AMA: RTG with Raja about AMD VEGA

[NumLock21]

AMD Radeon Technology Group had recently did a AMA answering some of the question we have about AMD VEGA.

 

 

Quote

 

I want to start things off today but saying thank you to everyone for all of your excitement, energy and enthusiasm for all things AMD, and in particular, for Vega.

Earlier this week we were thrilled to launch Radeon Vega Frontier Edition. We think it will have a big impact on machine intelligence and content creators. I also know some of you are disappointed that we didn’t launch RX Vega as well. I wanted to hold this AMA and have an open discussion with you about our Vega launches. And while we're not launching RX Vega today -- so I won’t be talking about pricing or launch date -- there are lots of rumors and innuendo I want to put to bed, and there are plenty of questions I can answer.

 

Official Link

In Raja We Trust

 

Alternative link

https://videocardz.com/69662/raja-koduri-explains-where-radeon-rx-vega-in-reddit-ama

[dalekphalm]

Some interesting tidbits hidden in there.

 

One thing confirmed: We will see some form of RX Vega that is faster at gaming then the Frontier 16GB HBM2 card announced earlier.

 

Raja also hinted at (but did not confirm) a 16GB HBM RX card - I think we can guarantee at least 8GB available for RX Vega, and I would expect a 16GB flagship card (Something perhaps comparable to 1080 Ti or Titan Xp in terms of pricing).

 

No hard specs or pricing or release dates.

 

Raja did indicate we'd see it at Computex, but that it would not be released that week... that week. Which could imply it would be released shortly afterwards, perhaps a week or two after Computex. But, that's pure speculation on my part.

[Taf the Ghost]

A 16GB card, especially with the new memory controller really isn't going to be needed, but they could do a "halo" card later.  A Vega X, pretty much.  Possibly with factory AIO.  Wouldn't surprise me if that got released in, say, October.  Especially when HBM2 yields are higher.

[dalekphalm]

5 minutes ago, Taf the Ghost said:

A 16GB card, especially with the new memory controller really isn't going to be needed, but they could do a "halo" card later.  A Vega X, pretty much.  Possibly with factory AIO.  Wouldn't surprise me if that got released in, say, October.  Especially when HBM2 yields are higher.

Frankly, people were worried that the card might only ship with 4GB (Which might still be true for "small" Vega).

 

The cold, hard, truth is that Vega will sell better if it has "more" RAM than the comparable competition. Regardless of if that extra RAM adds any performance benefit.

[Delicieuxz]

Gamer Vega is not that far away, since it's still being presented at Computex, end of May, just as everyone previously anticipated:

 

Quote

 

RK: We’ll be showing Radeon RX Vega off at Computex, but it won’t be on store shelves that week. We know how eager you are to get your hands on Radeon RX Vega, and we’re working extremely hard to bring you a graphics card that you’ll be incredibly proud to own. Developing products with billions of transistors and forward-thinking architecture is extremely difficult — but extremely rewarding — work. And some of Vega’s features, like our High Bandwidth Cache Controller, HBM2, Rapid-Packed Math, or the new geometry pipeline, have the potential to really break new ground and fundamentally improve game development. These aren’t things that can be mastered overnight. It takes time for developers to adapt and adopt new techniques that make your gaming experience better than ever. We believe those experiences are worth waiting for and shouldn’t be rushed out the door. We’re working as hard as we can to bring you Radeon RX Vega.

On HBM2, we’re effectively putting a technology that’s been limited to super expensive, out-of-reach GPUs into a consumer product. Right now only insanely priced graphics cards from our competitors that aren’t within reach of any gamer or consumer make use of it. We want to bring all of that goodness to you. And that’s not easy! It’s not like you can run down to the corner store to get HBM2. The good news is that unlike HBM1, HBM2 is offered from multiple memory vendors – including Samsung and Hynix – and production is ramping to meet the level of demand that we believe Radeon Vega products will see in the market.

 

 

It also sounds like it should be attractive, price-wise, and will have a BIOS that allows overvolting.

Quote

RK: One thing for certain is that Vega Instinct is well positioned to deliver dramatically better performance per dollar;, and TCO(total-cost-of-system ownership) is probably the most important metric to our machine learning and hyperscale customers and combined with Epic/Naples – Vega is extremely attractive..
You want a joke:)
Vega needs some extra Volta(ge) for overclocking:)

 

[Guest]

1 hour ago, NumLock21 said:

AMD Radeon Technology Group had recently did a AMA answering some of the question we have about AMD VEGA.

 

 

Official Link

In Raja We Trust

 

Alternative link

https://videocardz.com/69662/raja-koduri-explains-where-radeon-rx-vega-in-reddit-ama

I read his reddit post in his voice, am I weird?

[MoonSpot]

44 minutes ago, dalekphalm said:

...it would not be released that week... that week. Which could imply it would be released shortly afterwards...

Or....Earlier.

BAHAHAHAHAHA!

