gtx 1180 leaks and infofmation from wccftech

[leadeater]

34 minutes ago, cj09beira said:

i think they need to do both anyways because even if they go mcm they will endup at the 64rop/cu limit in no time flat (good chance it could happen in the second gen),which one will be first i have no clue, but vega already having IF gluing everything together could point to a design mcm in the works

The IF in Vega isn't used in the core CU structure of the GPU though, it's only used to interconnect things like the memory controller. They need to do what Nvidia has done with Volta, reduce the CUs per Compute Engine and deploy more clusters of them (SMs in Nvidia speak). If they did that they would need to improve the Hardware Schedulers and the ACEs to handle the extra Graphics Pipelines but it would result in less CUs being resource wasted. This can be done without going MCM or IF, it's just a logical structure change (well physical too).

 

Likely other knock on effects but more CUs per Compute Engine isn't scaling out actual performance.

[Taf the Ghost]

23 minutes ago, cj09beira said:

having 2 dies that small could be good or bad, it makes even the 580 tier need more than one die but it could allow for having 64rops and 32 cus which if they can have 4 dies it could be really powerful, hmm there might be a reason to only have 32 cus after all 

Unless they've gotten through the ROP issue, I don't think we're getting 4-way GPUs, as I imagine a 2-way is a lot easier to deal with. Also, with the HBCC that came in with Vega, memory management is actually controlled at the Driver level, so the issues with having to replicate the memory across both GPU dies should be manageable. 

 

I still would assume where likely to see a 48 CU design and that's what they mean by "scalable". While the GPUs will run with CU deactivated, since they've switched to GCN, you've gotten your choice of 8 CU, 16 CU, 32 CU or 64 CU designs.

 

Actually, the Ryzen Mobile parts is an 11 CU design, which is the first unbalanced design we've seen from GCN. Does that mean they can produce a Navi part with however many CUs they want? Rolling out 30 CU and 48 CU designs would probably make a lot more sense on 7nm than another 16 & 32 CU for the mainstream parts.

[cj09beira]

2 minutes ago, Taf the Ghost said:

Unless they've gotten through the ROP issue, I don't think we're getting 4-way GPUs, as I imagine a 2-way is a lot easier to deal with. Also, with the HBCC that came in with Vega, memory management is actually controlled at the Driver level, so the issues with having to replicate the memory across both GPU dies should be manageable. 

 

I still would assume where likely to see a 48 CU design and that's what they mean by "scalable". While the GPUs will run with CU deactivated, since they've switched to GCN, you've gotten your choice of 8 CU, 16 CU, 32 CU or 64 CU designs.

 

Actually, the Ryzen Mobile parts is an 11 CU design, which is the first unbalanced design we've seen from GCN. Does that mean they can produce a Navi part without however many CUs they want? Rolling out 30 CU and 48 CU designs would probably make a lot more sense on 7nm than another 16 & 32 CU for the mainstream parts.

we already got 44cus with the 290x which in my mind is a pretty dam good balance between cus and the rest (when at 64 rops), but you are probably right in that having more than 2 dies would be quite dificult if memory is handled by the driver which for me means they need bigger dies

the way i think is that if the go mcm they really need to push it as far as they can to take as much advantage of that as possible because nvidea is also going to do it sooner or later and it would be very helpful for amd if they had more market share by that time, plus amd really needs its mind share,

i was talking just yesterday with a collegue of my uni that while knowing about intel and amd cpus and having a ryzen gpu himself didn't know amd also made gpus

[laminutederire]

1 hour ago, Taf the Ghost said:

My guess is Nvidia is going to shift their entire stack up about 50USD in MSRP, or more, given the persistent GPU prices (Cryptos sell off in the beginning of the year then rise in the end of the year, for reasons I don't know). By launching into their big selling season at relatively good prices, they should sell out their expected sales stock. AMD's Navi won't effect them until it actually lands.

 

Given that 14nm -> 7nm generations are about 1.5 Nodes worth of previous shrink level, we're really unsure what AMD could be dropping on 7nm with Navi. We can assume it's going to be a fairly sizable generational improvement (node shrinks do that), but that's all where looking at right now. (I do hope it's a 48 CU Navi and then cut-down versions of that die, but we'll see.) 

