Support NeoGAF

[Seider]

33 Tflops FP16

16.7 Tflops FP32

Maybe they said 33 Tflops because in RDNA 3.5, FP16 is useful in games.

 

[TNT Sheep]

I like this, but it's not accurate. Although more powerful, Ultra Super Saiyajin was slower than SSJ.

The more accurate comparison would be when Piccolo fused with Nail.

Nah man, both comparisons are wrong. I would go with Second Form Frieza for base PS5 and Third Form Frieza for PS5 Pro. An incremental step towards the actual big transformation.

 

[Panajev2001a]

33 Tflops FP16

16.7 Tflops FP32

Maybe they said 33 Tflops because in RDNA 3.5, FP16 is useful in games.

That as well, but there are double the execution units in RDNA3 and when you can dual issue (limited scenarios in RDNA3, slightly expanded on RDNA3.5 IIRC) you can get to double that peak… when you can dual issue you can get then another bump by using FP16 and even more with INT8/4.

 

[Panajev2001a]

Correct. Anyone who said this would have dual-issue compute is completely clueless.

There is no reason to add lots of specific dedicated logic for RT and use dual-issue at the same time.

The leaks were from dev documentation. So pretty much confirmed by Sony.

Dual Issue is part of RDNA3.x and PS5 Pro GPU has a lot of new architectural changes… new ROP, new RT units, WMMA instructions (AI acceleration), etc… why do you draw the line at VOPD instructions? .

 

[RobRSG]

I see Xbots are still talking about Teraflops on Twitter...

They are still clueless, aren't they?

Let's see how the MS "mid-gen" from 2020 compares to the Sony mid-gen from 2024...

It will be hilarious

The Playstation roaches are the ones inflating those Teraflops from what I can see. Green rats are somewhat quiet, because Xbox is pretty much dead at this point.

 

[Elios83]

The Playstation roaches are the ones inflating those Teraflops from what I can see. Green rats are somewhat quiet, because Xbox is pretty much dead at this point.

What's to inflate?
The numbers are in the official leaked technical documentation for developers which was obviously right about everything?
It's 16.7tf and 33.4tf with dual issue with half precision.


It's also not relevant to talk about teraflops since the key features of the Pro are the PSSR and the new advanced ray tracing architecture. The rest is there just to support the new features.

 

[Panajev2001a]

It's 16.7tf and 33.4tf with dual issue with half precision.

The claim about dual issue theoretical peak is separated from FP16 usage. PS4 Pro, PS5, Xbox Series X|S, etc… and now PS5 Pro can use FP16 or lower precision to boost throughput.

Finding code that benefits from both dual issue and FP16 will be rare, but that would run at 4x the base FP32 rate for that section of the code.

 

[nowhat]

Gravity Rush 2 : 2160p ?

This was a curious case on PS4 Pro, it used "geometry upscale rendering" that was advertised (or at least mentioned) before the release of the Pro, but saw little use. Basically, the geometry runs at native 2160p, while pixel shaders run at 1080p. The resulting image is a mixture of sharp and blurry, but in case of Gravity Rush, it worked well enough.

Edit: of course this has already been discussed, should read the thread to the end before replying, but ehh, I'll still leave this here.

 

[MisterXDTV]

It's absolutely mind-blowing people are complaining about the dual issue teraflop number that AMD itself uses for their graphics cards since RDNA 3....

Check the 7800 XT teraflop number....

 

[jasondefaoite]

Besides the fact that it already rendered a game at 8K, it's part of the HDMI 2.1 spec.

Sony has to enable it for output but why would they?

Take the L. You made a mistake. It happens. No point moving the goalposts now.

 

[SplunkyMunkey]

I wonder if 8k is achievable through pssr when it's matured definitely not powerful for a complete 8k image but maybe a set resolution upscaled to 8k without sacrificing image quality. I'm more intrigued how capable pssr can be

 

[HeisenbergFX4]

Oh man, when is Nintendo announcing Switch 2? Give me your best guess

I get zero info on Nintendo sadly

 

[octos]

The claim about dual issue theoretical peak is separated from FP16 usage. PS4 Pro, PS5, Xbox Series X|S, etc… and now PS5 Pro can use FP16 or lower precision to boost throughput.

