AMD details future 7nm CPU and GPU designs

[WYP]

AMD has released details about their future 7nm CPU and GPU designs. ZEN 2 is "Design Complete".



Read more about AMD's future 7nm CPU and GPU designs.

 

[Giggyolly]

I just don't buy those GPU slide, how the hell is AMD going to pull out 3 generation by 2020?!
Even if that's Vega 7nm for Radeon Pro only, followed closely by Navi for consumer, I still think "next gen" before the end of 2020 is a stretch.

 

[NeverBackDown]

You don't have to buy it. It's a free roadmap!

 

[spolsh]

also, the speech marks on "Next gen" extend beyond 2020, and with AMDs history with graph scaling that could mean 2030.

 

[WYP]

I just don't buy those GPU slide, how the hell is AMD going to pull out 3 generation by 2020?!
Even if that's Vega 7nm for Radeon Pro only, followed closely by Navi for consumer, I still think "next gen" before the end of 2020 is a stretch.

AMD almost definitely means up to an including 2020, or that they will have silicon ready before 2020.

 

[AngryGoldfish]

It's quite possible Navi will only encompass one chip and only cater to the midrange and low-end (like Polaris did).

I have no idea what they're planning on doing with Vega in the gaming sector. It's such a shambles that I don't see how they could market it once Nvidia release their next generation and Navi comes out. Like Fury X, it'll be a fish out of water. 7nm Vega 20 for the compute sector might still live, but that's about it.

So, if Navi is only one or two midrange chips, and is basically just a redesigned Polaris, and Vega 7nm is a refresh, it's possible they've also been able to work on this "next-gen" simultaneously. As WYP said a few days ago, "next-gen" could have been in the works for years. I suspect we'll see "next-gen" in late 2020, if we do see it in 2020.

 

[ET3D]

"Improve on 'Zen' in multiple dimensions"

The dimensions in question are obviously width and height. It's a smaller process, after all.

 

[grec]

Navi is supposed to be scalable and have a mature implentation of Infinity Fabric. If the intention is for multi-die MCMs to be used in a similar way to Epyc/TR, then there's no reason why a single ~36 CU(Polaris-sized) chip can't cover the whole RX670-RX695 range of performance(A 4xMCM should comfortably outperform Vega64 CFX while appearing to the system as a single GPU).

 

[AngryGoldfish]

Navi is supposed to be scalable and have a mature implentation of Infinity Fabric. If the intention is for multi-die MCMs to be used in a similar way to Epyc/TR, then there's no reason why a single ~36 CU(Polaris-sized) chip can't cover the whole RX670-RX695 range of performance(A 4xMCM should comfortably outperform Vega64 CFX while appearing to the system as a single GPU).

Originally that's what I saw rumoured as well, but it was just a rumour. Now I see rumours that Navi will not use the Infinity Fabric in the way Zen has. If it does it will only see two RX 580-performance chips stacked together to hit GTX 1080/Vega 64 performance level at around the $250-300 price point.

https://wccftech.com/rumor-amd-navi-mainstream-gpu-to-have-gtx-1080-class-performance/

What's likely to me is Navi will be Polaris all over again, with only one midrange die that'll draw more power than the GTX 1160, overclock worse, be hotter, but will still perform better in DX12/Vulkan scenarios or where games are coded specifically with their cards in mind, as was the case with Far Cry 5.

I don't think we'll see AMD compete in the high-end until 2020 at the earliest.

 

[grec]

While there's been some indication the implementation of this tech has been pushed back a generation, it's existance itself is more than just a rumour; Multi-die Vega/Navi cards have long appeared in AMD slides and some tech demos have reportedly been shown to developers behind closed doors.

Raja has previously said: "It enables some really low latency and high-bandwidth interconnects.This is important to tie together our different IPs (and partner IPs) together efficiently and quickly. It forms the basis of all of our future ASIC designs. We haven't mentioned any multi GPU designs on a single ASIC, like Epyc, but the capability is possible with Infinity Fabric."