Yeah Navi won't impact sales then, that's probably why they're releasing it then. Because if rumours are to be true, that'd mean significant price cuts on their part, which hurts their bottom line.

If anything we could hope AMD pull a Pascal on Vega+Polaris and have those efficiency range being at significantly higher range and therefore end up with a significant increase in either power consumption or performance depending on the product. That should help them disassociate themselves from the whole not card image a bit. Ultimately 7nm may be more about price than performance crown. 250$ Navi vs 600 or 700$ 1170 for the same perf? That'd be quite a huge gap.

1 hour ago, Taf the Ghost said:

Yup. AMD has clearly aligned their focus to get on the node as fast as possible, while Nvidia could very easily do a 1 year generation on 12FF then move to TSMC's 7nm process by late 2019, if Navi comes out really good.

 

7nm "generation" is also where almost everyone catches up to Intel, which is pretty much the first time that's happened since about 1999 or something like that.

That's probably what Nvidia is up to. They don't seem to want to radically change their arch for the foreseeable future, so a node shrink would be the way to go for them. Doing it incrementally makes sense because that gives them a little bit more time to change the arch itself.

 

1 hour ago, cj09beira said:

7nm should be a great equalizer for the whole industry, it will be very fun 

Yeah, unfortunately that might get destabilized again with post silicon technologies.

[Taf the Ghost]

1 minute ago, cj09beira said:

we already got 44cus with the 290x which in my mind is a pretty dam good balance between cus and the rest (when at 64 rops), but you are probably right in that having more than 2 dies would be quite dificult if memory is handled by the driver which for me means they need bigger dies

the way i think is that if the go mcm they really need to push it as far as they can to take as much advantage of that as possible because nvidea is also going to do it sooner or later and it would be very helpful for amd if they had more market share by that time, plus amd really needs its mind share,

i was talking just yesterday with a collegue of my uni that while knowing about intel and amd cpus and having a ryzen gpu himself didn't know amd also made gpus

MCM might also end up being a ML/DL/AI only type of product, which is also a place that might profit more from it, since it's just acting as a "many, many cores" processing card at that point.

 

Per wiki, they can put up to 11 CUs in a Shader Engine, so the Vega in Raven Ridge isn't unbalanced. So maybe there is some internal reorganization with Navi of the CU & SE structure, allowing for more a more balanced design. The semi-custom piece for Intel is a 24 CU design, for instance. So maybe it will be something like a 48 CU design as the larger Navi part. We'll find out in about a year.

[Taf the Ghost]

6 minutes ago, laminutederire said:

Yeah Navi won't impact sales then, that's probably why they're releasing it then. Because if rumours are to be true, that'd mean significant price cuts on their part, which hurts their bottom line.

If anything we could hope AMD pull a Pascal on Vega+Polaris and have those efficiency range being at significantly higher range and therefore end up with a significant increase in either power consumption or performance depending on the product. That should help them disassociate themselves from the whole not card image a bit.

That's probably what Nvidia is up to. They don't seem to want to radically change their arch for the foreseeable future, so a node shrink would be the way to go for them. Doing it incrementally makes sense because that gives them a little bit more time to change the arch itself.

 

Yeah, unfortunately that might get destabilized again with post silicon technologies.

Launching Q3 2018 would let Nvidia get the New Generation + Christmas sales, which is their big period, then they could launch a version that's ported to 7nm the next year to eat up those holiday sales.

 

The speed at which they release probably is dictated by AMD, as Nvidia dropped the Titan Xp & lowered the 1080's price right before they expected AMD to launch RX Vega in April 2017. From that, we can get a pretty good idea of Navi's results from what Nvidia does in 2019.

[laminutederire]

Just now, Taf the Ghost said:

Launching Q3 2018 would let Nvidia get the New Generation + Christmas sales, which is their big period, then they could launch a version that's ported to 7nm the next year to eat up those holiday sales.

