Support NeoGAF

[transformer]

in the video the hairy dude stated "from a pure processing perspective this is 4x more powerful than the xbox1x". there's no room for interpretation.




if they meant real world gaming performance, they should have said real world gaming performance

So we are going to ignore what the head of Xbox game studios said about the 4x number? And how is it incongruent with the processing power statement? Games require processing power. The CPU processes, GPU processes, RT hardware will process, could be some other custom work in the APU to assist in processing. The processing power of the console, inclusive of ALL HW in the console, is 4x higher than Xbox one x.

 

[DemonCleaner]

oh boy

 

[Insane Metal]

Since NAVI is more efficient and 10TF of NAVI beats 12TF of VEGA, I'm changing my prediction to 10TF max for both PS5 and XBII.

 

[ResilientBanana]

So we are going to ignore what the head of Xbox game studios said about the 4x number? And how is it incongruent with the processing power statement? Games require processing power. The CPU processes, GPU processes, RT hardware will process, could be some other custom work in the APU to assist in processing. The processing power of the console, inclusive of ALL HW in the console, is 4x higher than Xbox one x.

That's pretty awesome, I expect both consoles to be 4x the Xbox One x. But we also have to take into consideration that we can't translate to TFLOPs the same way it used to with AMD. For instance the Xbox One X 6TF on GCN is technically around 4.75 TFLOPS on NAVI which makes Lockhart, if it exists, a fairly capable machine. Unless I'm interpreting Lisa Su wrong, that's what I got from the AMD conference.

One thing I noticed is that the GPU's all have 8GB GDDR6 so I can honestly believe we won't see the next-gen consoles surpassing 16GB GDDR6. More bandwidth allows you to use less memory to get the same results.

 

[CrustyBritches]

ya too little power

180-190W is too much power? I agree, that might be a little too high for a console.

 

[Munki]

PS5 GPU Will Use Brand New RDNA Architecture Developed by AMD

Information from the recent AMD Computex event indicates that Sony's PlayStation 5 console will be using brand new RDNA architecture for its AMD GPU.

screenrant.com

ok?

 

[DeepEnigma]

ok?

It is all marketing and what MS and Sony allow them to discuss at this time.

 

[ethomaz]

more CU = more expensive. consoles last gen had more cus but they were disabled to increase manufacturing yield. you wont see 64 CU consoles. 60 at most.

Neither console will have more than 64 CU

my guesstimate? the PS5 and xbox three will have around 56-60 CU. at 1.8 ghz. these would be around 13.8 TFLOPS. which is around what people have rumored

they will manufacture the consoles at lower than 64 CUs because they will 100% want higher yields.

if they come with the full 64 I'd be shocked lol.

tho if am Sony. I don't know why they shouldn't bring in more power with a graphics accelerator. or a cell processor side-chip that helps ps3 emulation.

1800Mhz with the power draw of RDNA/Navi? I will be surprised.

 

[ethomaz]

Since NAVI is more efficient and 10TF of NAVI beats 12TF of VEGA, I'm changing my prediction to 10TF max for both PS5 and XBII.

That is really true? Navi 10TFs beats Vega 12TFs?

 

[TeamGhobad]

i dont understand how one cards tflops is more effective than anothers. what metric will we use to differentiate the cards now?

 

[ethomaz]

i dont understand how one cards tflops is more effective than anothers. what metric will we use to differentiate the cards now?

Flops are the same.
How the card use them is more or less efficient than another.

 

[TeamGhobad]

Flops are the same.
How the card use them is more or less efficient than another.

so if a card uses bruteforce to get 12tflops than its still better than a card that uses a more efficient way to get 10tflops?

 

[ethomaz]

so if a card uses bruteforce to get 12tflops than its still better than a card that uses a more efficient way to get 10tflops?

That situation can happen well nVidia cards do more for games with 9TFs than AMD cards with 12-13TFs.
But all depends how the hardware is optimized to use these units and how the software is coded to make easy the hardware use them.

Now expect to use 100% of the SPs in a card reaching the max TFs output is utopian.
It is not possible to use 100% of the SPs units all the time... sometimes the game didn't even require these units and there are bootlenecks all the time... for example the data didn't reach in time from RAM so that chunk of SPs will stay iddle until the data reach the registers... so it is already holding the GPU to reach the max TFs performance.
There is a lot of others situations.