Finding code that benefits from both dual issue and FP16 will be rare, but that would run at 4x the base FP32 rate for that section of the code.

It doesn't really matter if the code is rare or not, because as always, <10% of the code is responsible for all the slowdown.
Good programmers will find, using profiling, the sections of code that need to be boosted, and such a feature will no doubt prove useful.

 

[shamoomoo]

33 Tflops FP16

16.7 Tflops FP32

Maybe they said 33 Tflops because in RDNA 3.5, FP16 is useful in games.

No. RDNA3 can dual issue some instructions/operations so the FP32 FLOPS is double.

 

[Panajev2001a]

It doesn't really matter if the code is rare or not, because as always, <10% of the code is responsible for all the slowdown.
Good programmers will find, using profiling, the sections of code that need to be boosted, and such a feature will no doubt prove useful.

Sure, might be also limited by bandwidth at some point, but it is not a waste of transistors if not Cerny would demand changes… see how they got AMD to redesign the Ryzen 2 FPU: https://chipsandcheese.com/p/the-nerfed-fpu-in-ps5s-zen-2-cores

 

[Lysandros]

There seem to be a lot of confusions, PS5 PRO is 33.4 TF with dual issue at full fp32 precision, not half precision. So the theoretical max of the machine is officially 33.4 TF. Dual issue feature exists, it is not to be dismissed wishfully. It can ad a significant ~15-20% of compute performance especially in a fixed spec APU.

 

[SenkiDala]

8K output on a base PS5 was never enabled via firmware...

PS5 Pro has 8K output natively

HDMI chip should be different

Of course it was. The Touryst plays at 8K/60fps on base PS5.



Please make researches before saying shit like this. It's alright to be wrong, it happens to me all the time.

 

[nowhat]

Of course it was. The Touryst plays at 8K/60fps on base PS5.



Please make researches before saying shit like this. It's alright to be wrong, it happens to me all the time.

Oh this is so highly ironic. If you'd bother to watch the video - it renders internally at 8K, but that is supersampled to 4K, because, yes, the base PS5 lacks the HDMI bandwidth to actually output in 8K. But as you say, it's alright to be wrong.

 

[Kamina]

Guys, we are almost there and Europe is the land of no disc drives.

 

[Rocco Schiavone]

Guys, we are almost there and Europe is the land of no disc drives.

Well said. Asked Expert in Germany and they told me that Sony isn't providing disc drives till early 2025.

 

[Kamina]

Well said. Asked Expert in Germany and they told me that Sony isn't providing disc drives till early 2025.

if true

 

[JackMcGunns]

I'll let you know once I see how Black Myth Wukong runs on it and how it compares to a budget PC. That game should've been the showcase for PS5 Pro, but it's strange how it's not on the list of enhanced games, but hopefully added soon.

 

[Euler007]

TLDR 32 Gbp caps 8k at 30fps with 4:2:2 chroma subsampling. Also forces you to pick between HDR or 4:4:4 chroma subsampling at 4K/120. I assume the Pro doesn't support DSC either. Kind of an own goal for a console that focuses on IQ as a selling point to not go with 40 Gbps. Never mind if it does support DSC.

 

[Kamina]

Well said. Asked Expert in Germany and they told me that Sony isn't providing disc drives till early 2025.

The funny part is that MediaMarkt keeps saying that it “will be back in stock soon” on the product page…

 

[RobRSG]

Secret sauce keeps giving.

 

[S0ULZB0URNE]

Take the L. You made a mistake. It happens. No point moving the goalposts now.

What was said incorrectly in the post you just quoted?

 

[Shake Your Rump]

So on the whole, this is what an additional $250 gets you:

I wish.

Ignoring the cost of the original PS5 we all paid four years ago, this thing is $500 more in Canada. I paid $720 after tax for the original. The Pro with stand and disc drive (both included with my PS5) is over $1200 after tax here.