 

The speed at which they release probably is dictated by AMD, as Nvidia dropped the Titan Xp & lowered the 1080's price right before they expected AMD to launch RX Vega in April 2017. From that, we can get a pretty good idea of Navi's results from what Nvidia does in 2019.

Yeah of course. Right now Navi could be exceptional or a flop. It could be released this summer, or in 2019, nobody knows. It wouldn't be impossible that Nvidia expects Navi this summer and plans to release something at the same time to try and shadow Amd's release. That'd indicate that Navi isn't that good (otherwise... price cuts), or that Nvidia doesn't know the level of the cards contrary to Vega since not much has leaked on Navi compared to Vega ( maybe because the marketing head has changed? ).

[cj09beira]

3 minutes ago, laminutederire said:

Yeah of course. Right now Navi could be exceptional or a flop. It could be released this summer, or in 2019, nobody knows. It wouldn't be impossible that Nvidia expects Navi this summer and plans to release something at the same time to try and shadow Amd's release. That'd indicate that Navi isn't that good (otherwise... price cuts), or that Nvidia doesn't know the level of the cards contrary to Vega since not much has leaked on Navi compared to Vega ( maybe because the marketing head has changed? ).

i wouldn't expect navi until at least january vega 7nm might see a summer release but i don't expect navi to do the same as for that they need much more volume, and then vega wouldn't help much as a pipe cleaner 

[leadeater]

23 minutes ago, Taf the Ghost said:

Per wiki, they can put up to 11 CUs in a Shader Engine, so the Vega in Raven Ridge isn't unbalanced.

That just for Raven Ridge, the 11 CUs per Shader Engine? Vega has 16 CUs and so did Fiji (Fury).

[laminutederire]

2 minutes ago, cj09beira said:

i wouldn't expect navi until at least january vega 7nm might see a summer release but i don't expect navi to do the same as for that they need much more volume, and then vega wouldn't help much as a pipe cleaner 

I personally don't care that much since I don't have any money for those haven't they starting producing prototypes for Navi though?

That and it is rumoured the new PlayStation based on Navi could arrived for Christmas. It would make sense that Navi would be ready before and released for PC as am incentive for developers to get to know its arch in practice before the PS5 would be released. It would also make sense that they'd produce the ps5 parts before and delaying the PC parts afterwards as well.

[Taf the Ghost]

10 minutes ago, cj09beira said:

i wouldn't expect navi until at least january vega 7nm might see a summer release but i don't expect navi to do the same as for that they need much more volume, and then vega wouldn't help much as a pipe cleaner 

There's been persistent rumors of something in Q3, but that's probably Vega on 7nm for ML. And it's probably going to be a display of the very few working silicon. I honestly wouldn't expect Navi until well into Q2 2019. AMD will want to get Zen2 and Epyc2 out the door first. There's also the little issue that AMD will still be selling every GPU they've allocated to make by that point.

[cj09beira]

1 minute ago, Taf the Ghost said:

There's been persistent rumors of something in Q3, but that's probably Vega on 7nm for ML. And it's probably going to be a display of the very few working silicon. I honestly wouldn't expect Navi until well into Q2 2019. AMD will want to get Zen2 and Epyc2 out the door first. There's also the little issue that AMD will still be selling every GPU they've allocated to make by that point.

agreed with the first part, but by then there is a good chance that mining wont last that long, profitability is so low right now

[Taf the Ghost]

10 minutes ago, laminutederire said:

I personally don't care that much since I don't have any money for those haven't they starting producing prototypes for Navi though?

That and it is rumoured the new PlayStation based on Navi could arrived for Christmas. It would make sense that Navi would be ready before and released for PC as am incentive for developers to get to know its arch in practice before the PS5 would be released. It would also make sense that they'd produce the ps5 parts before and delaying the PC parts afterwards as well.

Navi's design, and Nvidia's 7nm as well, should be well completed already. Testing & Validation take a while, and they're waiting for full production on the 7nm nodes. 

 

Though TSMC still seems to be on track for around July for full production on their 7nm node, so we'll see. We could see 7nm on PS5 before any other 7nm product to market.

