RTG Raja's going on a break, Lisa Su is stepping in for the next few months

[Doobeedoo]

8 minutes ago, cj09beira said:

Vega 20 will probably be Vega with fp64, so it will be really compute focused, also easy to see by the crazy amount of hbm2 on it

Yeah I guess so with like 32GB HBM2 also maybe dual GPU card with 16GB who knows. 

[samcool55]

10 minutes ago, Lathlaer said:

That was just rough conservative estimation, obviously. I just watched a video comparing Witcher 3 where 780ti was pulling about 37 fps whereas 1080ti was doing 87-89. That is about 137% increase in performance. 980ti was doing almost double, a bit less than 100% increase. 

 

 

The MSRP price increase for high end NVIDIA gpus has been almost 0% from 780ti.

 

I am leaving Titans out because people who pay for Titans don't really care about stuff like that. They are either techtubers who get it for free or are willing to pay premium for the cool name and to be able to claim that they have the current fastest thing a consumer can get. 

Well that's the difference between first and last point but it goes from 460 dollars up to almost 1000 dollars. Also back in the day before the titan those cards were the best that nvidia could possibly do. However that's no longer true so even if you say it changed 0% it's not completely correct because the 1080ti isn't their flagship while the geforce 2 ultra was. Also if you would choose an earlier starting point or finishing point the 0% would change because it's incorrect.

 

If you would go from the 8800 Ultra to the GTX 285 the conclusion would be with the same logic "flagship gpu's are decreased over 50% in price" which is clearly a wrong conclusion.

 

To compare it with cars, the fastest production car a few decades ago was the jaguar type E, if you would compare it to today's fastest supercar the result would be incorrect because these days there's basically a class above that (hypercars) and it's those that are the real fastest production cars so it's one of them you have to use to compare top speeds.

 

The Titan cards are the fastest cards out there for gaming today so that's what you have to use to compare even if you believe they are stupid, (which i do agree with) it's not a valid reason to ignore them.

 

The 1080 for example was meant to replace the 980 and not the 980ti. Also as time went on MSRP's changed. The 780 for example dropped to 500 MSRP once the 780Ti launched. 

 

Also about that percentage, what you say is 1 measuring point between 2 generations. That 50%-60% number i said however can be proven. (well it was a vague guess from a graph that's enough to calculate the exact average percentage improvement for a new generation.)

The graph is at 19:21.

 

 

[Sniperfox47]

4 hours ago, Misanthrope said:

No that's mostly accurate and plausible, but it would still kinda mean admitting defeat on Vega: There's still promised features that are flat-out not active in Vega so at the very least the engineers should be working hard at driver support at this point.

 

So if what you say turns out to be accurate it means Su thinks Vega is not salvageable and that they can just move the current stock to miners and workstation clients (who obviously do not need the missing features) which is worrisome.

 

Actually the entire situation with Radeon Group is fairly worrisome to me: for them to "fix" things we'd be talking very long term development of a brand new architecture from the ground up (like Zen) and that's 4 to 5 years in the making before it'd be ready.

How do you go from "They're pushing their APUs forward *with* Vega" to "They're abandoning Vega"? Part of moving forward with their APUs is getting the Vega drivers worked on. The arch isn't going to be substantially different and most of the driver work can be shared between the two (other than stuff like CPU HSA that only the APUs will have, and Crossfire which only the Graphics Cards will have).

 

It's not like they need Raja sitting there prodding them to build the drivers up, the scope of work is pretty clearly defined on the software/firmware side at this point.

 

 

1 hour ago, samcool55 said:

DX12/Vulkan would help a lot with this because the drivers wouldn't be as important anymore compared to today (DX11). However those games are still nowhere to be found and it wouldn't be unreasonable to think AMD hoped games with DX12/Vulkan support would start rolling out about now but that's not the case.

 

I wouldn't be surprised if Vega people in 2-3 years time would say "see Vega was a much better option! F nvidia!" because by then the drivers are optimized, more hardware functions are enabled (hopefully all by then) and DX12/Vulkan is (finally) common. But then again that's not now so it's pointless.

That's the thing though, if it's this hard for them to optimize their own OpenGL drivers, how hard is it going to be for Devs to optimize for it on the Vulkan/DX12 side of things?

 

Sure the drivers for Vulkan are still there assisting, but they're still a lot more reserved and out of the way than DX11.

 

And with DX12 already being incredibly difficult to develop for,  how is the extra complexity of optimizing for Vega going to help us see more titles?

 

44 minutes ago, samcool55 said:

Yeah volta will eventually arrive, but at what price? 

There's a very real possibility we won't see Volta on the consumer side at all. It's already out on the enterprise side but

A) it's really costly to manufacture

B) there's no real competition from Vega to encourage a release

C) honestly most of the changes with it aren't *super* impactful to gaming/prosumer performance anyways.

