Support NeoGAF

[Shin]

The console makers will use whatever manufacturer the chip designer uses.

Not necessarily if any of this is true:

SIE chose AMD 's small power - saving CPU core Jaguar architecture for PS 4 series.
However, this is different from the original plan, and in the initial plan was to become the "Steamroller (steam roller)" core of Bulldozer (bulldozer) system. And APU of Bulldozer core PS 4 was scheduled to be manufactured at 28 nm of GLOBALFOUNDRIES, not TSMC's 28 nm.

However, at the last minute GLOBALFOUNDRIES '28nm rise was doubtful, SIE and AMD were urged to change the plan in a hurry.
As a result, we moved the foundry to TSMC, due to time and engineering constraints, the CPU core was changed to a Jaguar core that is synthesizable already on the TSMC process.

TSMC is leading the pack that much is known, nVidia doesn't make APU's like AMD so assuming they stick with an APU then it will be AMD.
Plus the latter is cheaper than a nVidia GPU's, it wouldn't make sense either to use a GPU from your competitor in your own APU (albeit semi-custom).

 

[truth411]

Nothing is really 'true 7nm'. If by that you mean Intel 7nm, then you'd be looking at something like 5nm for the other foundries.

I dunno TSMC EUV 7nm+ does look like Intel 7nm. TSMC Regular 7nm is more like 10nm.

 

[E-Cat]

I dunno TSMC EUV 7nm+ does look like Intel 7nm. TSMC Regular 7nm is more like 10nm.

We don't even know the dimensions of Intel 7nm yet. But if the cadence stays the same, Intel 7nm will be comparable to 5nm, just like Intel 10nm is comparable to 7nm.

7nm+ is only like a 15-20% area reduction.

 

[blu]

Not to mention Nvidia openly said recently the console business means nothing to them.

Their words don't always match their actions.

 

[Carn82]

This isn't true at all. More money might be made in those segments in the future but right now the gaming segment is bigger than all of the other segments combined. The YOY growth in revenue for Tegra last quarter was actually bigger than the entire automotive segment. Where do you even get the idea that more money is being made in those segments?

I was referring to the console space, not PC gaming. If the Q1 results are an indication, datacenter revenue might leap across gaming revenue this year already.

 

[DieH@rd]

Imagine if they’d have skipped the mid gens and gone from 28 to 7nm.

IMO, midgen does not affect the power differnce between launch 8th gen consoles and 9th gen.

 

[E-Cat]

IMO, midgen does not affect the power difference between launch 8th gen consoles and 9th gen.

It literally makes no difference. Not sure what point he was trying to make.

 

[Taggen86]

Does anyone have an idea of what kind of Tflop numbers we could expect with next gen 7nm console GPUs and nvidia cards?

Looking at historic numbers. it seems that Tflops scale almost perfectly with transistor density

28nm: GTX 960 (2,3 TF) 16nm: GTX 1060 (3,8 TF)
28nm: GTX 980 (4,6 TF) 16nm: GTX 1080 GTX 1080 (8,2 TF)
(65 to 75 percent leap both in terms of transistor density and tflops)

If this is true next gen with 7nm technology should be around 120-130 percent more powerful than current gen (less for AMD card which already uses 14nm technology).

7nm GTX 1160: 8,5 TF
7nm GTX 1180: 18,4 TF

So if next gen consoles use mid range GPUs (a la GTX 1160/RX 680) we could expect 8,5 TF graphic cards in them. Combine that with a much better CPU a la RYZEN enabling 60 fps gameplay and we have a substantial leap (at least in relation to the original PS4 and Xbox one, 4,5-6 times as powerful in terms of GPU, I guess more than 10 times as powerful in terms of CPU). The 399-499 USD price range is clearly achievable I think. However, XB1X owners who do not care about 60 fps wont have much incentives to upgrade since they only get a slight 40 procent increase in GPU power (at the 399 price point)...

 

[KOHIPEET]

7nm non-EUV? Interesting..

EUV should be ready by 2020 shouldn't it?

 

[E-Cat]

Does anyone have an idea of what kind of TFlop numbers we could expect with 7nm console GPUs and high end nvidia cards?