The TFLOPS is the metric to tell you the max processing power a GPU can have... it is like the max speed your can shows on display... can you reach the speed displayed in the screen? At least in my cars most shows 220 or 250 km/h but even in optimal situations (in a race track with all the best conditions) I never could reach more than 205 or 210 km/h.

It is the same here... the peak performance is TFLOPS is basically impossible to be archived so the GPU that can use more of that TFLOPS for games (that means efficiency) will be a better architecture for games.

 

[ResilientBanana]

That is really true? Navi 10TFs beats Vega 12TFs?

According to Lisa Su's slide - 10 NAVI TFLOP x 1.25 = 12.5 VEGA TFLOP

 

[Negotiator]

i dont understand how one cards tflops is more effective than anothers. what metric will we use to differentiate the cards now?

There are 2 metrics: GPU compute shaders (GPGPU) and traditional 3D rasterization.

GPGPU-wise the more flops the merrier. GCN can reach 99-100% efficiency in those workloads.

But when it comes to traditional 3D rasterization, there are more factors at play (like memory compression aka DCC, tiled rendering -which Navi doesn't seem to have-, hardware tessellator performance, DX11/OpenGL driver tricks etc.)

 

[IceManCat]

So a 12 TF next gen system is equivalent in power to a 15 TF current generation system ?

 

[ResilientBanana]

So a 12 TF next gen system is equivalent in power to a 15 TF current generation system ?

I'd expect 10TF next gen systems.

 

[LordOfChaos]

So a 12 TF next gen system is equivalent in power to a 15 TF current generation system ?

In theory, given AMDs 1.25x IPC increase is in full effect. Has yet to be tested how well it translates.

 

[IceManCat]

In theory, given AMDs 1.25x IPC increase is in full effect. Has yet to be tested how well it translates.

I wonder how they would test it, seems like the only real possibility would be to run a last gen game on a new console. But even then it wouldn't be a good metric

 

[LordOfChaos]

I wonder how they would test it, seems like the only real possibility would be to run a last gen game on a new console. But even then it wouldn't be a good metric

They could test RDNA's IPC uplift entirely on PC, other console special sauce is harder but the 1.25x IPC is for everything using RDNA

 

[CrustyBritches]

8tf Panello redeemed!

As a gif lover I do have a soft spot for that "It's 8TF, Austin..." gif. Mr. McMahon is perfect for a dream-dashing suit.

 

[SonGoku]

Keep in mind that 9TF 5700XT beats 12.6Tflop Vega 64.

Old 12TF wishes for PS5 were based on Vega's perf/tflop figures.

Yeah but it doesn't make sense to use such small chip on PS5, the strength of the 7nm node is in the density increase not clock/performance.
I think we'll see 12TF RDNA2 minimum.

The CPU is based on the third generation of AMD’s Ryzen line and contains eight cores of the company’s new 7nm Zen 2 microarchitecture.

I think that's just the writer adding his own spin

7nm+ has a very small area lead over 6nm, but power isn't yet stated. I don't know why more EUV layers would lead to worse.

Where does it say 6nm has more EUV layers than 7nm EUV.

 

[LordOfChaos]

Where does it say 6nm has more EUV layers than 7nm EUV.

Wrong link, whoops

TSMC: Most 7nm Clients Will Transition to 6nm

www.anandtech.com

As previously reported, TSMC’s N6 process technologies adopts extreme ultraviolet lithography (EUVL) to lower manufacturing complexity by reducing the number of exposures required for multi-patterning (which is needed today as TSMC’s N7 uses solely DUV lithography). While TSMC's N7+ uses up to four EUVL layers, its N6 expands it up to five layers, whereas N5 expands usage of EUVL all the way to 14 layers.

 

[SonGoku]

Wrong link, whoops

TSMC: Most 7nm Clients Will Transition to 6nm

www.anandtech.com

As previously reported, TSMC’s N6 process technologies adopts extreme ultraviolet lithography (EUVL) to lower manufacturing complexity by reducing the number of exposures required for multi-patterning (which is needed today as TSMC’s N7 uses solely DUV lithography). While TSMC's N7+ uses up to four EUVL layers, its N6 expands it up to five layers, whereas N5 expands usage of EUVL all the way to 14 layers.

