Support NeoGAF

[Die Namek Ability]

BUT WE DO KNOW THESE THINGS! WE HAVE CONCRETE HARDWARE DETAILS OF THE SX AND PS5

and we also know that the SS has to be the cheapest system, otherwise it would be a pointless version

BUT WE DO KNOW THESE THINGS! WE HAVE CONCRETE HARDWARE DETAILS OF THE SX AND PS5

and we also know that the SS has to be the cheapest system, otherwise it would be a pointless version

 

[Bo_Hazem]

I think it tough to grasp, that something for video can be used for audio, that's why people are....ummm.....concerned. But GPGPU has been here for a long time.

Here, a very short video from AMD:

 

[M1chl]

Here, a very short video from AMD:

I can visualized it quite easily, just saying that people have problem to disconnect pixels and sound waves : ) But for others video is nice : )

 

[01011001]

sorry but stupidity upsets me.

so do you really think we do not know these things or are you trolling? I hope you're trolling

 

[Bo_Hazem]

I can visualized it quite easily, just saying that people have problem to disconnect pixels and sound waves : ) But for others video is nice : )

What AMD done previously was reserving 4 CU's to produce 32 sound sources fully raytraced. With the tempest it's capable of more than 5,000 sources of current PS4 VR and hundreds of heavily calculated, raytraced sound sources.

So it's a massive jump, as well as Dolby Atmos is good for only 32. It's a great push from Sony to drive the industry further.

 

[Die Namek Ability]

sorry but stupidity upsets me.

so do you really think we do not know these things or are you trolling? I hope you're trolling

Please dont get upset with yourself. Let me help you out.

My response was to the original post that
XSX=most powerful
XSS=least expensive

I said it was a stupid argument because you could also say

XSX=most expensive
XSS=weakest

None of the above has been confirmed as fact and yet you choose to believe the first set over the ones i have posted disregarding the FACT that i have reached my conclusion using the same STUPID logic the original poster used. But somehow you have found a way to twist this as "my brain doesn't work", or im stupid. Biased people will believe what suits thier agenda, case in point. Enjoy your Xbox dude!

 

[M1chl]

What AMD done previously was reserving 4 CU's to produce 32 sound sources fully raytraced. With the tempest it's capable of more than 5,000 sources of current PS4 VR and hundreds of heavily calculated, raytraced sound sources.

So it's a massive jump, as well as Dolby Atmos is good for only 32 as well. It's a great push from Sony to drive the industry further.

Well with dolby atmos is arbitrary limit, because "intense labor", will see how those technologies mature. However Atmos have really awesome feeling of space and it apply compression (like a audio amplitude compression, not bitrate), reverb, delay in an amazing wave. That it feels way bigger than having sound without it. This sound enhancement is also important I would say.

Great to see it being further improved.

I admit, I did not know that AMD has this kind of system in place. However with AMD, lot of technologies does not went into a real life use case.

 

[Tomeru]

I hope people wont call the series s the Xbox SS. Thats kinda nazi...

 

[M1chl]

I hope people wont call the series s the Xbox SS. Thats kinda nazi...

Not everything is StormTroopers, pardon Schutzstaffel.

 

[Kagey K]

Please dont get upset with yourself. Let me help you out.

My response was to the original post that
XSX=most powerful
XSS=least expensive

I said it was a stupid argument because you could also say

XSX=most expensive
XSS=weakest

None of the above has been confirmed as fact and yet you choose to believe the first set over the ones i have posted disregarding the FACT that i have reached my conclusion using the same STUPID logic the original poster used. But somehow you have found a way to twist this as "my brain doesn't work", or im stupid. Biased people will believe what suits thier agenda, case in point. Enjoy your Xbox dude!

Series X = Most powerful is indeed a FACT.

Series S (if it exists) being the weakest is also likely a fact. (Though not confirmed.) Everything is is conjecture and speculation at this point.

 

[Tomeru]

Not everything is StormTroopers, pardon Schutzstaffel.

 

[Gaz]

Those Digital Foundry comparisons gonna get real hot when they are out. My gosh do DF need to upgrade their servers to prepare for the onslaught of traffic.

The writing is on the wall on this one but I'm still going to enjoy the ride based on how much noise the internet is making about this:

 

[M1chl]

It's also SS for Super Sport.

 

[geordiemp]

Series X = Most powerful is indeed a FACT.

Most powerful paper terraflops, but will it be more performant in games ?