[Taf the Ghost]

2 minutes ago, cj09beira said:

agreed with the first part, but by then there is a good chance that mining wont last that long, profitability is so low right now

There's always a new coin.

[Taf the Ghost]

http://www.freepatentsonline.com/20180102338.pdf

 

@cj09beira

 

Speaking of Chiplet.

[vdc]

On 18/04/2018 at 11:44 AM, SC2Mitch said:

wccftrash so take it with a dumpster of salt 

aw yeah me likey this 

[laminutederire]

4 minutes ago, Taf the Ghost said:

Navi's design, and Nvidia's 7nm as well, should be well completed already. Testing & Validation take a while, and they're waiting for full production on the 7nm nodes. 

 

Though TSMC still seems to be on track for around July for full production on their 7nm node, so we'll see. We could see 7nm on PS5 before any other 7nm product to market.

Nvidia might need less validation time. Ok if they go to 12nm then 7nm, that'd be the forth node for the same arch, so validation becomes more of an habit than something else where you could really have problems with rad8cak changes.

[cj09beira]

12 minutes ago, Taf the Ghost said:

http://www.freepatentsonline.com/20180102338.pdf

 

@cj09beira

 

Speaking of Chiplet.

thats GREAT news, and i dont need to tell you why  the question is when will we see the first product with it

[SeriousDad69]

Neat, pretty much the same it's been the last couple of generations(80 non ti as good or slightly better than last gen 80ti). I don't personally care about mid-range gpus though, I cant wait to see specs of the 1180ti or 1150ti. Low end and high end are where the best bang for the buck is, at least when it comes to nvidia gpus.

[pyrojoe34]

On 4/18/2018 at 10:55 AM, AluminiumTech said:

We could see Tensor Cores in Nvidia's next gen. We'll just have to wait and see.

I don’t see why they would. As far as I can tell there’s no real use for them in gaming (at least not yet). Mostly just machine learning applications and this way they can force the professional market to commercial GPUs over gaming GPUs. That would be good news for gamers (less non-gamer market competition) but bad news for prosumers and budget academic labs.

 

At some point I could see mixed precision used in gaming for physics and real-time AI but the development of those into games would likely be a few generations away anyway and be very limited due to lack of experienced developers and inherent programming challenges with implementation of such complex code.

[AlTech]

Just now, pyrojoe34 said:

I don’t see why they would. As far as I can tell there’s no real use for them in gaming (at least not yet).

We don't know because game developers haven't tried to use them in gaming yet.

[Mira Yurizaki]

Using tensor cores for machine learning in games at runtime doesn't make sense to me. It makes more sense during development to fine tune things.

 

But I don't want AI (the enemy anyway) to get progressively better as the game ages or the game needs time to mature or something and I have to run it.

[Tieox]

Shall we take bets on if Linus drops one?

[Unixsystem]

5 hours ago, M.Yurizaki said:

Using tensor cores for machine learning in games at runtime doesn't make sense to me. It makes more sense during development to fine tune things.

 

But I don't want AI (the enemy anyway) to get progressively better as the game ages or the game needs time to mature or something and I have to run it.

I believe it was GN that said that most of the new real time raytracing Nvidia has been showing off utilizes tensors, which is why they always advertise as being a Volta-centric feature. If raytracing is going to be the next big "thing", I could see them pushing out consumer cards with at least some form of tensor on board, just to get an install base up so devs can start working with them. I'm not sure on a technical level if they can add less/nerfed tensors compared to what are on the compute cards, but it would make some amount of sense if they were able to.

[Mira Yurizaki]

23 minutes ago, Waffles13 said:

I believe it was GN that said that most of the new real time raytracing Nvidia has been showing off utilizes tensors, which is why they always advertise as being a Volta-centric feature. If raytracing is going to be the next big "thing", I could see them pushing out consumer cards with at least some form of tensor on board, just to get an install base up so devs can start working with them. I'm not sure on a technical level if they can add less/nerfed tensors compared to what are on the compute cards, but it would make some amount of sense if they were able to.

Which would make more sense given Microsoft's push to get ray tracing into Direct3D.