It's likely the only reason we've not yet at least seen a traditional dual-die variant of Vega so far(AMD has made at least a "prosumer" version of every top-end chip for many generations without fail, including Polaris) is the same reason we've not seen a Vega Nano: The existing SKUs are (For the time being) already in far higher demand than they expected+can keep up with.

I think we'll certainly see a dual-GPU variant that appears to the OS as a single GPU for Navi at some point, even if only in low volumes and with the full implementation coming in later drivers, to iron out the kinks in the technology as they move to become more dependant on it in the future.

Lets not forget this is technology NVidia has also publicly been researching and planning to use for years.

 

[AngryGoldfish]

While there's been some indication the implementation of this tech has been pushed back a generation, it's existance itself is more than just a rumour; Multi-die Vega/Navi cards have long appeared in AMD slides and some tech demos have reportedly been shown to developers behind closed doors.

Raja has previously said: "It enables some really low latency and high-bandwidth interconnects.This is important to tie together our different IPs (and partner IPs) together efficiently and quickly. It forms the basis of all of our future ASIC designs. We haven't mentioned any multi GPU designs on a single ASIC, like Epyc, but the capability is possible with Infinity Fabric."

It's likely the only reason we've not yet at least seen a traditional dual-die variant of Vega so far(AMD has made at least a "prosumer" version of every top-end chip for many generations without fail, including Polaris) is the same reason we've not seen a Vega Nano: The existing SKUs are (For the time being) already in far higher demand than they expected+can keep up with.

I think we'll certainly see a dual-GPU variant that appears to the OS as a single GPU for Navi at some point, even if only in low volumes and with the full implementation coming in later drivers, to iron out the kinks in the technology as they move to become more dependant on it in the future.

Lets not forget this is technology NVidia has also publicly been researching and planning to use for years.

All I've seen with Navi is "scaleability", and that wasn't shown in the most recent slide from AMD (as far as I remember). I haven't seen any confirmed or leaked slides showing anything more than that. Not saying they don't exist; just that I haven't seen them.

Stacking Vega seems highly implausible—and not because of the high demand. I always thought the design idea was to stack small dies on top of each other; not oversized ones. It'll be ridiculously powerful, yeah, but impossible to power and cool (unless they shrink down to 7nm and reduce the clocks a bunch).

 

[grec]

Assuming by stacking you mean side-by-side on an interposer(Realistically TSV-style direct stacking isn't going to happen with multiple GPU dies anytime soon), I don't see any reason for it to not be possible. Current Vega64 card designs & cooling can go upto 350W, while well binned Vega chips with slightly more conservative clock curves have been shown to require as little as 165W(With AMDs original and unreleased Vega64 nano card having a 175W max consumption). With Vega 7nm launching(For enterprise) in the coming months and Navi being mostly a mature refinement both architecturally and node-wise, chances are we'll see full fledged Navi dies varying from ~100-200W power consumptions consumptions, presumably with the better of those bins going to multi-GPU(Whether MCM or otherwise) designs.

 

[AngryGoldfish]

Assuming by stacking you mean side-by-side on an interposer(Realistically TSV-style direct stacking isn't going to happen with multiple GPU dies anytime soon), I don't see any reason for it to not be possible. Current Vega64 card designs & cooling can go upto 350W, while well binned Vega chips with slightly more conservative clock curves have been shown to require as little as 165W(With AMDs original and unreleased Vega64 nano card having a 175W max consumption). With Vega 7nm launching(For enterprise) in the coming months and Navi being mostly a mature refinement both architecturally and node-wise, chances are we'll see full fledged Navi dies varying from ~100-200W power consumptions consumptions, presumably with the better of those bins going to multi-GPU(Whether MCM or otherwise) designs.