 

I would not be surprised to see Nvidia skip over it on the consumer side and refresh Pascal again before jumping to their next Arch.

[TidaLWaveZ]

2 minutes ago, cj09beira said:

Gpus class is determined by die size not perf. And if you go back to older generations the 80 class chip is 450 mm2 or more. The 1080 Gpus is in the 300s

I get what you're saying, how the Ti has bumped all the 80 down a die size, but the fact that the classes have switched die sizes just proves that die size does not determine class.

[cj09beira]

1 minute ago, TidaLWaveZ said:

I get what you're saying, how the Ti has bumped all the 80 down a die size, but the fact that the classes have switched die sizes just proves that die size does not determine class.

It just means that they are selling mid range cards at high end prices

[Misanthrope]

1 minute ago, Sniperfox47 said:

How do you go from "They're pushing their APUs forward *with* Vega" to "They're abandoning Vega"? Part of moving forward with their APUs is getting the Vega drivers worked on. The arch isn't going to be substantially different and most of the driver work can be shared between the two (other than stuff like CPU HSA that only the APUs will have, and Crossfire which only the Graphics Cards will have).

 

It's not like they need Raja sitting there prodding them to build the drivers up, the scope of work is pretty clearly defined on the software/firmware side at this point.

It's fairly self explanatory: APU products are very different from dedicated gpu products: the goals are fairly different since one prioritizes power consumption and small die size while the other prioritizes raw performance where power requirements grow exponentially i.e. You can take a fairly amazing architecture for APUs but once you try to translate that into a high powered chip it would well, fizzle away using too much power much in the way we're seeing Vega here.

 

Having entirely separate divisions might not be necessary (or at least not without significantly downsizing) if you can end up focusing on only one side. And lastly, assuming you keep things as is, driver development for consumer cards vs workstation look very differently. The former requires a lot of software specific optimization while the later requires a lot more validation and stability testing, driver certification, etc. If you are forced to choose what to prioritize, the goals of the consumer driver team would be extremely different to those on the workstation team.

[TidaLWaveZ]

1 minute ago, cj09beira said:

It just means that they are selling mid range cards at high end prices

I agree with that, the 1080ti's price pretty much proved it.

[Sniperfox47]

8 minutes ago, Misanthrope said:

~snip~

Yeah, but APUs were already in their roadmap and are needed to really push Zen on the mobile side of things. Note this also includes for mobile GPGPU compute which is becoming increasingly common.

 

Is the work for certified drivers and consumer drivers different? Absolutely, I'm not debating that, but some of these features that aren't finished are incredibly useful for certain Enterprise workflows, and again that workload can be shared across both.

 

You can prioritize something all you want, but if you have money invested into something else that you intended to move, you kind of have to move it if you want to make money.  With their current debt, AMD doesn't have the luxury of sitting on this stuff. It's the reason why Vega had to come out with rushed drivers and the same reason why AMD can't afford to push the APU development.

[TidaLWaveZ]

9 minutes ago, huilun02 said:

RTG is doing badly so they might sack Raja. I mean just look at all those Polaris and Vega cards sitting on store shelves unable to be sold.

That doesn't necessarily mean that they're doing well, supply could be extremely low.

[samcool55]

13 minutes ago, Sniperfox47 said:

 

There's a very real possibility we won't see Volta on the consumer side at all. It's already out on the enterprise side but

A) it's really costly to manufacture

B) there's no real competition from Vega to encourage a release

C) honestly most of the changes with it aren't *super* impactful to gaming/prosumer performance anyways.

 

I would not be surprised to see Nvidia skip over it on the consumer side and refresh Pascal again before jumping to their next Arch.

I guess we'll see volta eventually tho. But probably a smaller GPU design (i don't expect them to put a full volta in the 1080 replacement, only 1080ti and up like they are doing now) which would reduce costs and we already know the consumer ones will probably use GDDR6 and not HBM2 which is again a big cost saving.

 

There is indeed basically no pressure change for nvidia because of vega, i agree. Actually quite sad tbh.

 

Changes are indeed not massive but i prefer any improvement over 0% improvement for the same price if they launch something new/better

 

I don't know if they would refresh Pascal tho, Pascal was actually not planned in the first place and is basically Maxwell on a new node. The only reason Pascal is a thing is because it was good enough to compete with AMD. Volta is probably much older and mature than we think...

[TidaLWaveZ]

1 minute ago, samcool55 said:

I guess we'll see volta eventually tho. But probably a smaller GPU design (i don't expect them to put a full volta in the 1080 replacement, only 1080ti and up like they are doing now) which would reduce costs and we already know the consumer ones will probably use GDDR6 and not HBM2 which is again a big cost saving.

