AMD Radeon Vega GPU Architecture

[RejZoR]

Anyone else getting that feeling Vega will be the next R9 290X ? No one really expected that card to be as savage as it ended up to be. It was so powerful they could afford rebranding it to R9 390X and it was still competitive against brand new NVIDIA cards with all new (or at least highly refreshed) architecture.

 

[Kanan]

Anyone else getting that feeling Vega will be the next R9 290X ? No one really expected that card to be as savage as it ended up to be. It was so powerful they could afford rebranding it to R9 390X and it was still competitive against brand new NVIDIA cards with all new (or at least highly refreshed) architecture.

For me it feels more like a revolution than a evolution, 290X was just a big 7970. I think Vega will be something special, something new.

 

[Xzibit]

Vega will be the new Fury

Vega
Vega X
Vega Nano
Vega Pro (Vegi)

Having it in Xbox Scorpio will make it age well like their current GCN cards. A higher clocked variant of whats inside could be what replaces Polaris down the line.

The memory sharing sets up there hints at Navi next gen memory. Raja hinted a while back on the possibility of Multi GPUs on a single interposer.

 

[Brusfantomet]

Anyone else getting that feeling Vega will be the next R9 290X ? No one really expected that card to be as savage as it ended up to be. It was so powerful they could afford rebranding it to R9 390X and it was still competitive against brand new NVIDIA cards with all new (or at least highly refreshed) architecture.

Until we have the actual specs we will not know, but I will try anyway, because i have been drinking.

The power of Hawaii (the chip for 290x/390x) comes from brute force, AMD showed as many cores as they dared into that chip together with an unheard of 512 bit GDDR memory buss with what was at the time fast chips, this resulted in a card that has the potential to melt plastic under oc, and loves water cooling, basically a chip hellbent on being fastest and doing it with brute force, since the power is there driver optimizing has good headroom and that is why the 290x launched against the 780 ti, but sill plays all games on 1080p just fine, as long as you do not care that it might break some standards (max 300 W power? ha). It is interesting to note that the last time Nv tried that it resulted in the GF100 powering the GTX480, it took Nv half a year to stamp out the GF110 powering the GTX 580 help, after that Nv went back to the drawing board and came up with the Kepler architecture in the GK100 powering the GTX 680 that finaly could match the WLIW 4/5 architecture from ATI/AMD in power/performance.

Vega on the other hand, has a different approach from Hawaii, first page, second paragraph:

The idea behind "Vega" is that conventional approaches to making GPUs, so AMD believes, will not be able to cope with emerging workloads.

Therefore they have completely changed the design philosophy from previews GCN chips, where the old bread relied on the fact that AMD had the theoretical computing advantage over the corresponding Nv architecture they were only able to match the smaller Nv chips. Now AMD is trying to spend less silicon on sheer compute power (the memory bandwidth is still massive) and in stead have introduced an local cash structure to feed their admittedly superior crunching cores.

That leaves us with a core that is less like what we have seen from GCN previously, in stead it relies more on the drivers and other tricks to get speedups. This will require closer supervision from AMD to ensure that the chip preforms optimally. therefore it will actually be the most Nv like chip from AMD in a long time, as Nv ever since the GTX 680 have been hard on the optimizing game.

Vega will NOT be at its optimal when lanced. the architecture is built to be smart, not brute force. Therefore it will not be as god as the 290X. that been said, a lot of the smarts is in the silicone, wile Nv still does most of its optimizing in the driver.

So in short, no, I do not think Vega will be a new 290x.
I am still going to seriously consider getting one after i see the benchmarks.

 

[Steevo]

Until we have the actual specs we will not know, but I will try anyway, because i have been drinking.

The power of Hawaii (the chip for 290x/390x) comes from brute force, AMD showed as many cores as they dared into that chip together with an unheard of 512 bit GDDR memory buss with what was at the time fast chips, this resulted in a card that has the potential to melt plastic under oc, and loves water cooling, basically a chip hellbent on being fastest and doing it with brute force, since the power is there driver optimizing has good headroom and that is why the 290x launched against the 780 ti, but sill plays all games on 1080p just fine, as long as you do not care that it might break some standards (max 300 W power? ha). It is interesting to note that the last time Nv tried that it resulted in the GF100 powering the GTX480, it took Nv half a year to stamp out the GF110 powering the GTX 580 help, after that Nv went back to the drawing board and came up with the Kepler architecture in the GK100 powering the GTX 680 that finaly could match the WLIW 4/5 architecture from ATI/AMD in power/performance.