So you think we'll see Vega-stacked (I do mean side-by-side, yeah, but stacked is easier to write) GPUs for gaming later this year or just Vega-stacked in general? It would be awesome to see Vega 64 with lower clock speeds 'stacked' and offering performance levels past a 1080Ti, but I don't see it happening, not for gamers. I see Vega's gaming influence being over. I do see its compute performance increasing because it's clear that's where AMD are more focused right now and can actually compete and earn some money. I don't see any reason to believe Navi is going to be anything more than a midrange product as AMD have not suggested any reason for it to be. To release a second 'high-end' product so closely after Vega without replacing Polaris would be very odd for such a small company with such meagre GPU sales. Polaris was not a replacement for Fiji. I don't think Navi will be a replacement for Vega. The only hint that it will be anything more is that Raja—who incidentally has said some highly debatable things—hinted at the Infinity Fabric being an integral part to future GPUs from Radeon.

That said, I think Navi SHOULD replace Vega. If AMD are hiding something big in hopes that Nvidia will release (what I think they will release) a mediocre new architecture later this year and then rain on their parade with Navi that scales from the low-end all the way up to 1080Ti and beyond performance levels at competitive prices, that'll be wonderful. Take GCN out with a bang. Awesome. Vega 20 7nm will remain the compute card and Navi will be two 7nm GTX 1080 performing chips 'stacked' together. Awesome. I just don't think it's going to happen.

 

[grec]

Oh no yeah, Vega 7nm won't release as a gaming product and it's now unlikely we'll see a consumer dual card(An enterprise one could certainly be possible to maximise space efficiency for racks though), and by the time Navi is released it will almost certainly fit into the area of the market polaris targeted. All past rumours and leaked slides have pointed to Navi being a die shrink and architectural refinement of Vega, with the same number of CUs. It would probably fit in the upper-mid end of the market(Where Polaris was when it launched) if it maintained Vega's 64CUs and refined its memory stack and multi-GPU performance.

 

[AngryGoldfish]

Oh no yeah, Vega 7nm won't release as a gaming product and it's now unlikely we'll see a consumer dual card(An enterprise one could certainly be possible to maximise space efficiency for racks though), and by the time Navi is released it will almost certainly fit into the area of the market polaris targeted. All past rumours and leaked slides have pointed to Navi being a die shrink and architectural refinement of Vega, with the same number of CUs. It would probably fit in the upper-mid end of the market(Where Polaris was when it launched) if it maintained Vega's 64CUs and refined its memory stack and multi-GPU performance.

A midrange product (Polaris type positioning) with 64 CUs? Similar to the R9 390X, except not a re-badging but a new architecture based on the previous one. I don't see why they would do that.

 

[grec]

If the key technology originally planned for Navi("scalability improvements"/Improved IF) isn't ready yet, but 7nm is and they need a new line to make use of it, then it makes a lot of sense to just tweak Vega's core for gaming performance on 7nm while keeping the overall design much the same. We know Vega has a lot of untapped potential when it comes to gaming performance(HBCC, DSBR, Primitive shaders, RPM, ect have been too hard to properly(Or at all) implement in gaming applications so far given Vega's relatively immature and compute-orientated implementations of these technologies), and Navi's refinements of these technologies and their APIs could allow them to finally be enabled fully in consumer drivers and deliver solid gains for AMD alongside the clock speed bumps from 7nm and the efficiency gains that would allow it to fit into the midrange and dual-card markets(Given 7nm yields will likely be much lower, heavily disabled variants of Navi could appear in place of a Polaris sized chip for the RX67/80 cards).

Essentially, Navi appearing as a tweaked Vega could allow them to be first on the market with 7nm gaming cards (If NVidia's cards are coming this summer they're almost certainly 12nm, a safe but solid jump from Pascals 16nm), while making reasonable IPC gains in the applications that matter with minimal R&D cost and time.

Navi probably won't launch as the fastest GPU core on the market, but it could put AMD back into their position of having the fastest gaming card on the market with a dual-GPU variant(in the same way the HD7970X2, R9 295X2, and Fury X2 were all technically the fastest GPU cards on the market at the time of their respective launches), just this time while being in the power draw range of a traditional single-GPU enthusiast card and hopefully while moving away from the use of AFR for mGPU gaming in modern titles.