[MoonSpot]

Quote

gfxchiptweeterIn Raja We Trust 43 points 7 hours ago 

 

6) Many argue that vega is just a refined polaris gpu, how would you respond to this ?

My software team wishes this was true:)

Vega is both a new GPU architecture and also completely new SOC architecture. It's our first InfinityFabric GPU as well

 

 

Not sure how to take that given what we've seen I.F. do between CCX complexes CPU side.  Think this hints at one of the VEGA GPUs being a dual gpu.

[Cartoons Plural]

Always exciting to see new gpu architectures.

[f22luke]

1 hour ago, dalekphalm said:

Frankly, people were worried that the card might only ship with 4GB (Which might still be true for "small" Vega).

 

The cold, hard, truth is that Vega will sell better if it has "more" RAM than the comparable competition. Regardless of if that extra RAM adds any performance benefit.

I agree, sadly the mindless masses still hold to the philosophy that bigger is always better.  That said if there marketing department repeats what they did with the FuryX it will put the bandwidth comparison front and center.  This of course will just confuse the mindless masses (MM) more; requiring most maddening meticulous musings.  Much mayhem will matriculate amongst the MM maxing when muddled mush melded with the melted mucus, that must have been their brains molds to solid mounds.  Many mops will be mandated to move the mounds with most malice before mold manages to mutilate the ........ woops I was doing it again wasn't I.  

[dalekphalm]

9 minutes ago, MoonSpot said:

 

Not sure how to take that given what we've seen I.F. do between CCX complexes CPU side.  Think this hints at one of the VEGA GPUs being a dual gpu.

I don't think so - at least, not exclusively.

 

I could certainly see a dual Vega GPU, especially aimed at the Pro market for content creators, akin to the Radeon Pro Duo. But I don't think, for example, the Vega Frontier card they showed off is a dual GPU - I firmly think that's a single GPU card.

 

Now, take 2 of those together for ~25 TFLOPS of gaming performance? I wouldn't say no.

[maybethisnamewillwork]

• Raja is keeping his beard until Vega launches.
  
  So we can assume the launch won't be that far away, or else he will look like santa claus 

[XenosTech]

Gotta love invisible text

 

On topic: I really hope vega delivers... I need to have a collection of GPU's again like I did when AGP was still a slot on the motherboard.

[captain cactus]

17 minutes ago, maybethisnamewillwork said:

• Raja is keeping his beard until Vega launches.
  
  So we can assume the launch won't be that far away, or else he will look like santa claus 

Paper launch during Computex, products launching in June. That still qualifies as H1 2017 so he's keeping it for at least 6 more weeks.

[Valentyn]

4 hours ago, Taf the Ghost said:

A 16GB card, especially with the new memory controller really isn't going to be needed, but they could do a "halo" card later.  A Vega X, pretty much.  Possibly with factory AIO.  Wouldn't surprise me if that got released in, say, October.  Especially when HBM2 yields are higher.

That's too far out. Neither Samsung or Sk Hynix have even shown or listed their ability to manufacture 8GB HBM2 modules, and now they're are managing to supply AMD. 

 

I expect 8GB Vega, so two 4GB modules to be sourced from both Samsung and Hynix at the moment. 

 

I expect a Computex  announcement and in shops within the month. 

[Valentyn]

3 hours ago, MoonSpot said:

 

Not sure how to take that given what we've seen I.F. do between CCX complexes CPU side.  Think this hints at one of the VEGA GPUs being a dual gpu.

That's Navi that will be multi GPU when they move to their scalable cores. 

 

Infinity Fabric was a listed for GPUs since they first unveiled it, and can scale to 512GB/s on them. 

I think it's part of what makes their HBCC work. 

[Trixanity]

58 minutes ago, Valentyn said:

That's Navi that will be multi GPU when they move to their scalable cores. 

 

Infinity Fabric was a listed for GPUs since they first unveiled it, and can scale to 512GB/s on them. 

I think it's part of what makes their HBCC work. 

Well, that and the fabric needs to be built into the architecture anyway since they plan to use it in data centers and high performance compute where they need to connect multiple GPU and CPU. So it trickles down automatically. They probably also want to prepare for an eventual multichip Navi that uses the fabric to connect them. So they need us to beta test it for them.

[Valentyn]

14 minutes ago, Trixanity said:

Well, that and the fabric needs to be built into the architecture anyway since they plan to use it in data centers and high performance compute where they need to connect multiple GPU and CPU. So it trickles down automatically. They probably also want to prepare for an eventual multichip Navi that uses the fabric to connect them. So they need us to beta test it for them.

Raja did state that Vega is the start of a new architecture for them, and that the driver team wishes it was more like Polaris to work with. 

 

It goes back to what Raja said during computex 2015. They see the future for GPUs to be as multi core as CPUs and as scalable. Especially since making smaller dies is easier and cheaper. 

 

Ryzen  is similar with their CCX, so we get "big" Vega, and then 7nm Vega 20, which leads to Navi on their road maps. 

