SK Hynix product catalogue or a simple readout-bug.

https://imgur.com/id%3D1AO2C%3Btype%3Dalbum

 

[SimBy]

Looks like they forgot to change memory into HBC at one point.

edit: possible explanation, why 8 GiByte HBM2 stacks were not in SK Hynix product catalogue or a simple readout-bug.
<blockquote class="imgur-embed-pub" lang="en" data-id="a/1AO2C"><a href="//imgur.com/1AO2C">gpu-z</a></blockquote><script async src="//s.imgur.com/min/embed.js" charset="utf-8"></script>

Or neither. Raja did confirm they are using multiple HBM sources.

 

[CarstenS]

Or neither. Raja did confirm they are using multiple HBM sources.

Hence „possible“.

 

[no-X]

As I know, Micron doesn't manufacture HBM (only HMC).