AMD Radeon VII Detailed Some More: Die-size, Secret-sauce, Ray-tracing, and More

[Jism]

Those CU's could be programmed to fill in raytracing. These vega cards are programmable till tokio.

 

[londiste]

Dedicated Hardware acceleration for RT is a smokescreen IMO, the key is if you can cut down your FP or INT instructions as small as possible and run as many as parallel as possible. AMD does have some FP division capability so its possible that some cards can be retrofitted for RT.

You mean essentially RPM or 4*INT8? Vega brought them into consumer space and got some shiny moments in game performance thanks to it. In the other camp Turing followed suit with including RPM and at least on GPU level 4*INT8 was in Pascal if not earlier.

Who said async compute was dead. It was primitive shader and dsbr that never worked on vega, not async compute. I think you are confused here. Never ever AMD said async compute was not supported or dead.

Async is alive and kicking. However its impact is fairly small. Following what is in the news post Strange Brigade actually gains a few % from Async Compute being enabled, at best. It is definitely a good thing to have but a game changer. It also works fine enough in both camps by now.

 

[INSTG8R]

You mean essentially RPM or 4*INT8? Vega brought them into consumer space and got some shiny moments in game performance thanks to it. In the other camp Turing followed suit with including RPM and at least on GPU level 4*INT8 was in Pascal if not earlier.

Good point. RPM often gets overlooked I know FC5 is using it and I’m gonna assume AC Odyssey would too.

 

[londiste]

Why is Radeon VII only 7.5% faster in hitman 2?

Isn't Hitman 2 fairly CPU-hungry? AMD's game test results are on i7-7700K.

I also have a feeling that AMD may be working secretly with Intel on RayTracing tech to sett up a unified standard against nvidias RTX.

Bullshit. DXR in DX12 and Vulkan-RT extensions is as standard as it gets. AMD will do their implementation of these if they know what is good for them (and extend on these if necessary).
Unless you are implying AMD would either try shoving RT into DX11 or into something proprietary?

 

[Midland Dog]

still 64 ROPs damn...
so just Vega but at 7nm and more hbm2 to boost up the price.
where is Navi??

apparently its 128

 

[Kaotik]

Isn't Hitman 2 fairly CPU-hungry? AMD's game test results are on i7-7700K.

Bullshit. DXR in DX12 and Vulkan-RT extensions is as standard as it gets. AMD will do their implementation of these if they know what is good for them (and extend on these if necessary).
Unless you are implying AMD would either try shoving RT into DX11 or into something proprietary?

Last time I cheked Vulkan doesn't have official RT-extensions at this time, NVIDIA was at least trying to push their solution [essentially RTX] as standard but at least so far that hasn't happened to my knowledge

apparently its 128

Unless proven otherwise it should be 64, as the Vega 20 diagrams from Instinct release clearly show 4 Pixel Engines per Shader Engine.
I think the 64/128 confusion comes from the fact that NVIDIA cards have their ROPs tied to memory controllers, so doubling memory controllers should double the ROPs and same logic is applied to Vega on some sites, even though in AMDs case the two aren't tied together

 

[Midland Dog]

Last time I cheked Vulkan doesn't have official RT-extensions at this time, NVIDIA was at least trying to push their solution [essentially RTX] as standard but at least so far that hasn't happened to my knowledge


Unless proven otherwise it should be 64, as the Vega 20 diagrams from Instinct release clearly show 4 Pixel Engines per Shader Engine.
I think the 64/128 confusion comes from the fact that NVIDIA cards have their ROPs tied to memory controllers, so doubling memory controllers should double the ROPs and same logic is applied to Vega on some sites, even though in AMDs case the two aren't tied together

pretty sure GN said 128

 

[INSTG8R]

Last time I cheked Vulkan doesn't have official RT-extensions at this time, NVIDIA was at least trying to push their solution [essentially RTX] as standard but at least so far that hasn't happened to my knowledge


Unless proven otherwise it should be 64, as the Vega 20 diagrams from Instinct release clearly show 4 Pixel Engines per Shader Engine.
I think the 64/128 confusion comes from the fact that NVIDIA cards have their ROPs tied to memory controllers, so doubling memory controllers should double the ROPs and same logic is applied to Vega on some sites, even though in AMDs case the two aren't tied together