Looking at historic numbers. it seems that TFlops scale almost perfectly with transistor density

28nm: GTX 960 (2,3 TF) 16nm: GTX 1060 (3,8 TF)
28nm: GTX 980 (4,6 TF) 16nm: GTX 1080 GTX 1080 (8,2 TF)

If this is true next gen should be around 120-130 percent more powerful than current gen (less for AMD card which already uses 14nm technology).

7nm GTX 1160: 8,5 TF
7nm GTX 1180: 18,4 TF

So if next gen consoles use mid range GPUs (a la GTX 1160/RX 680) we could expect 8,5 TF graphic cards in them. Combine that with a much better CPU a la ryzen and we have a substantial leap (at least in relation to the original PS4 and Xbox one)

A couple things:

1. Historically, it's not worth doing a new generation until you can scale both computing and memory by at least 8x. That would set the floor at 14-15 TFLOPS for PS5.

2. However, due to the manufacturing of advanced nodes getting slower and more expensive, that bar will likely be lowered. Computing power now doubles every 2-2.5 years, so a PS5 launching in 2020 should easily at least double the power of Scorpio (= 12 TFLOPS). Another way to look at possible future console specs is to take the high-end AMD card out in the market at that point in time and divide its power by some constant. At the launch of PS4 it was 3x, though Scorpio is getting closer to 2x. By 2020 we'll have GPUs in the ballpark of 30 TFLOPS, so you do the math.

EUV should be ready by 2020 shouldn't it?

Sooner than that. Samsung's 7nm will employ EUV and go into risk production in 2H-18. Similarly, TSMC 7nm+ and GF 7LP+ will have it circa 2019.

 

[mrklaw]

They would have still moved to 16nm for their slim refreshes

Mint if they hadn't increased power they could have pretty heavily reduced size and cost

I wonder if they'll naturally refresh PS4 slim with 16nm anyway for that reason (or are they already?)

 

[Taggen86]

If you only use a mid range card, you would probably need 5nm transistors to get a console twice as powerful as Scorpio (12 TF). I very much doubt that we have 5nm cards in 2020 given that the 7nm cards are expected to arrive in 2019 (3 years after 16-14nm) and it is getting more and more difficult to shrink transistors. Furthermore, consoles usually use GPU transistor technology that is 0.5-1.5 years old (based on PS4 pro and PS4). But we will see. I hope you are correct =)

A couple things:

1. Historically, it's not worth doing a new generation until you can scale both the computing power and memory by at least 8x. That would set the floor for PS5 at 14-15 TFLOPS.

2. However, due to the manufacturing of advanced nodes getting slower and more expensive, that bar will likely have to be lowered. Computing power doubles every 2-2.5 years now, so a PS5 launching in 2020 should easily at least double the power of Scorpio. Another way to look at possible future console specs is to take the high-end AMD card out in the market at that point in time and divide its power by some constant. At the launch of PS4 it was 3x, though Scorpio is getting closer to 2x. By 2020 we'll have GPUs in the ballbark of 30 TFLOPS, so you do the math.

Sooner than that. Samsung's 7nm will employ EUV and go into risk production in 2H-18. Similarly, TSMC 7nm+ and GF 7LP+ will have it circa 2019.

 

[G.ZZZ]

You can't realistically call this 7nm with a 40-50% gain on 14 nm srsly. Halving the node should double the efficiency. At least be mathematically consistent on those arbitrary numbers jeez

 

[AmyS]

Why not just launch early on 14nm then do a 7nm refresh in 2024?

Surely that would be cheaper while still providing a substantial jump from the Jaguars in the current gen consoles.

That doesn't make sense.

PS4 Slim, Xbox One S, PS4 Pro and Xbox One X all use 16nm now.
So going to 14nm wouldn't be of much benefit for future consoles since 14nm too similar to 16nm.

The next consoles will use some flavor of 7nm in 2019/2020 and shrink to some flavor of 5nm in ~2023.

 

[Taggen86]

Agree, 5nm mid gen upgrades in ~2023 seems likely. That is also when I expect 399 USD next gen consoles will be twice as powerful as scorpio.