Vega on the other hand, has a different approach from Hawaii, first page, second paragraph:
Therefore they have completely changed the design philosophy from previews GCN chips, where the old bread relied on the fact that AMD had the theoretical computing advantage over the corresponding Nv architecture they were only able to match the smaller Nv chips. Now AMD is trying to spend less silicon on sheer compute power (the memory bandwidth is still massive) and in stead have introduced an local cash structure to feed their admittedly superior crunching cores.

That leaves us with a core that is less like what we have seen from GCN previously, in stead it relies more on the drivers and other tricks to get speedups. This will require closer supervision from AMD to ensure that the chip preforms optimally. therefore it will actually be the most Nv like chip from AMD in a long time, as Nv ever since the GTX 680 have been hard on the optimizing game.

Vega will NOT be at its optimal when lanced. the architecture is built to be smart, not brute force. Therefore it will not be as god as the 290X. that been said, a lot of the smarts is in the silicone, wile Nv still does most of its optimizing in the driver.

So in short, no, I do not think Vega will be a new 290x.
I am still going to seriously consider getting one after i see the benchmarks.

Unless AMD is planting some very intuative AI and enough memory to back up what it learns or use driver updates to feed the AI new information with DX12 they will play less of a role unless they are going to start using their drivers to start prefetching data into vmem, and if that is the case the drivers will play a huge part, but current frame stalls in their whole product stack are caused by shitty latency times in their cache and obtusely deep pipelines for out of order execution mixed with mediocre prefetch logic. I seriously hope this doesn't cause fewer but more noticeable stalls when the AI or whatever fails and causes a whole queue to dump to get the needed data in.

 

[Xzibit]

Raja interviewed at CES 2017 mostly about Vega arch and some other interesting questions

 

[hat]

This isn't a review, this is a hype article.

I really hope AMD has something big to release, but I've seen this before from them. Lots of hype surrounding a product and when it comes time for the product to release we find out it's not that great... but I honestly hope they do make a great graphics card this time, and likewise I hope they make a great processor with Zen.

 

[Chaitanya]

 

[jesdals]

What is the strange card extension on that VEGA card - see the blue connector and extra length of circuit board??? Ahhh - got it from the video But 6 month away :-(

 

[the54thvoid]

It's going to be June. Isn't it.

 

[Frick]

Since its a multi chip module I don't think it will be under $600 at launch.
If it beats th 1080 and touches the titan, like what we are hoping for it will probably be $700+.

Nv will then be able to lower the price of the 1070/1080 to be competitive, since there is no way a complex chip solution like Vega is cheaper to make than a mature GP104, but it could aslo be that Nv wants to keep their good margins on the GP104 series.

This is what I'm thinking too. I see way to many "1080 perf $400 OMGOMGOMG" posts on the tubenets. Ditto for Ryzen. That kind of thinking is what will sink AMD.

 

[Vlada011]

Depend of price and performance I'm open and for Vega option, not only GTX1080/GTX1080Ti.
Best option would be if GTX1080 price drop of Founders Edition.

 

[Super XP]

Vega the king of the Graphics Card. Can't wait.

 

[Captain_Tom]

Anyone else getting that feeling Vega will be the next R9 290X ? No one really expected that card to be as savage as it ended up to be. It was so powerful they could afford rebranding it to R9 390X and it was still competitive against brand new NVIDIA cards with all new (or at least highly refreshed) architecture.

I would say it will be something like the new 290X with the efficiency revolution the 7970 was.


Based on leaked specs it should be at least trading blows with the 1080 Ti, and if its arch really is as much better as they say it is: it stands a decent chance of just straight up crushing the everything Nvidia has...now. Remember that they still could launch a full GP100 card if they wanted to.


Titan Pascal Black could have 3840-SP's, 11000 MHz x 384-bit, and 10 - 20% higher core clocks than what they have now. So if AMD really wants to dominate, they will need a card more than 20% stronger than the 1080 Ti. Vega could pwn that hard, but idk that would be at the upper limits of my optimistic estimates...

 

[RejZoR]

NVIDIA won't be releasing any new Pascals. Next one is GV100, the Volta core.

 

[Captain_Tom]

NVIDIA won't be releasing any new Pascals. Next one is GV100, the Volta core.