Efficiency is just as important and ps5 GPU is more efficient

 

[Die Namek Ability]

Series X = Most powerful is indeed a FACT.

Series S (if it exists) being the weakest is also likely a fact. (Though not confirmed.) Everything is is conjecture and speculation at this point.

Because it has more CUs? Or more TFlops?

You could extrapolate from what has been provided by Sony or MS this is the case but it is far from fact which was my point!

 

[Gaz]

Most powerful paper terraflops, but will it be more performant in games ?

Efficiency is just as important and ps5 GPU is more efficient

Man how on earth is PS5 gonna outperform the XSX? Please dont say the SSD

 

[Kagey K]

Most powerful paper terraflops, but will it be more performant in games ?

Efficiency is just as important and ps5 GPU is more efficient

After this Hot Chips info, most powerful.

Trying to downplay it does nothing, in fact it comes across as embarrassing.

 

[Kagey K]

Because it has more CUs? Or more TFlops?

You could extrapolate from what has been provided by Sony or MS this is the case but it is far from fact which was my point!

It's a fact. End of. No amount of denial will change it.

End of.

 

[geordiemp]

Man how on earth is PS5 gonna outperform the XSX? Please dont say the SSD

After this Hot Chips info, most powerful.

Trying to downplay it does nothing, in fact it comes across as embarrassing.

Glad you asked, 15 % behind in shader performance....what do to catch up a bit (I never said out perform did I)?

What if ps5 had a compression in the GPU that made the vertex shaders transfer to the pixel shaders faster and more efficient ?

Well thats the naughty dog and Cerny patent, Geometry engine unique to ps5. What does it solve, well see below

Add in direct apis vs abstract Dx12 and it will be interesting, but it is no white wash.

But keep with your power crown, it will be fun to see the melt downs (again). games shown dont lie.

I still think XSX will be slighly edge, but nothing in the realms you guys are boasting about.

I would wait to see results before shouting about power. Notice MS have not been crowing about power so much lately and toned it down............Just wait and see.

 

[Lysandros]

We always need to be careful with MS statements. SPFP means Single precision Float Point, so fp16.

Exactly.. And that makes the xsx audio dsp around 70 gflops full (fp32) block. Including the decode harware.

 

[TBiddy]

We always need to be careful with MS statements. SPFP means Single precision Float Point, so fp16.

Isn't Single Precision Floating Point FP32?

Single-precision floating-point format - Wikipedia

en.wikipedia.org

Edit: For clarity, SPFP is absolutely FP32 and not FP16. FP16 is half precision.

 

[MasterCornholio]

Man how on earth is PS5 gonna outperform the XSX? Please dont say the SSD

What's wrong with the I/O being a strength of the PS5? It's going to give the console some advantages just like the additional CUs of the XSXs GPU will give it some as well.

 

[Marlenus]

Efficiency is just as important and ps5 GPU is more efficient

Citation needed.

The 5500XT and 5700XT have the same perf/watt and run similar clocks. The 5700 has lower clocks and more CUs than the 5500XT and also has higher perf/watt.

I doubt this will change which means a higher clocked lower CU GPU will have worse perf/watt than a lower clocked higher CU GPU.

 

[Die Namek Ability]

It's a fact. End of. No amount of denial will change it.

End of.

 

[VideodromeX]

Is this a topic about XSX or PS5?

 

[TBiddy]

Is this a topic about XSX or PS5?

War never changes. Whatever it takes to promote the console of your choice, you konw.

 

[01011001]

Please dont get upset with yourself. Let me help you out.

My response was to the original post that
XSX=most powerful
XSS=least expensive

I said it was a stupid argument because you could also say

XSX=most expensive
XSS=weakest

None of the above has been confirmed as fact and yet you choose to believe the first set over the ones i have posted disregarding the FACT that i have reached my conclusion using the same STUPID logic the original poster used. But somehow you have found a way to twist this as "my brain doesn't work", or im stupid. Biased people will believe what suits thier agenda, case in point. Enjoy your Xbox dude!

again, how can you in any way claim that we do not know if the SX is the most powerful console, and how can you claim the SS os the not the cheapest?

I in fact believe that people who state this are indeed a bit dense. because we know for a fact the hardware setup of the SX and PS5,
and we know for a fact that the SS HAS TO BE the cheapest system going by hardware and the whole purpose of the system.

so I say it again, how can you justify saying we do not know these things?
because you can't

 

[geordiemp]

Citation needed.