 

[Wolzard]

I like this, but it's not accurate. Although more powerful, Ultra Super Saiyajin was slower than SSJ.

The more accurate comparison would be when Piccolo fused with Nail.

PS5 Pro had an increase in strength when they improved the GPU, but the whole is held back by the weak CPU.

I think the analogy is perfect, because Ultra Super Saiyan is kind of an imperfect SSJ2, in the same way that the PS5 Pro is an imperfect PS6.

 

[Mibu no ookami]

PS5 Pro had an increase in strength when they improved the GPU, but the whole is held back by the weak CPU.

I think the analogy is perfect, because Ultra Super Saiyan is kind of an imperfect SSJ2, in the same way that the PS5 Pro is an imperfect PS6.

Show me the CPU constrained games... this makes zero sense.

PS5 will get cross gen games with PS6 and these games are gonna be obviously running on the same CPU...

 

[Wolzard]

Show me the CPU constrained games... this makes zero sense.

PS5 will get cross gen games with PS6 and these games are gonna be obviously running on the same CPU...

Any game with RT, games that use Unreal Engine 5, recent Capcom games like Dragons Dogma 2, etc.
PS4 received multis from PS5, even though it has a tablet CPU. Your argument doesn't make much sense.

 

[gooner4life_uk]

There's no 33.anything tflops, it's an RX 6800 with RDNA 4 RT and ML Bolted on, it's 16.7 teraflops.

 

[Nankatsu]

Probably it was already answered here but I couldn't find it: GPU wise, what's the equivalent on the current PC market?

 

[MisterXDTV]

There's no 33.anything tflops, it's an RX 6800 with RDNA 4 RT and ML Bolted on, it's 16.7 teraflops.

So it's an RDNA 2 GPU Now, from 2020

Just shut up

 

[IllegalLemon]

I see Xbots are still talking about Teraflops on Twitter...

They are still clueless, aren't they?

Let's see how the MS "mid-gen" from 2020 compares to the Sony mid-gen from 2024...

It will be hilarious

Many aren't taking into account the fact that Sony is using FP32 to describe the performance of the Gpu in the pro, not FP16, which is what gpu vendors typically use when AI is involved in the pipeline (PSSR), which frees up gpu resources for other uses. It's why it's rated at 16.7 TFLOPS & not 33.4, not taking into account the design differences between the RDNA 3 compute unit and the RDNA 2 compute unit, which already makes the ps5 pro GPU much different than that in the base console. Perhaps Sony don't use FP16 in their metrics since the pro uses a hybrid rdna 3 and 4 gpu.

Nvidia has been using FP16 numbers to describe the performance of their GPUs since the Turing arch debuted (due to dslr becoming a factor in the pipeline). Amd doesn't do it yet, since they don't have bespoke ML hardware in their GPUs (that might change soon enough though).

 

[gooner4life_uk]

So it's an RDNA 2 GPU Now, from 2020

Just shut up

Yes, it's an RDNA 2.0 RX 6800 with added RDNA 4.0 ML and RT cores added, that's where all the evidence points.

 

[MisterXDTV]

Yes, it's an RDNA 2.0 RX 6800 with added RDNA 4.0 ML and RT cores added, that's where all the evidence points.

Sure they used AMD GPUs from 4 years ago...

 

[FireFly]

Yes, it's an RDNA 2.0 RX 6800 with added RDNA 4.0 ML and RT cores added, that's where all the evidence points.

Presumably the 7800 XT is just the 6800 with VOPD and cache changes. And the 8800 XT will be the 6800 with VOPD and cache changes and extra RT and ML hardware. Or maybe we can go back to RDNA 1 and say the 6800 is just the 5700 XT with 50% more CUs, support for DirectX 12 and Infinity Cache. So that the RDNA 4 cards are really just souped up RDNA 1 parts.

 

[Games of Today]

I'll be testing Red Dead 2 v1.00 to see if there are any improvements.

