Baseless Next Generation Rumors with no Technical Merits [post E3 2019, pre GDC 2020] [XBSX, PS5]

[DavidGraham]

On the PC space, FP16 was used in about 6 games, notably among them is Wolfenstein 2, Wolfenstein YoungBlood, Resident Evil 2, The Division 2, Far Cry 5 and maybe Devil May Cry 5, so it's logical that PS4pro has benefited from such implementation, although limited in number and scope.

 

[Deleted member 13524]

Edit: XBX has no fp16, so if you can use it for most, PS 8 TFfp16 could be faster than XBX 6 TFfp32. But i don't know if this ever happened in practice.

If the XBoneX uses a Polaris GPU (which it probably does) then it does "have FP16". Polaris has instructions for FP16 by using only the needed cache for FP16 instead of having to perform a full "promotion" of every variable to FP32, i.e. it takes less bandwidth to do FP16 operations on XBoneX so it's still better to use FP16 when you can. It also takes less power for the operation, though that's less relevant to a console.

What Polaris (and by extension XBoneX) doesn't have is rapid packed math, or the ability to do FP16 at twice the rate of FP32.

Never happened cause 4Pro is missing the needed bandwidth and RPM FP16 only helping in limited situations over the big picture.

Source for FP16 performance boosts never happening? And because of bandwidth nonetheless?

Last I checked FP16 is extensively used for upscaling and reconstruction, and we saw a large number of developers claiming that 2xFP16 is very useful for a substantial percentage of pixel shaders (IIRC 20-30%).

As for the bandwidth, it doesn't really make much sense to claim FP16 throughput is bottlenecked by bandwidth, since its operations are less bandwidth sensitive than FP32 ones.

Could you be mistaking this with the 64 ROPs on the Pro never being fully utilized because there's not enough bandwidth to feed them? That was indeed a claim made by a Sony dev IIRC, but it had nothing to do with 2xFP16.

 

[Shifty Geezer]

I'm curious how widespread fp16 is at all, and for what it's used.

My expectation is very little. It's a niche in a niche, only good for PSPro titles which won't sell more on account of it being written in versus not. The economic reasons to rework parts of the shaders to use RPM just aren't there where the existing PS4 base shaders can be used. Some devs may have dabbled or done so out of wanting to achieve a better target, but it makes no sense for devs to push this tech. Come next-gen when it's a standard feature, games will be better optimised around this standard feature.

 

[ultragpu]

no one knows.
you don’t know everything about the chip. You don’t know everything about the system.
Have all the flops you want but without a massive amount of bandwidth to feed it; it’s all going to be bandwidth bound anyway. You can rely on other technologies to make up a a difference but we saw how that turned out for 4Pro and X1X.
big TF means nothing without big bandwidth. And big bandwidth and big TF means big price.

Hope HBM2 isn't off the table yet for those mouth watering bandwidth, the rumor did mention Sony got a sweet deal from some manufacturer awhile back. But yeah who knows.

Reading what Jim said about fastest transition to next gen PS ever, I would think BC is of extreme importance to Sony. Inability to play your PS4/Pro library on PS5 would be massive fuck up. Imagine MS giving you game pass and hundreds of last gen game on it, and PS5 doesn't even give you ability to play your disk or digital games from last gen on it? Its probably much more important for them then ~5% higher TF number.

If I could speculate, I think this is big reason why Sony started earlier to work on Navi with AMD. They absolutely have to nail BC for last gen, otherwise pool of 100+ million current costumers might just buy next Xbox.

Why can't they just keep their PS4s?

 

[N2O]

Use HBM2 doesn't really help much for the console, and increase the cost by a lot.
It's not DDR3 vs GDDR5 anymore.

 

[ultragpu]

Klee is still as confident as ever, the latest scoop from the dude when replying to his status.

Oh don't worry. I'm still around.

Just lurking in the background allowing people to say things they will regret later.

For now.

 

[AbsoluteBeginner]

Klee is still as confident as ever, the latest scoop from the dude when replying to his status.

I mean if he has the specs he can easily say "This is not PS5/XSX". Its not gonna get anyone in trouble, but he will confirm black on white this is not it.

 

[iroboto]

I think we all have general rules whenever someone boasts how smart/sure of something they are.

What was the saying? The uninformed always boasts about how sure they are of something and the wise are always full of doubt.