The 5500XT and 5700XT have the same perf/watt and run similar clocks. The 5700 has lower clocks and more CUs than the 5500XT and also has higher perf/watt.

I doubt this will change which means a higher clocked lower CU GPU will have worse perf/watt than a lower clocked higher CU GPU.

Quite correct.

So Cerny said due to RDNA2 limitations and logic propagation in GPU, limit of performance gains was 2.23 Ghz so they capped it there.

What do you think the frequency limit of gains for the 5700 is and prpagation delay / other resitrictions of RDNA1 are ? There is your answer.

If you took either of those 5500 / 5700 and clocked them 20 % lower so all frequencies are not limited by something else what would happen ? I think you know dont you ?

 

[Apollo Helios]

Not really, the light (UV or other wavelengths) is used to shine a pattern on a material covering teh wafer which can selectively remove it at next stage, so it leaves a pattern of protection on the surface so the next stage only adds that process in the uncovered areas.

EUV is much shorter wavelength and hence is more accurate and precise at nanometres, but the machines that do this cost more than god.

Yes, but layers is all done with light, no? You don't take a layer of circutry and in the end you physically take it and stack them together to create X-layers of silicon, after it exist the machine, is already finished product, if I am not mistaken, because I am not sure how would you connect those layers with physically stacking.

I am not really that articulate and it was probably stupid premise for question from the get go. You set some depth into material, do the layer, then set another depth into material and do other layer and so on, right?

OK watch this


It does really cost A LOT to R&D and make those machines!

 

[JLMC469]

Is this a topic about XSX or PS5?

Fanboys.

 

[Die Namek Ability]

again, how can you in any way claim that we do not know if the SX is the most powerful console, and how can you claim the SS os the not the cheapest?

I in fact believe that people who state this are indeed a bit dense. because we know for a fact the hardware setup of the SX and PS5,
and we know for a fact that the SS HAS TO BE the cheapest system going by hardware and the whole purpose of the system.

so I say it again, how can you justify saying we do not know these things?
because you can't

Yeah im done here

 

[michaelius]

Sounds like Raytracing is pretty marginal in RDNA2 hardware. Looks like Nvidia is going to demolish AMD again in desktop space.

 

[thicc_girls_are_teh_best]

I think it tough to grasp, that something for video can be used for audio, that's why people are....ummm.....concerned. But GPGPU has been here for a long time.

I think you mean something for audio can be used for video ?

Yes it is unusual and Tempest has a lot of restrictions in that regard, but creative devs can still find a way to squeeze out a bit of that performance. FWIW they can probably do something vaguely similar for Series X's audio if they wanted to get creative enough, the the specific degree of types of non-audio graphics-type tasks would likely vary since they aren't repurposing a CU core the way Sony is.

I base this on the fact a game actually did do this, Shining Force III on the Saturn, over two decades ago. They used the Saturn's Yamaha audio DSP for some graphics and logic work when able. I should try finding more information on that but it was really cool to find out devs went to such lengths to work within the design of such a system.

The better question is will any dev utilize Tempest (and to somewhat lesser extent, Series X audio core) for non-audio tasks? I don't think we'll see anything like that until well into the generation, and we have to keep in mind we're only talking about a "few" hundred GFLOPs at most with either. It seems like a lot in isolation but it really isn't when compared to the capabilities of their GPUs. Still though, anything that can be squeezed out for extra performance, even if you have to get a bit creative with it, is ultimately good to have.

We always need to be careful with MS statements. SPFP means Single precision Float Point, so fp16.

Or we can take them at their word and I believe most people knew it was referring to FP16. But that only really means it would likely scale WRT FP32 of One X CPU cluster, if anything. FWIW Sony didn't specify FP16 or FP32 when comparing Tempest to PS4 CPU.

 

[geordiemp]

OK watch this


It does really cost A LOT to R&D and make those machines!

They get lots of particles from me after I have had a few puffs

Automation is hypnotic...

ASML stuff is nuts, 3 cargo planes for 1 machine and people think the chips will be cheap

 

[Psyrgery ES]

Truth is nobody knows enough, unless Sony gets more transparent.

The only metric comparable is by their jagcore cpu examples.

336gbs is crazy fast just for audio, there is a reason MS partition it this way...
Besides, sorry i doubt PS5 developers will want to eat too much into that 448gbs for audio....

I swear, some people see the 9gbs/22gbs numbers of the PS5 ssd and be Holy Fuck ITS FAST and then watch the 336gbs of ram for purposes other than gpus and be man thats slow AF.