 

There is indeed basically no pressure change for nvidia because of vega, i agree. Actually quite sad tbh.

 

Changes are indeed not massive but i prefer any improvement over 0% improvement for the same price if they launch something new/better

 

I don't know if they would refresh Pascal tho, Pascal was actually not planned in the first place and is basically Maxwell on a new node. The only reason Pascal is a thing is because it was good enough to compete with AMD. Volta is probably much older and mature than we think...

There's one thing you can count on, that it will be announced at a huge convention.

 

My guess CES 2018. I bet we'll here ramblings of the card late this year and it will release in the first half of 2018.

 

[Dylanc1500]

10 minutes ago, samcool55 said:

I guess we'll see volta eventually tho. But probably a smaller GPU design (i don't expect them to put a full volta in the 1080 replacement, only 1080ti and up like they are doing now) which would reduce costs and we already know the consumer ones will probably use GDDR6 and not HBM2 which is again a big cost saving.

They aren't going to use a full fat gv100 when those are costing them ~$1000 just to manufacture. Unless people open up their pockets.

Quote

Jen-Hsun Huang - NVIDIA Corp. 

And so the price of Volta is driven by the fact that, of course, the manufacturing cost is quite extraordinary. These are expensive things to go and design. The manufacturing cost itself, you guys can estimate it, is probably in the several hundred dollars to close to $1,000. However, the software intensity of developing Volta, the architectural intensity of developing Volta, all of the software intensity associated with all the algorithms and optimizing all the algorithms of Volta is really where the value-add ultimately ends up. And so I guess the pricing – your question relates is pricing. We expect pricing to be quite favorable for Volta.

 

[Sniperfox47]

27 minutes ago, samcool55 said:

I don't know if they would refresh Pascal tho, Pascal was actually not planned in the first place and is basically Maxwell on a new node. The only reason Pascal is a thing is because it was good enough to compete with AMD. Volta is probably much older and mature than we think...

Umm... Not sure where you're getting your information from... Pascal's been in the roadmap since early 2014, and has substantially more architectural changes than just a node shrink.

 

They may not manifest as higher IPC for your average gamer, but the changes for VR multiprojection, Supersampling, and how the hardware handles async Compute are pretty substantial, among others.

 

Pascal wasn't just something they threw together last minute, it was a decision to split the changes planned for Volta to help bring it to market faster.

[Humbug]

12 hours ago, djdwosk97 said:

 

Except didn't Raja have (relatively) little to do with Vega (in terms of its final state -- sure he was managing RTG during that time, but there's a limit to what he could do after the architecture has been developed) as it was mostly through production? Isn't Navi the first architecture he's played a big role in? 

Correct Navi will be more indicative of Raja's and RTG's work.

But while RTG and Raja cannot be blamed for the performance of Vega, they can be blamed for the messed up launch schedules and not finishing the driver in time.

[DildorTheDecent]

#RajaOUT

[MEC-777]

RTG is in a rough spot. Raja, I'm sure, has been working his a$$ off with the recent Vega launch (and unfortunately not able to de-throne the 1080Ti). Then all this controversy over pricing and the current mining craze, buying up all AMD cards and hiking prices... It's a big, hot mess right now. 

 

Honestly, I don't blame him for taking a break. I also don't think we should read too much into it either. 

[samcool55]

18 minutes ago, Dylanc1500 said:

They aren't going to use a full fat gv100 when those are costing them ~$1000 just to manufacture. Unless people open up their pockets.

 

Well yes but that price includes 16GB HBM2 which is easily 300 dollars on its own. And the GV100 is more than the GPU itself, it's the complete module. There's also the NVlink stuff which might be expensive as well. 12nm is also still new so i can imagine yields are quite bad, especially if you want a 100% working one, it's based on 12nm which nothing else uses. (polaris and vega are 14 and pascal is 16) And there's also the deep-learning stuff which might be irrelevant for gamers so they could ignore that as well.

 

It's high-end difficult to produce low yield stuff with expensive memory. The 1000$ price tag isn't as crazy as you might think.

 

If there will be a new titan it could be launched with GDDR6 and by then yields will be up and cost will be down so it's not completely insane that they would put a GV100 GPU in a titan or replacement of the 1080ti.

[samcool55]

17 minutes ago, Sniperfox47 said:

Umm... Not sure where you're getting your information from... Pascal's been in the roadmap since early 2014, and has substantially more architectural changes than just a node shrink.

 

They may not manifest as higher IPC for your average gamer, but the changes for VR multiprojection, Supersampling, and how the hardware handles async Compute are pretty substantial, among others.

 

Pascal wasn't just something they threw together last minute, it was a decision to split the changes planned for Volta to help bring it to market faster.