 

[ultragpu]

I mean if he has the specs he can easily say "This is not PS5/XSX". Its not gonna get anyone in trouble, but he will confirm black on white this is not it.

He already said it maybe? PS5 is still double digit RDNA TFs, slightly stronger than Xbox even after the apparent 12 TF announcement and he even said XsX is a little above 12 TF. No point repeat the same thing over and over.

 

[cheapchips]

I do wonder if MS will come out as cost competitive with Sony. The chip is the only place where we know they differ significantly on BOM, other than minor thing like DS5 triggers. If that gap is narrow enough then maybe they'll absorb the chip costs. That few billion lost to hardware margins is more palatable than the market share roasting they had this gen.

 

[N2O]

Is Klee a developer, or an insider that just heard things from developers?

 

[iroboto]

Is Klee a developer, or an insider that just heard things from developers?

He is the former editor for a gaming magazine IIRC

 

[AbsoluteBeginner]

He already said it maybe? PS5 is still double digit RDNA TFs, slightly stronger than Xbox even after the apparent 12 TF announcement and he even said XsX is a little above 12 TF. No point repeat the same thing over and over.

My problem with Klee is he was vague. He even said, back in August, people will be disappointed with specs once they find them out. Now, who would be disappointed with 12-13TF Navi GPU inside console? That is far cry from what was expected back on August when he said it.

I already said what I think, he is posturing with vague leaks, thats all. He even posted MS dev kits are later so this might be the differential, its all a bit up in the air.

Same with rest of the leaks. Reiner got info of PS5 being stronger then Scarlett the minute it was announced at E3. Nothing much, just "got info from developer", but that is very convinent time to say such thing as 90% of big sites reported PS5 > Scarlett after reveal even with Sony not being there. Free publicity and at a same time bad one for your competitor. Win - Win.

Him saying its even slightly above 12TF means nothing when he said it after announcement and TomWarren leaking specs a day before. Its easy to tell it after the fact, really. Even if it ends up 11.9TF he can easily say "they had to downclock it" "It was dev kit" etc.

Same thing with Hqmgg, who said it matches leaked Dante specs (64CU * 1475MHz was posted) and Klee saying "bingo". Except Arden doesnt look anything like 64*1475MHz and it never made any sense (as Navi scales differently from Vega).

All in all, alot of vague unconfirmed "leaks". Same song and dance as the one previous gen, so lets just wait for official confirmation. Oberon being "size smaller" then Arden, as per very credible Taiwanese leaker, tells much more then whatever these celebrity leakers leaked and collaborates 1:1 with leaked AMD repo.

 

[Esrever]

So? Are we at 15TF yet?

When do we get to the secret 2nd GPU?

 

[RDGoodla]

That leaker Aquariuszi, who said OBR is 300mm2 and Arden is 350mm2, also attributed the difference of area to something like RT ASIC (of course it may be other kinds of processor), not the GPU size.

This is likely a sign the latest PS5 has similar CUs to xbox. Imo OBR will have 4~8 less CUs, using 7nm+ and being able to clock at 2GHz. This spec is equivalent to 12.3~13.3 TFflops which is in line with what insiders previously said.

The most important thing is: it is small and fast, smart and beautiful design as described by some of our beyond3d members.

 

[N2O]

That leaker Aquariuszi, who said OBR is 300mm2 and Arden is 350mm2, also attributed the difference of area to something like RT ASIC (of course it may be other kinds of processor), not the GPU size.

This is likely a sign the latest PS5 has similar CUs to xbox. Imo OBR will have 4~8 less CUs, using 7nm+ and being able to clock at 2GHz. This spec is equivalent to 12.3~13.3 TFflops which is in line with what insiders previously said.

The most important thing is: it is small and fast, smart and beautiful design as described by some of our beyond3d members.

Don't read too much for his predictions part.
And I am sure if Sony can 7nm+, Microsoft can too.

 

[AbsoluteBeginner]

That leaker Aquariuszi, who said OBR is 300mm2 and Arden is 350mm2, also attributed the difference of area to something like RT ASIC (of course it may be other kinds of processor), not the GPU size.

This is likely a sign the latest PS5 has similar CUs to xbox. Imo OBR will have 4~8 less CUs, using 7nm+ and being able to clock at 2GHz. This spec is equivalent to 12.3~13.3 TFflops which is in line with what insiders previously said.

