Radeon Vega Architecture Preview Thread

[Veradun]

There will be at least two. I don't think AMD ever releases just one chip, usually its chip + slightly cut down (which is the best bang/buck). Often at release the cut down ones are even full or mostly full chips and can be software unlocked, and then the later releases are physically cut dies.

The only HBM product shipped by AMD so far had 3 flavours: full, cut and special ops (nano). I agree we'll have at least 2 SKUs :>

 

[Phynaz]

I don't think AMD ever releases just one chip

I started to make a list of AMD GPU launches that only released one chip. I stopped after I got to 7.

 

[crisium]

As far as new AMD high end goes, they only release 1 chip at launch but the cut-down within the next few weeks. Tahiti, Hawaii, and Fiji are the last 3 AMD high end, and they all did this. But less than a month makes it essentially a launch window, so it depends on how literal you want to be, but it's almost certain there will be a slightly cut-down chip in less than a month after the full-fat arrives.

 

[Bacon1]

I started to make a list of AMD GPU launches that only released one chip. I stopped after I got to 7.

Glad to see you listed them here for us. Please try to post more than just garbage / troll posts. Contribute.

I'll help you out though:

• 2900 GT, Pro, XT
  
• 3850, 3870 (and X2 versions)
• 4830, 4850, 4860, 4870, 4890 (and a few X2 versions)
• 5830, 5950, 5870, 5970 (dual)
• 6930, 6950, 6970, 6990 (dual)
• 7870 XT, 7950, 7970, 7990 (dual)
• 290, 290x, 295x2 (dual)
• 390, 390x, Fury, Fury X
• 470, 480

So I don't see a single time they only had one version of their dies

 

[Glo.]

I agree with Phynaz. If anything we will see first its the big Vega. Small(er) Vega will come later. With possibly cut down versions, of big Vega.

 

[Phynaz]

Glad to see you listed them here for us. Please try to post more than just garbage / troll posts. Contribute.

I'll help you out though:

• 2900 GT, Pro, XT
  
• 3850, 3870 (and X2 versions)
• 4830, 4850, 4860, 4870, 4890 (and a few X2 versions)
• 5830, 5950, 5870, 5970 (dual)
• 
  
• 6930, 6950, 6970, 6990 (dual)
• 7870 XT, 7950, 7970, 7990 (dual)
• 290, 290x, 295x2 (dual)
• 390, 390x, Fury, Fury X
• 470, 480

So I don't see a single time they only had one version of their dies

Despite your insulting me I'll make one final reply to you before you join my ignore list.

https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units

There's the launch dates. Man up and admit you were wrong to the rest of the forum. I wont see your reply.

 

[MrTeal]

You two seem to be discussing different things. Bacon1 was replying to Dave2150's post about whether there would just be one card or another with a cut down version of the chip. AMD often launches like this, or the xx50 version is delayed by a week or two. Phynaz and Glo seem to be saying that AMD will only launch the one die now, and the smaller Vega will follow later.

Both are quite possibly (and likely, IMO) true.

 

[ElFenix]

Glad to see you listed them here for us. Please try to post more than just garbage / troll posts. Contribute.

I'll help you out though:

• 2900 GT, Pro, XT
  
• 3850, 3870 (and X2 versions)
• 4830, 4850, 4860, 4870, 4890 (and a few X2 versions)
• 5830, 5950, 5870, 5970 (dual)
• 6930, 6950, 6970, 6990 (dual)
• 7870 XT, 7950, 7970, 7990 (dual)
• 290, 290x, 295x2 (dual)
• 390, 390x, Fury, Fury X
• 470, 480

So I don't see a single time they only had one version of their dies

Despite your insulting me I'll make one final reply to you before you join my ignore list.

https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units

There's the launch dates. Man up and admit you were wrong to the rest of the forum. I wont see your reply.

As MrTeal points out you're both clearly talking past each other. Try to keep it civil.

AT Moderator ElFenix.