In 2013 it was missing http://www.anandtech.com/show/7900/nvidia-updates-gpu-roadmap-unveils-pascal-architecture-for-2016

Maxwell -> volta, no pascal.

 

And yes of course they would improve the architecture design, but it's nowhere near substantial.

Hardware async compute basically went from broken to sort of working but still a step behind from what a 7970 for example can benefit from async compute.

 

The node shrink allowed a much higher core clock which is the biggest reason for the performance improvement for Pascal compared to Maxwell. 

 

Also, to put the launch date vs announcement date in comparison. Pascal was announced in 2014. It launched in 2016. 2 years is a really small time to develop a new GPU.

[ScratchCat]

13 hours ago, MechPilot524 said:

I think for Vega it's largely too late for anything really noticeable. Drivers, sure, but that's it aside from maybe improving HBM2 availability and fab deals. My hope is that Lisa Su might influence the next generation of cards (or the generation after), finding the right people and laying the right groundwork as she did for Zen.

Technically Vega has 13 TFLOPs of compute performance, more than the Titan XP, it seems they are mainly limited by drivers and the seemingly non functional check-for-covered-pixel/ pre rasteriser  thing (forgot the name )

[Dylanc1500]

42 minutes ago, samcool55 said:

Well yes but that price includes 16GB HBM2 which is easily 300 dollars on its own. And the GV100 is more than the GPU itself, it's the complete module. There's also the NVlink stuff which might be expensive as well. 12nm is also still new so i can imagine yields are quite bad, especially if you want a 100% working one, it's based on 12nm which nothing else uses. (polaris and vega are 14 and pascal is 16) And there's also the deep-learning stuff which might be irrelevant for gamers so they could ignore that as well.

 

It's high-end difficult to produce low yield stuff with expensive memory. The 1000$ price tag isn't as crazy as you might think.

 

If there will be a new titan it could be launched with GDDR6 and by then yields will be up and cost will be down so it's not completely insane that they would put a GV100 GPU in a titan or replacement of the 1080ti.

Oh believe me I know, but most people don't realize the cost of manufacture for something like the full fat g(x)100 GPU package. That also doesn't take consideration for the rest of the board either as the cost is different whether you go PCIe card or mezzanine card.

[Sniperfox47]

1 hour ago, samcool55 said:

And there's also the deep-learning stuff which might be irrelevant for gamers so they could ignore that as well.

Sadly AMD kinda made that relevantish for gamers with having games optimized for Rapid Packed Math on Vega. The deep learning stuff is mostly about FP16 and FP8 throughput, which is what RPM is all about too.

 

Not all of the Deep Learning stuff they cut out for the consumer cards is relevant, but enough of it is that we'll likely either see them ignore RPM at the cost of a few fps in certain games, or (less likely) keep the full throughput fp16 hardware in the consumer chips, at the risk of canibalizing other markets.

[nicolas2465]

15 hours ago, YongKang said:

Well, that's odd. Raja from RTG is going on a break until Q4 (start of 2018) and the CEO is stepping in.

News link: https://www.tweaktown.com/news/59104/rtg-boss-takes-sabbatical-lisa-su-over/index.html

Bear in mind that the RX Vega series has just launched and it's strange that such an important executive such as Raja is taking a kinda long break right now. Could it be an order from the top brass? This is just some speculation but feel free to drop your thoughts down below. This is getting interesting.

He was so busy creating cards that are barely competitive and use tons more power, and also need to be undervolted/bios flashed with other cards' bios'/etc.  He needs a few months to relax man!

[nicolas2465]

Can AMD just fire the entire team that makes the gaming radeon cards and hire some ex-Nvidia employees or something?  The CPU team did awesome with Ryzen.  They need a team like that for Radeon.  AMD GPUs are a fucking mess.  I'd never recommend their cards at this point

[MageTank]

2 hours ago, MEC-777 said:

RTG is in a rough spot. Raja, I'm sure, has been working his a$$ off with the recent Vega launch (and unfortunately not able to de-throne the 1080Ti). Then all this controversy over pricing and the current mining craze, buying up all AMD cards and hiking prices... It's a big, hot mess right now. 

 

Honestly, I don't blame him for taking a break. I also don't think we should read too much into it either. 

Yeah, he caught a ton of flak over the pricing, and the recent driver release that made Vega better for mining (despite AMD saying they were anti-mining). This, on top of the massive hype Vega was unable to live up to, has surely put an insane amount of stress on him. A break would be perfectly reasonable in that kind of situation. 

[YongKang]

1 hour ago, nicolas2465 said:

He was so busy creating cards that are barely competitive and use tons more power, and also need to be undervolted/bios flashed with other cards' bios'/etc.  He needs a few months to relax man!

he had alrdy taken two weeks off some time ago this year. Someone alrdy mention this in the convo.