Smfh.

"Furthermore If we go down the rabbit hole of CU vs MHZ, recall that increasing the CU count to increase TFLOP is not a linear increase either. Compare the 2080 vs the 2080ti, 50% more CUs for 17% extra performance. For AMD, just compare the R9 390X vs the Fury X, 41% increased CU count for 23% increased performance. This is partially related to a concept in computing called Amdahl's law. You can read more about it here. Basically the higher the parallelization of a workload (such as thousands of GPU CUs) the harder it is to extract perfect performance from it.

The 18% number that is being thrown around for the difference in performance in the XSX and PS5 is purely theoretical and is based on the raw TFLOP numbers. In reality, the difference might even be smaller then that. The PS5 GPU does have advantages in Pixel fill rate and triangle culling as well."

T H I S

A 5700 with faster frequencies is not more powerful than a 5700 XT with the same number of Teraflops, because RDNA1 Ocs like shit. If RDNA2 follows the same trend the power difference might even be higher than 18%

 

[SlimySnake]

Still the Tempest Engine CU discussion. Concern trolls come and go, get schooled, and then come back some time later

The tempest engine is not one of the 40CUs, it's a separate hardware unit BASED on a CU. A child can grasp this.

i would love to know where this tempest engine actually resides. Would be crazy if its own on the APU. a Dual CU on the 5700xt is around 5-8mm2, on the enhanced node its probably 3-4 mm2 meaning this 1 CU tempest engine might actually be just a 1-2 mm2 piece on the die offering insane amounts of audio processing power.

 

[Bitmap Frogs]

Those Digital Foundry comparisons gonna get real hot when they are out. My gosh do DF need to upgrade their servers to prepare for the onslaught of traffic.

They'll get real boring real fast like current gen comparisons got.

Xbox games will have a framerate or resolution advantadge. There, that's the entirety of DF comparisons for the next 7 years.

 

[geordiemp]

I think you mean something for audio can be used for video ?

Yes it is unusual and Tempest has a lot of restrictions in that regard, but creative devs can still find a way to squeeze out a bit of that performance. FWIW they can probably do something vaguely similar for Series X's audio if they wanted to get creative enough, the the specific degree of types of non-audio graphics-type tasks would likely vary since they aren't repurposing a CU core the way Sony is.

I base this on the fact a game actually did do this, Shining Force III on the Saturn, over two decades ago. They used the Saturn's Yamaha audio DSP for some graphics and logic work when able. I should try finding more information on that but it was really cool to find out devs went to such lengths to work within the design of such a system.

The better question is will any dev utilize Tempest (and to somewhat lesser extent, Series X audio core) for non-audio tasks? I don't think we'll see anything like that until well into the generation, and we have to keep in mind we're only talking about a "few" hundred GFLOPs at most with either. It seems like a lot in isolation but it really isn't when compared to the capabilities of their GPUs. Still though, anything that can be squeezed out for extra performance, even if you have to get a bit creative with it, is ultimately good to have.



Or we can take them at their word and I believe most people knew it was referring to FP16. But that only really means it would likely scale WRT FP32 of One X CPU cluster, if anything. FWIW Sony didn't specify FP16 or FP32 when comparing Tempest to PS4 CPU.

Time stamped, yes games can use tempest if computationaly expensive workloads are better suited to SPU like logic and lots of FFT.

 

[01011001]

Yeah im done here

because you have no arguments to back up you ridiculous claim

WE KNOW THESE THINGS

 

[Physiognomonics]

Isn't Single Precision Floating Point FP32?

Single-precision floating-point format - Wikipedia

en.wikipedia.org

Edit: For clarity, SPFP is absolutely FP32 and not FP16. FP16 is half precision.

Yes, you seem to be right about this.

 

[Marlenus]

Quite correct.

So Cerny said due to RDNA2 limitations and logic propagation in GPU, limit of performance gains was 2.23 Ghz so they capped it there.

What do you think the frequency limit of gains for the 5700 is and prpagation delay / other resitrictions of RDNA1 are ? There is your answer.

If you took either of those 5500 / 5700 and clocked them 20 % lower so all frequencies are not limited by something else what would happen ? I think you know dont you ?

I think Cerny is selling a story on why the PS5 GPU is designed how it is.

When you say efficiency are you talking perf/watt or perf/flop? Generally I default to perf/watt but I want to make sure we are discussing the same thing before we have a fanboi fight.