Eh I am not sure about that. In the short term I would agree that they won't be able to just retool a GP100 for gaming immediately. However if Vega and the 500 series gave them a lot of trouble, I would expect a full Pascal refresh series either late summer or early fall in order to try and hold AMD off till Volta. It would be led by a new Titan card 20-30% stronger than what they have now...

 

[kruk]

A device named AMD 687F:C1 (possibly Vega) appeared in OpenCL benchmark CompuBench: https://compubench.com/device.jsp?benchmark=compu15d&os=Windows&api=cl&D=AMD+687F:C1&testgroup=info (you can use the site to compare cards side by side: example)

It has 64 compute units and seems to be currently working at 1000 MHz base with 1200 MHz boost. It currently doesn't seem near the MI25 specs (sub 10 TFLOPS vs 12.5 TFLOPs of FP32). With higher clocks it could be really interesting.

We will see, not long now ...

 

[Captain_Tom]

A device named AMD 687F:C1 (possibly Vega) appeared in OpenCL benchmark CompuBench: https://compubench.com/device.jsp?benchmark=compu15d&os=Windows&api=cl&D=AMD+687F:C1&testgroup=info (you can use the site to compare cards side by side: example)

It has 64 compute units and seems to be currently working at 1000 MHz base with 1200 MHz boost. It currently doesn't seem near the MI25 specs (sub 10 TFLOPS vs 12.5 TFLOPs of FP32). With higher clocks it could be really interesting.

We will see, not long now ...

Nice tip!

I am currently comparing it to a Fury X, and it seems to be about 15 - 30% stronger at the same clockspeeds of 1000 MHz. If AMD can indeed clock this at 1500 MHz, it would destroy everything out right now lol. At the very least it looks like it will touch the 1080 Ti.

 

[Super XP]

Eh I am not sure about that. In the short term I would agree that they won't be able to just retool a GP100 for gaming immediately. However if Vega and the 500 series gave them a lot of trouble, I would expect a full Pascal refresh series either late summer or early fall in order to try and hold AMD off till Volta. It would be led by a new Titan card 20-30% stronger than what they have now...

I don't think Nvidia would bother with a refresh IMO, trying to squeeze out 20-30% more would be one hot running card.

Based on what I've read with Vega, AMD is concentrating on high efficiency and trying to eliminate waste. If they can achieve such a thing, it might be able to clock quite high.

New Design. Hopefully it's a Ryzen all over again.

 

[FordGT90Concept]

Anyone else getting that feeling Vega will be the next R9 290X ? No one really expected that card to be as savage as it ended up to be. It was so powerful they could afford rebranding it to R9 390X and it was still competitive against brand new NVIDIA cards with all new (or at least highly refreshed) architecture.

Don't want to get that hyped. It theoretically could but AMD is getting trounced in D3D11 where most of the benchmarks still are. I can't see AMD putting much/any effort into D3D11. AMD is clearly focused on D3D12, Vulkan, VR, HDR, and 4K. Knowing that, I expect something like Fury where it is okay'ish at 1920x1080 (so many idle compute units and hurt by the low clockspeed) but kicks ass at 3840x2160 for a $300-500 price.

It has 64 compute units and seems to be currently working at 1000 MHz base with 1200 MHz boost. It currently doesn't seem near the MI25 specs (sub 10 TFLOPS vs 12.5 TFLOPs of FP32). With higher clocks it could be really interesting.

Seems legit for Vega. Big chip, low clockspeed.

New Design. Hopefully it's a Ryzen all over again.

It's GCN 5.0. Not new, just improved.


Pretty sure they're planning to rebrand Polaris chips with GDDR5X and faster clocks (1.3-1.4 GHz) as RX 580/RX 570 too.

 

[RejZoR]

I always laugh at how "bad" Fury X was in DX11 and then you see it's all like 60+ FPS ultra settings. Like that's "bad" somehow...

Btw, Vega is not "just improved" anymore. Polaris was just that, Vega, it has basically everything new from tile based pixel engine, all new vertex unit, new memory controller paired with HBM2 and can utilize other memory for rendering, flexible compute units that can shift modes and mix them depending on workloads etc. And it's not GCN 5.0, it's NCU 1.0. You may argue it's just a name change, but when something is so drastically different, it needs a whole new name.