Actually the ROPS are tied to the memory which is why it has 16GB. I can’t provide a direct source for this quote but it rings true
`Unfortunately, you can't scale down the HBM2 any further and still retain the 128 ROPs, so 16 GB is the smallest capacity AMD can offer, which is why the pricepoint on this is so close relative to the 2080.
`

 

[Kaotik]

pretty sure GN said 128

I'm aware many have said 128, but none of the press material provided by AMD suggests such and Radeon Instint block diagrams say 64, so until AMD itself says 128 or we get benchmarks showing ROP capabilities past 64 units it's more probable option than 128

Actually the ROPS are tied to the memory which is why it has 16GB. I can’t provide a direct source for this quote but it rings true
`Unfortunately, you can't scale down the HBM2 any further and still retain the 128 ROPs, so 16 GB is the smallest capacity AMD can offer, which is why the pricepoint on this is so close relative to the 2080.
`

I'm pretty sure they're not in AMDs case, they're just assuming it because they're tied on NVIDIA and most AMD chips use same ROP:Memory Controller -ratio. Fiji for example has 4096-bit HBM memory controller and 64 ROPs, while it should have 128 if the memory controllers and ROPs were tied together. Also, with Vegas HBCC they're even less connected than before, they're actually behind Infinity Fabric -bus now.

For the 16 GB, it's the smallest capacity you can have with 4096-bit HBM2 because no-one makes smaller than 4GB HBM2-stacks.

 

[londiste]

Last time I cheked Vulkan doesn't have official RT-extensions at this time, NVIDIA was at least trying to push their solution [essentially RTX] as standard but at least so far that hasn't happened to my knowledge

You are right, my bad. There are only NV_raytracing extensions for Vulkan that went out of beta. The official answer was that Vulkan allows doing RT already.
I thought Vulkan was supposed to improve on how (badly) OpenGL dealt with extensions

 

[INSTG8R]

I'm aware many have said 128, but none of the press material provided by AMD suggests such and Radeon Instint block diagrams say 64, so until AMD itself says 128 or we get benchmarks showing ROP capabilities past 64 units it's more probable option than 128


I'm pretty sure they're not in AMDs case, they're just assuming it because they're tied on NVIDIA and most AMD chips use same ROP:Memory Controller -ratio. Fiji for example has 4096-bit HBM memory controller and 64 ROPs, while it should have 128 if the memory controllers and ROPs were tied together. Also, with Vegas HBCC they're even less connected than before, they're actually behind Infinity Fabric -bus now.

For the 16 GB, it's the smallest capacity you can have with 4096-bit HBM2 because no-one makes smaller than 4GB HBM2-stacks.

We really just need to wait for proper product spec sheets at this point because right now both are being tossed around and nobody seems to have a concrete answer.

 

[Rahmat Sofyan]

hats off for Dr.Lisa Su ..

still calm and responded with great answers ..

DXR not ready yet 100%, still plenty of time for RTG and AMD to get ready.

my RX 570 and RX 480 still Okay

 

[renz496]

Good luck with RayTracing in software, if that was viable we would have had that already. If they do it it is just a desperate move not to look obsolete.

Do not expect RayTracing in hardware until end of 2020 and even then they will be years behind nVidia who will, by that time, be in the process of readying their third gen RTX cards for release.

We need Intel to enter the market with RayTracing from the get go in 2020.

I also have a feeling that AMD may be working secretly with Intel on RayTracing tech to sett up a unified standard against nvidias RTX.

RTX is just nvidia fancy name for their hardware implementation. just like when they call their tessellation engine as "polymorph engine". the unified standard for ray tracing already exist in DirectX called DXR. RTX is not some exclusive API like mantle where it can only run on certain hardware. since DXR is the standard in DirectX AMD and Intel will have to follow that standard instead of coming out with new standard.

 

[Assimilator]

Probably the most useful thing about this card is that, if it is able to perform anywhere near the RTX 2080, it might well induce NVIDIA to drop the latter's price. Considering how expensive Vega 56/64 were, and remain, I'm pretty sure NVIDIA has a lot more wiggle-room in terms of pricing - and they surely would love to shut AMD out from the high-end GPU market completely, because that woudl guarantee them an effective monopoly on that market segment going forward.

tl;dr NVIDIA might well be willing to drop the price on RTX 2080 to allow them to hike the price on RTX 3000 and all its descendants.