Curious indeed

 

[ethomaz]

There are some interesting estimate for die-size of Anaconda happening in ERA right now.

380-400mm2.

Mother of God if true.

 

[Negotiator]

There are some interesting estimate for die-size of Anaconda happening in ERA right now.

380-400mm2.

Mother of God if true.

SonGoku's 72 CUs (64 active) dream redeemed!

 

[ethomaz]

I will use DukeBlueBall estimates here:

First based in the RX 5700 chip.

5700:
GDDR6 phy controller: 4.5mm x 8
Dual CU: 3.37mm x 20
4 ROP cluster: .55mm x 16
L1+L2+ACE+Gemotry processor+empty buffer spaces + etc: 139mm

After that he used the GDDR6 size in the Scarlett video and compared with the APU to reach these numbers:

370mm2-390mm2

And finally he tried to estimate how many DCUs can fit the Anaconda with that size:

75mm for CPU
45mm for 10 GDDR6 controllers
8.8mm for ROPs
140mm for buses, caches, ACE, geometry processors, shape etc. I might be over estimating this part as the 5700 seems to have lots of "empty" areas.

We have ~110mm left for CUs + RT hardware. There is enough there for ~30 dual CUs and RT extensions.

I found interesting.

30DCUs = 60CUs.

SonGoku's 72 CUs (64 active) dream redeemed!

He is estimating total 60CUs (including disabled) with that size.

 

[SlashBringingHasher]

Im working with 8 disabled for each tier
64 total -56 enabled
72 total - 64 enabled
80 total - 72 enabled

Theoretically you could fit up to 80CUs on a 350mm2 APU (launch PS4 size) on 7nm
On 7nm EUV no problem at all

I expect 72CUs max 64 enabled leaving space for audio processor, rt hw

i dont think you cn make a blueprint to manufacture more than 64 because thats how the architecture is tho.

 

[llien]

This escalated quickly.
We went from "probably under Vega 64" to definitely beats it.

The way I understood Microsoft, though, was that they'd use infinity fabric to connect to RT thing.
If so, it's a separate chip.

ethomaz
Which pic is used as the base for the estimates?

 

[Lort]

These machines take years to be designed, are you saying that MS will simply wait for Sony to reveal their specs then they simply build better machine? What if Sony never revealed their Specs? Microsoft will cancel next gen? C'mon man.

Your other point about Lockhart selling more than PS5 because its cheaper , That might happen in the US only, WW forget about it. Its all about brand recognition and perceived value and in both metrics PlayStation is way stronger than Xbox. Hell Xbox One has been selling for $199 and less since 2015 and it never sold as well as PS4. Lockhart will be the same.

So your saying sony could sell more $500 boxes than MS could sell $300 boxes? Brave call but thats yours to make. The xbox one has always been about the same price as the PS4 .. being slightly cheaper dosent make much difference.

Microsoft like sony would have multiple prototypes with ability to change yeild and power envelope... otherwise they wouldnt be both playing i wont show you mine if you dont show me yours.

 

[ethomaz]

ethomaz
Which pic is used as the base for the estimates?

I believe the part that shows the memory and the APU in the same screen.

"A rougher estimate using the 12x14mm GDDR6 chips next to the SOC."

 

[Ellery]

Keep in mind that 9TF 5700XT beats 12.6Tflop Vega 64.

Old 12TF wishes for PS5 were based on Vega's perf/tflop figures.

Those are just excuses. The way to calculate TFLOPs doesn't change and the hardware required stays the same.

It doesn't really have anything to do with the architecture. It is mostly a somewhat meaningless number, which was said in this thread countless times before, and I am certain nobody in this thread even remotely is knowledgeable enough to be able to know the difference from a developer standpoint for it, what it would actually mean and how it translates to what we see on screen.

There are plenty of people arguing over it. Here and on many other forums. Most of them have no idea about it at all and just throw numbers around and say "oh 10TF would be super disappointing".

