-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
PlayStation 5 Pro Unboxed, 16.7 TFLOPs GPU Compute Confirmed
SmokSmog
Member
You can't make game on it alone, you can use FP16 for some low precision stuff, it's called half precision for a reason., you will gain a little here and there.
kevboard
Member
well, if we look at the spec sheet sony included, it is absolutely possible that they actually aren't including that feature in their custom RDNA3 GPU.
the reason AMD added it is because Nvidia did. and on Nvidia GPUs it actually seems to have at least some benefits. but that's also hard to judge because the gains of their GPUs could also stem from other improvements of the architecture
Last edited:
Lysandros
Member
Sue this serial wipper's ass Sony!Please don't do this at home.
Mr Moose
Member
When was this confirmed?
DaleinCalgary
Member
So true. I will be tearing the ps5 pro box most likely just as I tore the shitty boxing in the elite headphones and the disk drive sitting in my closet.
digdug2
Member
Fess
Member
Instant PTSD!!Please don't do this at home.
XXL
Member
JimboJones
Member
Why? Packaging will be going straight into the bin.
ap_puff
Banned
That's incorrect, it's because there's a specific function in the hardware that lets it do certain kinds of math twice as fast. So "theoretically" it has double the tflops, but unless someone is specifically using that math, those extra tflops are worthless.
According to chips n' cheese this is limited to some fp32 addition commands (even though there are other cases where it could possibly be used) - either because the compiler wasn't good enough to detect more beneficial use cases, or the function simply isn't that important.
kevboard
Member
I guess it wasn't... maybe... not sure. but people just assumed it is RDNA3. not that it would make a difference tbh
skit_data
Member
10.28 IIRC
LavitzSlambert
Member
Please don't do this at home.
This was actually painful to watch, WTF?
As far as I remember He worked for some Sega magazine which probably got shut down when dreamcast died.richard used to be a dreamcast fanboy hence the animosity for playstation
Top_Gunner_82
Member
Reviews are from Wednesday.... Not sure on time but I'd think midday gmt love unboxing was today
Last edited:
Putonahappyface
Gold Member
The issue is accurate reporting. With Digital Foundry supposed developers information behind the scene,one would think they would've understand with dual issuing,the Pro's TFLOPS are 33 tflops with dual issue or 16.5 flops if there isn't a scenario where dual issuing couldn't be leverage.
Last edited:
SweetTooth
Gold Member
PS5 has 10.23TF that punches above 12TF
Magic Carpet
Gold Member
I say the Series X has a 12 that fails below 10.
SweetTooth
Gold Member
That works for me
DeepEnigma
Gold Member
You'd think we all learned this from this past gen already, even without ML/RT improvements to the Pro.
But here we are, the same circular arguments.
RT and ML performance is still measured in TFlops/TOPs.
OverHeat
« generous god »
ChiefDada
Member
Sure let's ignore context and pretend we weren't all on the same page discussing Vector ALU TFLOPS
S0ULZB0URNE
Member
Which lines up with the 4070/4070S comparisons.
Gaiff
SBI’s Resident Gaslighter
He didn’t say it meant nothing. He said it was not a good way to measure an entire system’s performance. The military analogy works because the size of a military force is still critical to its might, but it’s certainly not the sole factor.
Last edited:
PaintTinJr
Member
That is complete nonsense, of course it uses dual issue.
The reason they don't calculate it that way is that the vast majority of conventional GPU compute - beyond PSSR and nanite - that bog down the GPUs the most - like Lumen does - won't be able to benefit in doing more compute work in that time slice than the 16.7TF/s rate.
So the dual issue benefit is limited to tight and highly condensed simplistic compute or at best might allow some scatter Lumen type calculations to reduce their latency by using dual issue unless they some how alter the workload with AI training to become inferencing, but for the vast majority of the time in most games the GPU will not get that benefit.