 

[Camm]

tl;dr NVIDIA might well be willing to drop the price on RTX 2080 to allow them to hike the price on RTX 3000 and all its descendants.

RTX die sizes are huge, I dont think there is much wiggle room at all. Conversely, Vega VII die is much smaller, but HBM is still expensive, so AMD probably doesn't have much room either.

Interesting times.

 

[Aquinus]

Probably the most useful thing about this card is that, if it is able to perform anywhere near the RTX 2080, it might well induce NVIDIA to drop the latter's price. Considering how expensive Vega 56/64 were, and remain, I'm pretty sure NVIDIA has a lot more wiggle-room in terms of pricing - and they surely would love to shut AMD out from the high-end GPU market completely, because that woudl guarantee them an effective monopoly on that market segment going forward.

tl;dr NVIDIA might well be willing to drop the price on RTX 2080 to allow them to hike the price on RTX 3000 and all its descendants.

That's predicated on the idea that nVidia's RTX offerings aren't having yield issues which I find hard to believe for the 2080 Ti. As for the 2080, I'm not sure, but it's still a pretty good size die (bigger than a Vega 64.) Honestly, I think nVidia's problem is old inventory. Between that and the less than stellar reception of the RTX chips, investors were not amused.

 

[Mysteoa]

Probably the most useful thing about this card is that, if it is able to perform anywhere near the RTX 2080, it might well induce NVIDIA to drop the latter's price.

The price for RTX 2080 is around 750€ where I live and Radeon 7 will probably cost more then 700€ initially.

 

[btarunr]

Here is a better look

I still only count 64 ROPs in that graphic, since each "RB" (render backend) crunches 4 pixels per clock.

 

[Ruru]

Weird that there's just 3840 shaders

 

[Vya Domus]

That die size is pretty small and also not all of it is enabled, for the first time in many years AMD has a card that likely has better margins that Nvidia's equivalent. That's a pretty big deal.

 

[ValenOne]

I highly doubt it have 128ROPs. If it did have 64 ROPs then old Vega56/64 are memory bandwidth starved. Then again, all AMD recent GPU are bandwidth starved for example RX470 that uses the same memory as RX480 performs very close to it. GCN is reaching its limits, its good for compute but not as a gaming card. They need to put more than 4 Shader Engines, which in return increase geometry units and number of ROPs.

From https://www.techpowerup.com/gpu-specs/radeon-rx-vega-m-gh.c3056
Recent "RX Vega M GH" has 64 ROPS

 

[Fluffmeister]

According to AMD the cost of 7nm is significant, with 16 hbm2 I can't imagine it's cheap for them, but i assume they are least making some money.

 

[Deleted member 158293]

AMD does have pretty much all of the gaming space to consider when developing AMD implementing new features like RT. Having a half-@$$ noisy hybrid ray tracing implementation with little traction like seen presently wouldn't do them, Sony, or Microsoft any favours, nor impress them IMO.

Hopefully AMD learned this from their Async push, which is still great technology, but the software ecosystem wasn't ready a few years ago.

 

[FordGT90Concept]

According to AMD the cost of 7nm is significant, with 16 hbm2 I can't imagine it's cheap for them, but i assume they are least making some money.

Which is why Vega 20 isn't bigger than Vega 10. I think Huang's explosion is because he realizes he made a "big" mistake with Turing. AMD is focusing on where the money is at, not winning performance crowns that mean little in the larger context of things. Turing is substantially larger (and more costly to produce) than even Vega 10 is.


On topic, Vega 20 doesn't really impress but it really wasn't intended to impress either. Vega 7nm w/ Fiji memory bandwidth.

 

[Aquinus]

Vega 7nm w/ Fiji memory bandwidth.

I don't think that the added bandwidth is going to make a difference on this chip, that was just a side-effect of putting 16GB on it. I've played around with my own Vega 64 a bit to realize that not only does HBM overclock really well, it also makes practically zero difference in terms of performance, even with a 20% overclock on it. Vega was never starved for memory bandwidth and to me, that says this is totally about capacity. If power consumption could be improved, that would go a long way for Vega.