 

[Jon Neu]

Those Digital Foundry comparisons gonna get real hot when they are out. My gosh do DF need to upgrade their servers to prepare for the onslaught of traffic.

The writing is on the wall on this one but I'm still going to enjoy the ride based on how much noise the internet is making about this:

If the Series X multiplats run better, there's going to be a giganourmous shitshow around here.

We are not ready.

 

[TBiddy]

Exactly.. And that makes the xsx audio dsp around 70 gflops full (fp32) block. Including the decode harware.

Actually, it puts it at about 140 GFlops, since SPFP is fp32, not fp16, as evidenced above.

 

[Kagey K]

Yep that's basically how you are acting.

I'm done wasting my time. Thanks for the laugh though.

 

[geordiemp]

I think Cerny is selling a story on why the PS5 GPU is designed how it is.

When you say efficiency are you talking perf/watt or perf/flop? Generally I default to perf/watt but I want to make sure we are discussing the same thing before we have a fanboi fight.

2 different topics, amd already claimed the perf /. watt improvement and from hot chips we already seen a IPC improvement in RDNA2 over GCN.

All we know from RDNA2 is xsx is clocked at 1.825 ghz (no explanation of clock chosen) and Ps5 is clocked at 2.23 Ghz (Cerny said capped at 2,.23 ghz as that was limit of propagation logic for ps5), and we have seen 2 Sony cooling patents already so take a wild guess.

So without any other valid information on RDNA2 until pc parts appear, I am going with Cerny information for now for max clocks.

Now performance per TF, that is a different topic all together and aboutr efficiencies in rendering and shading. What do you think of the recent Cerny / naughty dog shading patent and compressing GPU vertices ?

If you want to call Cerny a liar fair enough, most of the time he has the last laugh.

 

[Marlenus]

2 different topics, amd already claimed the perf /. watt improvement and from hot chips we already seen a IPC improvement in RDNA2 over GCN.

All we know from RDNA2 is xsx is clocked at 1.825 ghz (no explanation of clock chosen) and Ps5 is clocked at 2.23 Ghz (Cerny said capped at 2,.23 ghz as that was limit of propagation logic for ps5), and we have seen 2 Sony cooling patents already so take a wild guess.

So without any other valid information on RDNA2 until pc parts appear, I am going with Cerny information for now for max clocks.

Now performance per TF, that is a different topic all together and aboutr efficiencies in rendering and shading. What do you think of the recent Cerny / naughty dog shading patent and compressing GPU vertices ?

If you want to call Cerny a liar fair enough, most of the time he has the last laugh.

RDNA can already exceed 2.23Ghz in the 5700XT with the right cooling and voltages.

What I do know is that for the 5700XT to see gains from that clockspeed it needs more memory bandwidth. With the RDNA2 colour compression enhancements I would not be surprised if 2.23GHz is the point where the memory bandwidth gets saturated so clocking higher uses more power and creates more heat for practically no performance uplift.

 

[Gaz]

If the Series X multiplats run better, there's going to be a giganourmous shitshow around here.

We are not ready.

If anyone is expecting a different result...I don't know what to say. But....the lotion has been ordered

 

[Die Namek Ability]

Yep that's basically how you are acting.

I'm done wasting my time. Thanks for the laugh though.

The irony

 

[geordiemp]

RDNA can already exceed 2.23Ghz in the 5700XT with the right cooling and voltages.

What I do know is that for the 5700XT to see gains from that clockspeed it needs more memory bandwidth. With the RDNA2 colour compression enhancements I would not be surprised if 2.23GHz is the point where the memory bandwidth gets saturated so clocking higher uses more power and creates more heat for practically no performance uplift.

Cerny did say specfically

With this new paradigm we're able to run way over that in fact we have to cap the GPU frequency at 2.23 GHz so that we can guarantee that the on chip logic operates properly.

So I am interpreting propagation logic delay but it could be multiple factors....as it is quite a vague statement.

But agree we dont have much to go on. It wont be long before PC chips on RDNA2 come out and we will know a shit ton more.

Also that Cerny ND patent removing the bottlenecks in GPU and compressing the vertices probably helps get up to the 2.23 as well. It does mention all the bottlenecks solved in the patent.

 

[Panajev2001a]

We always need to be careful with MS statements. SPFP means Single precision Float Point, so fp16.

No, that would be half precision then if it were FP16.

DP FP: FP64
SP FP: FP32
HP FP: FP16