 

[Headfoot]

Despite your insulting me I'll make one final reply to you before you join my ignore list.

https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units

There's the launch dates. Man up and admit you were wrong to the rest of the forum. I wont see your reply.

Well since AMD bought ATI in 2006, AMD's releases began with HD 3000.

• 3850/3870; Nov 19 07
• 4850/4870; Jun 25, 2008. Only 1 die, but 2 skus based on memory differentiation
• 5870/5850; Sep 23, Sep 30. Launched within a week of each other
• 6970/6950; Dec 15, 2010
• 7970/7950; Jan 9 2012, and 31 Jan 2012. 3 weeks apart
• 290x/290; 24 Oct 2013, 5 Nov 13, a couple weeks apart.
• FuryX/Fury; 24 June 2015, 14 July 2015. 3 weeks apart.

So technically they tie with either: simultaneous launch of full die SKU + cut down version of SKU; or they launch full die SKU then wait between 1 and 3 weeks and launch the cut down SKU.

Given recent history I predict the full die Vega will launch and then 3 weeks later the cut down version of that die will launch.

ATI's earlier Radeons had a different strategy and you can find a lot that launched the flagship alone. AMD has diverged from that strategy.

 

[Bacon1]

So technically they tie with either: simultaneous launch of full die SKU + cut down version of SKU; or they launch full die SKU then wait between 1 and 3 weeks and launch the cut down SKU.

Exactly. Both companies do it because it makes sense to both 1) Get more media coverage, and 2) Get more upsells because people won't wait the extra weeks. We wouldn't say that Polaris had a single chip because 480 and 470/460 launched later. Same thing on the other side.

 

[raghu78]

AMD will have Vega GPUs top to bottom complete stack according to Lisa Su. This pretty much confirms there is no Polaris refresh. Its all Vega with GFX IP 9.0 and NCU going forward. My guess is we will have a next gen Rx 6xx series as Rx 5xx is already a OEM rebrand of Rx 4xx (Polaris).

https://videocardz.com/65538/amd-radeon-rx-570-spotted-in-samsung-notebook

http://seekingalpha.com/article/404...-results-earnings-call-transcript?part=single

Lisa T. Su - Advanced Micro Devices, Inc.

"Yes. So, in terms of our strategy, I think on both the CPU and the GPU side, we have been on a fairly deliberate path to ensure that we got to a very competitive roadmap. So, on the CPU side, with Ryzen and Naples, we believe we will be quite competitive. On the GPU side, as we launch Vega, we will have a full stack, sort of top to bottom with new hardware. We continue to invest in software, and our approach to software is really around open source and using the ecosystem and using the community and focused on sort of the new APIs. So in gaming, we are very focused on DX12 and Vulkan and on the professional graphics and on the GPU server side, really using our GPUOpen. So, we'll continue to invest in software. No question that that's really critical for the graphics market, but we feel we are making good progress."

 

[crisium]

Interesting. Polaris 10 future definitely in doubt if Vega 11 is very similar, though I'm certain Polaris 11 will be here for a while, as most low end chips are.

 

[Headfoot]

I dont think that is the only interpretation at all. "We will have a full stack, sort of top to bottom with new hardware" easily can mean "With Vega's release (understanding that Polaris has already been released, but it is still "new") we will now have a full stack top to bottom"

I very much doubt Polaris is EOL'd 9-12 months after release.

 

[Snarf Snarf]

I find it curious that Lisa didn't specify what kind of new hardware. Normally there would be a statement like "Top to bottom featuring Vega uArch." but she just says "new hardware" which does not imply that Polaris is EOL. AMD has already shown their modular approach to GPU design allows them to bring in features of another graphics IP into an older arch. Is it possible the lower end will be P10 with some Vega goodies added on (similar to the PS4 Pro APU) to enhance perf/w?

 

[Glo.]

I dont think that is the only interpretation at all. "We will have a full stack, sort of top to bottom with new hardware" easily can mean "With Vega's release (understanding that Polaris has already been released, but it is still "new") we will now have a full stack top to bottom"

I very much doubt Polaris is EOL'd 9-12 months after release.