TheThreadsThatBindUs
Member
"Dual-issue" has become my current pet peave, because RDNA was already dual issue, as was GCN before it. Both did 2 Ops per clock for each ALU in FP32.
What this is is actually quad-issue, when the architecture can do 4 Ops per clock for each ALU. I haven't read the RDNA 3 white paper so I don't know if there are any limitations, but to claim that quad-issue is not usable for games is ignorant as all hell.
It's not useful for the majority of currently available games that are not programmed to take advantage of it. That will be true also for the Pro, bear in mind; because all Pro games will be written for the base console and tweaked to run on the Pro. So unless a dev takes the time to re-write large portions of their code to utilize quad-issue wave32, then it will just do the nominal 2-Ops per clock that the base hardware does.
I think the quad-issue not usable in games thing is more of a mistranslation. It will be hard to utilitze fully the quad-issue ALU computation if the registers aren't also twice as wide, or the internal bandwidth twice the size, or the GPU Front-End twice as wide. Clearly, the GPU will be bottlenecked somewhere, and so quoting TF figures using quad-issue is even more deceptive than with dual issue because the utilization of the former will be much worse than the latter; e.g. if in real games my actual utilization of the max. theoretical 16TFLOPs is 35%, then for quad-issue, it will be likely something like 5% utilization of the maximum theoretical 32TFLOPs.
What this is is actually quad-issue, when the architecture can do 4 Ops per clock for each ALU. I haven't read the RDNA 3 white paper so I don't know if there are any limitations, but to claim that quad-issue is not usable for games is ignorant as all hell.
It's not useful for the majority of currently available games that are not programmed to take advantage of it. That will be true also for the Pro, bear in mind; because all Pro games will be written for the base console and tweaked to run on the Pro. So unless a dev takes the time to re-write large portions of their code to utilize quad-issue wave32, then it will just do the nominal 2-Ops per clock that the base hardware does.
I think the quad-issue not usable in games thing is more of a mistranslation. It will be hard to utilitze fully the quad-issue ALU computation if the registers aren't also twice as wide, or the internal bandwidth twice the size, or the GPU Front-End twice as wide. Clearly, the GPU will be bottlenecked somewhere, and so quoting TF figures using quad-issue is even more deceptive than with dual issue because the utilization of the former will be much worse than the latter; e.g. if in real games my actual utilization of the max. theoretical 16TFLOPs is 35%, then for quad-issue, it will be likely something like 5% utilization of the maximum theoretical 32TFLOPs.
SonGoku
Member
Screaming about flops as the be all is as bad as dismissing them all together
Flops are a good indicator of performance gauge within any said architecture because it takes two important metrics (clock speed & stream processors count) besides general architecture improvements.
PSSR would have been worthless if the Pro did not have the raster/rt grunt to back it up
I cant believe this needs to be spelled out, Sony did the sensible thing by providing non dual issue flops considering the dual issue number is incredibly misleading. RDNA3 cards dont perform different than their equivalent non dual issue TF RDNA2 card
Last edited:
Zathalus
Member
It might have it, it might not. Clearly Sony doesn’t think it’s worth mentioning in the manual and has made zero mention of it anywhere else.
But it really doesn’t matter as we already know what the expected performance gains are.
SportsFan581
Member
Still smaller than the OG.
Agreed. The change hasn't amounted to much in the PC space.
Agreed. The change hasn't amounted to much in the PC space.
Last edited:
ResilientBanana
Member
This is the difference of utilizing 24V and 12V. Sure, one looks better, but they are the fuckin' same. 32-Bit vs 16-Bit - More Data vs Less Data - even if you had 2 times the amount of FLOPS, you still have to do twice the work to achieve it.
Last edited:
I agree that TFlops is important. But as DF said , tflop comparison, does not matter that much anymore. But still we have to wait and see, once we see all the ps4 and ps5 games enhance and those will games coming out soon that are specifically designed for ps5 pro capabilities. Let that games do the talking.