Moving the goalpost here doesn't help and is childish. All those fake leaks and rumors always talked about TFLOPs. Even those going all the way up to stupidly 16.8TF and nobody meant GCN 1.0 TFLOPS by that which would translate to XX Navi RDNA TFLOPs.

Also looks like I was wrong aswell. I thought around 10TFLOPs with the chance of pushing 10.8 TF to beat Stadia, but I guess looking at the 5700 cards it seems that between 8-9.2 is much more likely than over 10 unless those chips shown are not comparable to what we get with the PS5.

 

[Aceofspades]

So your saying sony could sell more $500 boxes than MS could sell $300 boxes?

Yes . Easily

 

[joe_zazen]

Since NAVI is more efficient and 10TF of NAVI beats 12TF of VEGA, I'm changing my prediction to 10TF max for both PS5 and XBII.

If $399, i am going with 8-9tf and 1TB ssd
If $499, 10-11 TF and 2TB ssd

Either, we wont know TF until after release because marketing wont touch that. Less than a 2x gpu jump wont sound very good.

 

[LordOfChaos]

This escalated quickly.
We went from "probably under Vega 64" to definitely beats it.

The way I understood Microsoft, though, was that they'd use infinity fabric to connect to RT thing.
If so, it's a separate chip.

Where was it that indicated this? That would be an interesting way to go. IF is fast but not monolithic chip fast, so RT over IF just wasn't something I was expecting.

(afaik, each chip to chip link has 42GB/s bandwidth)

 

[Lort]

Yes . Easily

And so why did they loose xbox 360vs ps3? ...

 

[KingT731]

And so why did they loose xbox 360vs ps3? ...

Because they didn't...The PS3 outsold the 360 every single year they were both on the market...

 

[SonGoku]

There are some interesting estimate for die-size of Anaconda happening in ERA right now.

380-400mm2.

Mother of God if true.

Relevant speculation:
A theory i had is consoles might go with a big APU die on 7nm and shrink to 6nm (refined 7nm) as soon as is available.

N6 'leverages new capabilities in extreme ultraviolet lithography (EUVL)' gained from N7+. N6 uses the same design rules as N7 and enables developers of chips to re-use the same design ecosystem (e.g., tools, etc.), which will enable them to lower development costs. Essentially, N6 allows to shrink die sizes of designs developed using N7 design rules by around 15% while using the familiar IP for additional cost savings.

Launch die size at 7nm	"6nm" die size (15% reduction)
400 mm2​	340 mm2​
390 mm2​	331.5 mm2​
380 mm2​	323 mm2​
370 mm2​	314.5 mm2​
360 mm2​	306 mm2​
350 mm2​	297.5 mm2​

TSMC will start risk production of chips using its N6 fabrication technology in the first quarter of 2020. Keeping in mind that it usually takes companies about a year to start high-volume manufacturing (HVM) after the beginning of risk production, expect N6 to be used for mass products starting from 2021.

A quick way to bring console costs down with little investment.

 

[SonGoku]

i dont think you cn make a blueprint to manufacture more than 64 because thats how the architecture is tho.

That was GCN limit, RDNA2 (big Navi) is expected to break that limit

He is estimating total 60CUs (including disabled) with that size.

He is lowballin CUs using 5700 empty spaces and adds up to 378.8 mm2
There's also the very real possibility of PS5/XB2 going with 7nm EUV

Using his stimates and maintaining the empty spaces on 7nm
80CU APU = 412.6 mm2
72CU APU = 399.1 mm2
64CU APU = 385.6 mm2

As I've said i dont see next gen consoles GPU being lower than 56CUs enabled
Bolded imo is the perf/buck sweetspot

For reference
The RTX 2060 is 445mm2 $349 with Nvidia's high profits
Launch PS4 roughly 350 mm2 at $399

400mm2 doesn't seem so outlandish to think considering it can be quickly shrunk to 340mm2 when 6nm is available.

 

[Aceofspades]

And so why did they loose xbox 360vs ps3? ...

PS3 outsold x360 every year since launch. It took longer because it launched in the US a year after 360, and in EU a year and half, and also suffered production shortages due to RSX yeilds reducing launch units from 2m to just 500k for all markets.