 

[RaySoft]

Just got confirmation that I'll receive my Pro in about 3 hours

 

[M1chl]

Many aren't taking into account the fact that Sony is using FP32 to describe the performance of the Gpu in the pro, not FP16, which is what gpu vendors typically use when AI is involved in the pipeline (PSSR), which frees up gpu resources for other uses. It's why it's rated at 16.7 TFLOPS & not 33.4, not taking into account the design differences between the RDNA 3 compute unit and the RDNA 2 compute unit, which already makes the ps5 pro GPU much different than that in the base console. Perhaps Sony don't use FP16 in their metrics since the pro uses a hybrid rdna 3 and 4 gpu.

Nvidia has been using FP16 numbers to describe the performance of their GPUs since the Turing arch debuted (due to dslr becoming a factor in the pipeline). Amd doesn't do it yet, since they don't have bespoke ML hardware in their GPUs (that might change soon enough though).

Yeah, I think it make sense, if you could do FP16 for Graphics and FP16 for AI in one clock, if that it is possible, remains to be seen. That way you could achieve something which does not look that good into something which look nice at almost the speed, as there wouldn't be even any AI model running. It would probably also help with latency.

FP16 driven graphics is something which does looks kind of trashy, but if you would have model to mask those downsides, you would gain a shit-ton of performance. For a visualisation, a lot of mobile games operates at half precision, which means a lot of artifacts, but given the super high density screens, its not as noticeable. Or people just don't care.

 

[Physiognomonics]

Probably it was already answered here but I couldn't find it: GPU wise, what's the equivalent on the current PC market?

That GPU isn't released yet.

PS5 Pro is RDNA3.5 + RDNA4 (with RT + ML) features. The latest AMD GPU is still RDNA3.

 

[saintjules]

Maybe I made a mistake and should have had my PS5 ship instead of pickup. Eliminated any chance of getting it early in the latter method.

 

[jasondefaoite]

What was said incorrectly in the post you just quoted?

Really?

The HDMI spec of the PS5 supported 8K hence the one title outputting 8K.

Not one single title outputted 8k. And only a single title rendered in 8K which is what you tried to move the goalposts to.

 

[MisterXDTV]

Really?



Not one single title outputted 8k. And only a single title rendered in 8K which is what you tried to move the goalposts to.

Can't believe in 2024 some people don't know the difference between rendering resolution and output resolution...

Unreal...

 

[Lysandros]

Many aren't taking into account the fact that Sony is using FP32 to describe the performance of the Gpu in the pro, not FP16, which is what gpu vendors typically use when AI is involved in the pipeline (PSSR), which frees up gpu resources for other uses. It's why it's rated at 16.7 TFLOPS & not 33.4, not taking into account the design differences between the RDNA 3 compute unit and the RDNA 2 compute unit, which already makes the ps5 pro GPU much different than that in the base console. Perhaps Sony don't use FP16 in their metrics since the pro uses a hybrid rdna 3 and 4 gpu.

Nvidia has been using FP16 numbers to describe the performance of their GPUs since the Turing arch debuted (due to dslr becoming a factor in the pipeline). Amd doesn't do it yet, since they don't have bespoke ML hardware in their GPUs (that might change soon enough though).

33.4 TF figure has nothing to do with fp16 though, it's fp32 with VOPD.

 

[IllegalLemon]

33.4 TF figure has nothing to do with fp16 though, it's fp32 with VOPD.

With some Rapid Packed Math thrown in there probably.

 

[Lysandros]

With some Rapid Packed Math thrown in there probably.

That would be 66.8 TF with RPM (fp16).

 

[blastprocessor]

If Sony had put 33.4 tflops on the box the 45 percent faster rendering wouldn't make sense would it?

 

[HeisenbergFX4]

If Sony had put 33.4 tflops on the box the 45 percent faster rendering wouldn't make sense would it?

 

[IllegalLemon]

If Sony had put 33.4 tflops on the box the 45 percent faster rendering wouldn't make sense would it?

Pretty much. But, nowadays, floats aren't only used for raster alone, Rt is also part of that computation and since Amd has the RT core and the compute Unit together as one, their compute performance is measured differently. Nvidia has the compute and rt units on the same chip, but separate and thus their performance is calculated separately.