HeWhoWalks
Gold Member
Outside of theoretical peak, no, teraflops are not that important. Dismiss them altogether? No. Make them the focal of a device's power? Silly because, again, it's about peak performance. When you see 16.7TF, it just means "more than 10.23TF". Cool. But, what's important? What can the device actually do? Well, then that teraflop number becomes far less important.
Me? I look at pixel rate, shader rade, texture fill rate, RAM amount, RAM speed, memory bandwidth/bus, and clock speeds (CPU and graphics card). The terflop number comes last in all of that.
Me? I look at pixel rate, shader rade, texture fill rate, RAM amount, RAM speed, memory bandwidth/bus, and clock speeds (CPU and graphics card). The terflop number comes last in all of that.
Last edited:
Panajev2001a
GAF's Pleasant Genius
Nowhere outside the developer docs that leaked out. I remember how Sony out the register file size of SPEs in PS3’s box manuals or the dual register file of the GS in the PS2 box manuals or the kill bit to strip degenerate triangles…
Maybe they give people numbers depending on the target audience? Devs would care about dual issue features and so they disclose it.
thicc_girls_are_teh_best
Member
How are some people so illiterate about tech they're still falling for propaganda 4 years later?
No, the Series X CPU isn't "locked" 3.6 GHz; no CPU in existence locks to a specific clock at all times unless it's malfunctioning. Series X's CPU is the same "up to" case as PS5's; it's just that Series X doesn't reallocate power from the GPU power budget to the CPU in order to help it hit its peak 3.6 GHz (SMT) or 3.8 GHz (non-SMT, which is useless for 9th-gen games because almost all of them are multi-threaded).
That is the only difference between the two systems when it comes to their CPUs. Well, at least in clocks anyway. PS5 = Smart Shift dynamically reallocates power budget from GPU to CPU (and vice-versa) to hit peaks IF required. Series X = No Smart Shift. If a game's running logic that only needs clocks of 2.2 GHz on the CPU, then it'll run that logic at 2.2 GHz on both PS5 and Series X.
It's like some people don't ever bother to check a Task Manager on their own PCs.
That account's usually full of BS. They were trying to pass off leaks & rumors a while ago, and IIRC none of them came true. Also tried hyping of SIE making a couple 3P publisher acquisitions (Square-Enix and From Software/Kadokawa IIRC); obviously neither happened and likely will never happen at this point.
Anything involving inner-workings of SIE studios as they are today, I wouldn't listen to that account for an informed take.
Last edited:
Zathalus
Member
I have no idea either way. It’s simply a possibility based on how Sony has engineered their GPUs in the past, they never align perfectly with what is out on the PC front. It’s just speculation that doesn’t really impact anything, we already know the expected performance increase from Sony and early testing.
Panajev2001a
GAF's Pleasant Genius
Come on… the dev docs leaked, it is not speculation. If you want to talk precedent look at PS4 Pro. It was a generation and something beyond PS4’s GCN family.
kevboard
Member
that is simply not true. first of all, CPUs that adjust their clock speed wasn't always a thing that existed. up until the late 2000s, PC CPUs all ran at a set clock speed that only changed if the user intervened by manually changing the clock speed.
secondly, every single console except the PS5 has a locked clock speed for both the GPU and the CPU. on Switch these clock speeds can be set by the developers, who can chose between 3 clock speed profiles in handheld mode, and are locked to a single profile in docked mode.
game consoles are not PCs. a game console has to always be predictable for the developers. and a locked clock speed is the only way to do that.
the PS5 also always clocks as high as it can. it only adjusts to such a negligible degree, and only if it hits its TDP limit, that it's almost not worth even taking note of.
AngelMuffin
Member
Wouldn’t limit that statement to just console users. Plenty of ignorant PC gamers out there as well.
Mr Moose
Member
Damn, wasting a lot of power at idle.