2 GPU designs(well, three, if you count the APU).

4096 GCN core, 3072 GCN core, and 1024 GCN core chip. "Sort of top to bottom".

 

[Bacon1]

 

[Puffnstuff]

The last radeon I had was a hd 4970 and it did okay at the time but wasn't any kind of powerhouse. If vega can consistently best a gtx 1080 ti I'm down with using one in my next build and if pairing it with a ryzen 7 yields great performance then I'm down with that too.

 

[w3rd]

At 6:11 in video, He begins to explain the rational.


I would also like to add, that too many people are overlooking AMD's new "fabric" technologies and AMD's heterogeneous pursuits. This non-volatile solution would be ideal for feeding 2 GPU's on the same card. (ergo: Vega x 2 on a single card/Vega2).

And didn't a Game Developer at Capsaicin & Cream, talk about near 100% scaling..? Now add an area where BOTH CHIPS can pull from non-volatile VRAM..? I think AMD is holding back and playing a slow hand. If I am right, Volta might not even matter, & next year around this time, the 1080ti will become the 8800GT of this era, at $299. Competing with Vega11x2 cards. While Vega & Volta battle it out upstream.

2017 is going to be a great year for gamers and hardware enthusiasts. Xbox Scorpio this Dec will be the biggest electronic seller in history... Game developers are going to ride that wave. Thnx AMD..

 

[Krteq]

And didn't a Game Developer at Capsaicin & Cream, talk about near 100% scaling..? Now add an area where BOTH CHIPS can pull from non-volatile VRAM..? I think AMD is holding back and playing a slow hand...

It's more complicated than you think. There is a long way to multiple GPU dies acting as one "monolithic GPU". Navi seems to be a first step to that direction, not Vega IMHO.

 

[Glo.]

It's more complicated than you think. There is a long way to multiple GPU dies acting as one "monolithic GPU". Navi seems to be a first step to that direction, not Vega IMHO.

Truth is that biggest problem for getting 100% scaling on dual GPU setups is always syncing the memory, and feeding the resources, accordingly. Vega has EVERYTHING to lift this bottleneck up(Unified Memory on hardware level, and HBCC which allows for direct access to data on the other GPU, and pull required data from it, etc).

Vega is most interesting architecture I have seen for very long time. On paper looks like engineering masterpiece, but... this is AMD. They can screw up best design, with idiotic decision. Lets hope this time everything goes well for them with GPUs.

 

[w3rd]

AMD has cohesive memory technology though...

 

[itsmydamnation]

i would expect multi die GPU to act as one thanks to using an active interposer. With all the Data routing etc being active and having massive bandwdith each SOC just scales sideways cross connected at L2 etc. In that model the bigger question is how/where setup/memory/scan out is handled.

 

[Magee_MC]

It's more complicated than you think. There is a long way to multiple GPU dies acting as one "monolithic GPU". Navi seems to be a first step to that direction, not Vega IMHO.

What if the rumored Vega X2 is AMD's way of testing it out. If they're able to make it work, they can work out all of the bugs so that when Navi comes along they have the process dialed in. Very much along the lines of how they used the Fury chips as a test bed for integrating HBM with an interposer. By doing it this way, they position themselves to have Navi be a completely modular chip using as many processors as they need without having all of the growing pains at the same time.

 

[Magee_MC]

i would expect multi die GPU to act as one thanks to using an active interposer. With all the Data routing etc being active and having massive bandwdith each SOC just scales sideways cross connected at L2 etc. In that model the bigger question is how/where setup/memory/scan out is handled.

This may be the reason that AMD is referring to the memory controller on Vega as a cache controller instead. It's built to route the memory to all of the processors instead of to just one.

 

[Puffnstuff]

New technology is always exciting and I'm hoping that Vega will put Radeon back on the map with stellar performance. NVidia has been too long without real competition just like Intel and its well past time that somebody provided it. Too bad we can't get a new upstart in the game to really heat things up.