I know it's all just speculations, but the thing is, it HAS TO match GTX 1080Ti. Roughly at least. If they don't, then they are majorly screwed. Now, the GTX 1080Ti is out, RX Vega still isn't. Either RX Vega is already awesome and AMD is just chuckling in evil as they are finalizing it or they are working with 150% to do something with it last minute so to speak to better match GTX 1080Ti. I sure hope it's the first one or at least something in between, so that final product will be basically trading blows with GTX 1080Ti and they could 100% focus on next product that could counter NVIDIA's Volta. That would be ideal for us consumers and for AMD.

 

[Kanan]

It's rather unlikely that it'll match the 1080 Ti. I only think that'll be the case when it's fully utilized, meaning DX12/Vulkan + usage of Primitive Shaders and the other new stuff exclusive to Vega, then I expect it even to be faster. Other than that, it'll most likely stick between 1080 and 1080 Ti, so they can price it at 500-600 bucks. It will be the best GPU for users who want a GPU for a long time. Pascal is just a refresh + some tweaks of Maxwell and will be kinda outdated once Vega is released.

 

[FordGT90Concept]

I always laugh at how "bad" Fury X was in DX11 and then you see it's all like 60+ FPS ultra settings. Like that's "bad" somehow...

Bad compared to Pascal and Maxwell which get a lot higher framerates at 1920x1080. 60 fps isn't good for a 144 Hz monitor, for example.

Btw, Vega is not "just improved" anymore. Polaris was just that, Vega, it has basically everything new from tile based pixel engine, all new vertex unit, new memory controller paired with HBM2 and can utilize other memory for rendering, flexible compute units that can shift modes and mix them depending on workloads etc. And it's not GCN 5.0, it's NCU 1.0. You may argue it's just a name change, but when something is so drastically different, it needs a whole new name.

Official AMD press release:
http://www.amd.com/en-us/press-releases/Pages/vega-amds-new-2017jan05.aspx
NCU describes the "Next-generation compute engine" section perfectly. The whole package is GCN 5.0. It's really no more significant than previous iterations of GCN.

I know it's all just speculations, but the thing is, it HAS TO match GTX 1080Ti. Roughly at least. If they don't, then they are majorly screwed. Now, the GTX 1080Ti is out, RX Vega still isn't. Either RX Vega is already awesome and AMD is just chuckling in evil as they are finalizing it or they are working with 150% to do something with it last minute so to speak to better match GTX 1080Ti. I sure hope it's the first one or at least something in between, so that final product will be basically trading blows with GTX 1080Ti and they could 100% focus on next product that could counter NVIDIA's Volta. That would be ideal for us consumers and for AMD.

Why? AMD's aim is market share and profits, not to play king. $700 cards like GTX 1080 Ti hardly make any sense in achieving that goal. Radeon has been doing great without trying to outdo NVIDIA.

Vega is not out yet because, as they said at the Capsaicin event, they're still working on it with partners. Vega is more aimed at Volta than Pascal.

 

[kruk]

If it can't compete with Titan X or 1080 Ti, why would they name it RX Vega and not RX 580/590? And Volta should have x70 running as fast 1080 Ti! Making fun of it with a chip twice as large would be just bad marketing. Also 1080 Ti price surprises me greatly. It could be $800 (and have a $900 FE) if big Vega is really not a threat, and people would still buy it. I doubt we will see top Vega under $600 ... it's big and expensive, so it won't be able to undercut nVidia cards by that much.

I'm still cautiously optimistic, because I know most of their R&D cash was probably spent on Ryzen, but I think another 980 Ti vs Fury X is perfectly doable.

 

[Captain_Tom]

I always laugh at how "bad" Fury X was in DX11 and then you see it's all like 60+ FPS ultra settings. Like that's "bad" somehow...

Btw, Vega is not "just improved" anymore. Polaris was just that, Vega, it has basically everything new from tile based pixel engine, all new vertex unit, new memory controller paired with HBM2 and can utilize other memory for rendering, flexible compute units that can shift modes and mix them depending on workloads etc. And it's not GCN 5.0, it's NCU 1.0. You may argue it's just a name change, but when something is so drastically different, it needs a whole new name.

I know it's all just speculations, but the thing is, it HAS TO match GTX 1080Ti. Roughly at least. If they don't, then they are majorly screwed.

Haha I just like how we are completely on the same page here!


I agree 100% that it HAS to meet or beat the 1080 Ti. I mean the specs are over double the 480, and it has a COMPLETELY new architecture. If it doesn't beat the 1080 Ti they screwed up as bad as bulldozer!