 

A quad die Navi with improved  compute would be fantastic in Deep Learning and Enterprise, especially since Naples already sports 128 PCIe lanes and will also use Infinity Fabric between it all. 

[captain cactus]

1 hour ago, Valentyn said:

Raja did state that Vega is the start of a new architecture for them, and that the driver team wishes it was more like Polaris to work with. 

 

It goes back to what Raja said during computex 2015. They see the future for GPUs to be as multi core as CPUs and as scalable. Especially since making smaller dies is easier and cheaper. 

 

Ryzen  is similar with their CCX, so we get "big" Vega, and then 7nm Vega 20, which leads to Navi on their road maps. 

 

A quad die Navi with improved  compute would be fantastic in Deep Learning and Enterprise, especially since Naples already sports 128 PCIe lanes and will also use Infinity Fabric between it all. 

Forget that, how about 4 1024-SP dies on a single interposer for a 4096-SP part at the price of like an RX580.

[leadeater]

5 hours ago, MoonSpot said:

 

Not sure how to take that given what we've seen I.F. do between CCX complexes CPU side.  Think this hints at one of the VEGA GPUs being a dual gpu.

The Infinity Fabric is used for much more than that type of thing. It's actually a protocol stack that can be carried over many different transport mediums and allows for system components to communicate using a common native protocol. This allows for all components involved to act as a single SoC, that could be anything.

 

This is all based on Gen-z technology and is deeply intertwined with AMDs technologies including the High Bandwidth Cache Controller. Think about how useful it would be to be able to grab information from storage but treat it as memory address space in the SoC then move that information in to RAM then do some processing then move that over to a GPU for more processing then back all using a single protocol all acting as one system.

[leadeater]

2 hours ago, Valentyn said:

That's too far out. Neither Samsung or Sk Hynix have even shown or listed their ability to manufacture 8GB HBM2 modules, and now they're are managing to supply AMD. 

 

I expect 8GB Vega, so two 4GB modules to be sourced from both Samsung and Hynix at the moment. 

 

I expect a Computex  announcement and in shops within the month. 

There are real Nvidia cards right now in the hands of customers with HBM2 memory. The thing that isn't clear is that AMD said they would be using faster HBM2 chips than is available and those a projects to arrive much later than the indicated Vega launch.

[Valentyn]

2 minutes ago, leadeater said:

There are real Nvidia cards right now in the hands of customers with HBM2 memory. The thing that isn't clear is that AMD said they would be using faster HBM2 chips than is avaliable and those a projects to arrive much later than the indicated Vega launch.

 

Those NVIDIA cards are Samsung 1.4Gbps 4GB Modules.

AMD are partnered with SK Hynix, that promised 2.0Gbps by Q3 2016, and only show 1.6Gbps for Q2 2017 now.

It looks like the Radeon Frontier Edition has 1.88Gbps modules, so those are most likely even lower quantities; hence 480GB/s Bandwidth as opposed to the original Vega spec of 512GB/s for 2 HBM2 modules.

[Valentyn]

13 minutes ago, lots of unexplainable lag said:

Forget that, how about 4 1024-SP dies on a single interposer for a 4096-SP part at the price of like an RX580.

Most likely the plan as well, as Raja during Computex 2016 talked about it all, and how many smaller dies are the way to go again.

So by 2019 I expect Radeons to be like Ryzens with multiple cores.

 

History hopefully doesn't repeat itself, and we have DX12 and Vulkan multi-gpu support more.

[Carclis]

6 hours ago, maybethisnamewillwork said:

• Raja is keeping his beard until Vega launches.
  
  So we can assume the launch won't be that far away, or else he will look like santa claus 

So a Christmas launch confirmed then...?

[Taf the Ghost]

5 hours ago, Valentyn said:

That's too far out. Neither Samsung or Sk Hynix have even shown or listed their ability to manufacture 8GB HBM2 modules, and now they're are managing to supply AMD. 

 

I expect 8GB Vega, so two 4GB modules to be sourced from both Samsung and Hynix at the moment. 

 

I expect a Computex  announcement and in shops within the month. 

Right, my expectation is we get a RX Vega Fury or RX Vega Starburst in November.  Like a Ti model.

5 hours ago, Valentyn said:

That's Navi that will be multi GPU when they move to their scalable cores. 

 

Infinity Fabric was a listed for GPUs since they first unveiled it, and can scale to 512GB/s on them. 

I think it's part of what makes their HBCC work. 

More than likely.  Infinity Fabric isn't a technology; it's an entirely different design approach.  Which is what makes is so awesome.  It's actually a sign of some really, really good engineers doing something awesome.

 

4x RX 560 with a bit lower clocks could fit on one PCB would pretty much toast a current Titan Xp. That is current, cheap chips against a $1200 card.  Or 4x $200 ($800) vs $1200.  Whereas the new Volta is yielding about 1 chip per wafer in their first runs for the professional cards.  When you're talking upwards of >20 Billion transistors in future, large chips, that's just a failure rate that's going to be too high.