The most important thing is: it is small and fast, smart and beautiful design as described by some of our beyond3d members.

Someone close to fab can have info on chip size, verification process and general state of fabrication, but whats inside chip is solely his prediction. Like his prediction that Sparkman layout indicates semi custom chip ~50% size of Arden but cannot be next gen because its not powerful enough.

We know that to be wrong, but chip that is size ~200mm² and has 4TF with 128-192bit bus can certainly exist for next gen. And it most likely does, but it doesnt make his info untrue, only his prediction.

BTW 7nm EUV does not provide that kind of jump. Not over N7P, which they will likely use (as you do not have to change anything in chip design to get ~10% improvement in TDP at same clocks - its just 2nd gen 7nm node).

That is obviously assuming Sony uses 7nm EUV, while MS decides to use worse node (never happened between the two).

 

[RDGoodla]

Don't read too much for his predictions part.
And I am sure if Sony can 7nm+, Microsoft can too.

Because he already sees the layouts of Arden/OBR APUs. And he found some other part in Arden making it 50mm2 larger.

The key point is GPU size is not the main reason of the 50mm2 difference.



Xbox will have 60~64 CUs including redundancy and I always think it will go 7nm+ to have 350mm2.

Someone close to fab can have info on chip size, verification process and general state of fabrication, but whats inside chip is solely his prediction. Like his prediction that Sparkman layout indicates semi custom chip ~50% size of Arden but cannot be next gen because its not powerful enough.

We know that to be wrong, but chip that is size ~200mm² and has 4TF with 128-192bit bus can certainly exist for next gen. And it most likely does, but it doesnt make his info untrue, only his prediction.

Imo you have seen a misleading translation. He said "it's not practical to expect it (sparksman) to have the performance of next-gen consoles".

He is right spakesman can't be on par with those next-gen consoles already announced.

 

[Deleted member 13524]

My problem with Klee is he was vague. He even said, back in August, people will be disappointed with specs once they find them out.

The average Joe who only looks at high-level specs can only be disappointed with reallistic next-gen consoles.
On the GPU front we went from a 0.19 TFLOPs PS3 to a 1.84 TFLOPs PS4, ~10x more powerful. From 512MB RAM to 8GB, 16x more RAM. CPU was considered a let-down last gen.
This time we're going from a 1.8 TFLOPs GPU to a very obviously <18 TFLOPs GPU, as it looks like it's a lot closer to 5x more powerful. We're apparently going from 8GB RAM to 16GB, so 2x more RAM unless there's an unmentioned additional large pool of memory in there. CPU is going from a 1.6GHz 8-core to a 3.2GHz 8-core.


If you are someone who ignores improvements in ALU utilization ratios, power efficiency, effective bandwidth, etc. and only look at raw throughput, then the perfect GPU for 2020 consoles would have been a 8000sp / 400 TMUs / 160 ROPs, TeraScale GPU at 1.8GHz resulting in a whopping 28 TFLOPs. Which is about how many execution units on the TeraScale architecture you could cram inside a 10 Billion transistor GPU if you consider that a RV770 with 800 sp / 40 TMUs / 16 ROPs was 950 million transistors.


So yeah, someone will always be disappointed.

I mean just look at this very forum. There are people here who will be disappointed if 60 FPS aren't mandatory for all 9th-gen titles. Those guys are always going to be sorely disappointed, probably for the rest of their lives.

 

[AbsoluteBeginner]

Because he already sees the layouts of Arden/OBR APUs. And he found some other part in Arden making it 50mm2 larger.

The key point is GPU size is not the main reason of the 50mm2 difference.



Xbox will have 60~64 CUs including redundancy and I always think it will go 7nm+ to have 350mm2.




Imo you have seen a misleading translation. He said "it's not practical to expect it (sparksman) to have the performance of next-gen consoles".

He is right spakesman can't be on par with those next-gen consoles already announced.

I dunno, 50mm² on same process (without ESRAM) is pretty substantional. There is no chance Oberon is gonna match Arden with 50mm² deficit.

From what can be deducted from AMD repo, seems like RT will be inside GPU compute units, not outside, so perhaps ~50mm² he is seeing is just what leaked doc shows - additional 10WGPs