 

[Lort]

Because they didn't...The PS3 outsold the 360 every single year they were both on the market...

Naw still bitter about it taking sooo long to catch up i see. So even about the same price the 360 can sell as many as a ps3 ... so i totally proved my point and destroyed yours .. unless you want to claim that the 360 would have sold less if it was $200 cheaper.

 

[SlashBringingHasher]

That was GCN limit, RDNA2 (big Navi) is expected to break that limit

He is lowballin CUs using 5700 empty spaces and adds up to 378.8 mm2
There's also the very real possibility of PS5/XB2 going with 7nm EUV

RDNA is still GCN. its still just better

the 64 CU limit still exists.

 

[SonGoku]

RDNA is still GCN. its still just better

the 64 CU limit still exists.

Im talking about RDNA2 "next gen" (big Navi)

 

[demigod]

Did anyone else catch this? Stolen from other place about Lockhart.

Wollan said:
For others like myself catching up to latest news. Richard Leadbetter in Digital Foundry's Microsoft E3 / Xbox Scarlett initial impressions video had this comment:

"I've spoken to developers about this across the year once this sort of concept [dual-SKU approach w/Lockhart] became known and there has always been to everyone I've spoken to about it, a certain degree of suspicion and worry. If we look back at the console generations the lowest spec machine has always been a pain in the bum to work with. ... [Based on Xbox's E3 conference] Is it [Lockhart] gone? That is the question because it certainly seems like it."

With Microsoft describing it as a single console at this years E3 makes me think the Lockhart has indeed been scrapped. And now Microsoft is no longer promising the strongest console outright which makes me believe the next-gen Xbox might now be a singular device with a price ceiling at $499. They might actually have been betting that Sony was releasing in 2019.

 

[TeamGhobad]

There are some interesting estimate for die-size of Anaconda happening in ERA right now.

380-400mm2.

Mother of God if true.

thats a lot of heat. double sided vapour chamber?

With Microsoft describing it as a single console at this years E3 makes me think the Lockhart has indeed been scrapped. And now Microsoft is no longer promising the strongest console outright which makes me believe the next-gen Xbox might now be a singular device with a price ceiling at $499. They might actually have been betting that Sony was releasing in 2019.

my belief is that Lockhart has morphed into a budget Anaconda. No disc drive, smaller storage, comes with xbox one controller.

 

[SlashBringingHasher]

Im talking about RDNA2 "next gen" (big Navi)

i really doubt they are getting rid of the CU limit. and it would just be too expensive to have that many CUs on a console.

 

[SonGoku]

thats a lot of heat. double sided vapour chamber?

RTX 2060 is 445mm2 do you consider that a hot card?

i really doubt they are getting rid of the CU limit. and it would just be too expensive to have that many CUs on a console.

Not really, see my post #4,288

Using his stimates and maintaining the empty spaces on 7nm
80CU APU = 412.6 mm2
72CU APU = 399.1 mm2
64CU APU = 385.6 mm2

As I've said i dont see next gen consoles GPU being lower than 56CUs enabled
Bolded imo is the perf/buck sweetspot

For reference
The RTX 2060 is 445mm2 $349 with Nvidia's high profits
Launch PS4 roughly 350 mm2 at $399

400mm2 doesn't seem so outlandish to think considering it can be quickly shrunk to 340mm2 when 6nm is available.

 

[LordOfChaos]

PS4's die was 348mm² at launch, so not unbelievable. By 2020 the node would also be as mature as 28 was going into 2013.

XBO was actually bigger at 363, though I think eSRAM would be easier to bin for defects than logic.

 

[SlashBringingHasher]

RTX 2060 is 445mm2 do you consider that a hot card?

Not really, see my post #4,288

chip size is meaningless when its exponentially more expensive to make chips at a small process node. (7nm).

also. 16GB+ of super fast RAM is gonna be super expensive.

 

[SonGoku]

PS4's die was 348mm² at launch, so not unbelievable. By 2020 the node would also be as mature as 28 was going into 2013.

Not to mention the quick 6nm shrink for cost reductions
400mm2 -> 340mm2

chip size is meaningless when its exponentially more expensive to make chips at a small process node. (7nm).

15% is not exponentially more expensive