That doesn't make sense.

PS4 Slim, Xbox One S, PS4 Pro and Xbox One X all use 16nm now.
So going to 14nm wouldn't be of much benefit for future consoles since 14nm too similar to 16nm.

The next consoles will use some flavor of 7nm in 2019/2020 and shrink to some flavor of 5nm in ~2023.

 

[E-Cat]

Agree, 5nm mid gen upgrades in ~2023 seems likely. That is also when I expect next gen consoles will be twice as powerful as scorpio.

That is way too conservative. Performance is roughly doubled every node, so you would expect 5nm consoles to be at least four times as powerful as Scorpio -- unless you want to dramatically lower their price point, but then it would make no sense to launch them in the first place.

 

[Taggen86]

srry my misstake. 5nm mid gen consoles in 2023 should be at least three times as powerful as Scorpio (16.8 TF) if we assume performance scales with transistor density in a similar way as it did 28nm vs 16/14nm, and if we assume a similar 399-499 USD price point.

That is way too conservative. Performance is roughly doubled every node, so you would expect 5nm consoles to be at least four times as powerful as Scorpio -- unless you want to dramatically lower their price point, but then it would make no sense to launch them in the first place.

 

[b0bbyJ03]

nVidia is fairly doubtful. They burned quite a few bridges in the past and it seems nobody is letting them forget it.

I wasn't really paying attention at the time. What did Nvidia do to burn bridges?

 

[mrklaw]

If everyone is assuming some flavour of ryzen and sticking with an APU, then I don't think you can just multiply the current GPUs so easily. Ryzen will mean a lot more space used by the CPU vs jaguar so leaving less effective space for the GPU

 

[Apparition]

I suppose we will first see the superslims on 7nm before the next gen ones arrive.

Wonder if next-gen will utilize any MCM-GPU designs.

 

[RaijinFY]

If everyone is assuming some flavour of ryzen and sticking with an APU, then I don't think you can just multiply the current GPUs so easily. Ryzen will mean a lot more space used by the CPU vs jaguar so leaving less effective space for the GPU

No not really... Ryzen (8c/16t) at 7nm should be ~ the same size as the Jaguar was at 28nm.

 

[Renekton]

No not really... Ryzen (8c/16t) at 7nm should be ~ the same size as the Jaguar was at 28nm.

Not sure about that, Jaguar cores are super tiny (3.1mmsq per core)

 

[ethomaz]

Not sure about that, Jaguar cores are super tiny (3.1mmsq per core)

Ryzen core is ~6.6mm2... 7nm will reduce 50-60% compared with 14nm.

Yes it can be even smaller... 2.7 to 3.3 mm2... there is no way to know until AMD put Ryzen in 7nm.

And Ryzen cores are probably bigger than mobile Zen cores... I expected even smaller size for APUs.

 

[Renekton]

Ryzen core is ~6.6mm2... 7nm will reduce 50-60% compared with 14nm.

Yes it can be even smaller... 2.7 to 3.3 mm2... there is no way to know until AMD put Ryzen in 7nm.

And Ryzen cores are probably bigger than mobile Zen cores... I expected even smaller size for APUs.

I thought Ryzen also has that large L3

 

[ethomaz]

I thought Ryzen also has that large L3

Yeap but it is not part of the core... it is outside and linked to 4 cores.

In any case the best case we can have for size of what will be on PS5 are the Zen based APUs to launch in late 2017 or early 2018.

 

[hanspampel]

I thought Ryzen also has that large L3

this is a Ryzen CPU

one CCX is 44mm² in 14nm LPP from GlobalFoundry



Jaguar Core at 28nm is 3mm² (without cache)

edit:
Xbox One SOC (red is the 8 Jaguar cores)

 

[nkarafo]

So what's the physical limits then? I heard this can't ever be as small as 3nm.

 

[ethomaz]

So what's the physical limits then? I heard this can't ever be as small as 3nm.

With actual material and tech yeap.

Of course new things will happen making chips even smaller... there are a lot of alternatives in research already and you won't see 5nm or 3mm before 2023... so there are a lot of time yet for new solutions matured.