ethomaz
Banned
The ideia of HBCC in a single memory pool dumbs me lol
-
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.
Wolfenstein II and Far Cry 5 will support FP16 Rapid Packed Math
- Thread starter ethomaz
- Start date
dr_rus
Member
Pro has some 1GB of DDR3 along with the fast GDDR5 but even in this case I can't think of any reason HBCC would be better than hand tuning to the memory architecture in case of consoles.
HBCC's main reason to exist is for the GPU to be able to operate on huge amounts of data, overspilling not only in system RAM but on SSDs and even into the network. Needless to say that this isn't something any console platform will be doing any time soon - and it does comes with a cost of a much higher latencies and lower bandwidths for any data not present in the VRAM "cache". Basically, not a good idea for real time graphics rendering although I can see it being used for data streaming at some point - HBCC is basically a much more generic and thorough alternative to what is called Tiled Resources in DX1x so it will probably substitute that at some point in the future.
kungfuian
Member
What's "radically"? Because it will be less than 2X in any case.[/QUOTE]
There seem to be quite a few people in this tread convinced that the benefits from RPM will be small, but Cerny very specifically said radically, which implies substation/large benefits not small ones. Call me crazy but I'll take Cerny's word over anyone's in this thread.
There seem to be quite a few people in this tread convinced that the benefits from RPM will be small, but Cerny very specifically said radically, which implies substation/large benefits not small ones. Call me crazy but I'll take Cerny's word over anyone's in this thread.
dr_rus
Member
Will you take his words over the actual results AMD have shown by themselves from the upcoming Futuremark Serra benchmark too?
Sad Affleck
Member
Eh...
http://www.gamasutra.com/view/feature/191007/inside_the_playstation_4_with_mark_.php
I've yet to see what is supercharged about the PS4 so I'd rather wait for actual results. I certainly take any promises made by AMD about their hardware with a pinch of salt since what usually follows is disappointment.
GPGPU! They make up for the shitty processors. Or something like that. Just shift your workload over, it's as easy as that.
Sad Affleck
Member
I understand, but that doesn't really mean that the hardware is punching above its weight. It means that power is not wasted and developers are taking advantage of existing features of the hardware. If Vega cards perform better in games that utilize rapid packed math I wouldn't consider that punching above their weight.
I don't know if they will be AMD spnsored or not but it's almost certain they will use Vega's Rapid Packed Math for increased performance. I'm eager to see benchmarks.
dr_rus
Member
UMA 8GBs of very fast GDDR5 memory can be considered so when comparing to PC split RAM/VRAM pools with RAM pools generally being several times slower than what PS4 has. This is less of an advantage now though, with PC videocards reaching above 8GB of VRAM alone.
All Bethesda games are now "AMD sponsored" and as for Far Cry 5 - remains to be seen as Ubisoft generally work with NV on PC enhancements but RPM/16bit math isn't exactly PC enhancement.
Negotiator
Member
Irony is cool and all, but isn't that exactly what Horizon Zero Dawn and Uncharted Lost Legacy do? I don't understand the skepticism, especially in 2017. Would they have the same graphics/physics with no GPGPU usage?
Like it or not, it was a wise decision to increase ACEs from 2 to 8 and adding extra CUs as well to increase future-proofness. That's what he meant by "supercharged". GCN 1.0 doesn't have the same GPGPU features.
Cerny hasn't lied at all, I remember him saying that it would take 3 years at least to reap the benefits of the PS4 semi-custom GPU and Uncharted 4 came out in 2016.
Sad Affleck
Member
Did it ever lead to any meaningful difference in performance though?
I don't know. Can you point me to official slides that show how they used GPGPU?
Hardly as the shared memory pool took a very hard hit in terms of bandwidth when both CPU and GPU accessed the pool.
Negotiator
Member
It's not just performance, but also memory savings.
GPGPU in a discrete memory setup (99% of PCs) requires copying data from the main RAM to the VRAM and vice versa. This means that you need double the amount of memory for certain data sets and you're also bottlenecked by PCI-e (consoles have custom buses for the CPU-GPU interconnection).
When people say that consoles punch above their weight, they mean that it's hard to buy a similar PC for the same price. PS4 Slim already costs 199 in certain retailers and 299 PCs have 4GB DDR3 with crappy IGPs. You pay 50% more and you'll never have the same gaming performance.
Scorpio costs 499 and you'll probably need a much more expensive PC to compete with it.
Soldering everything (APU, DRAM chips) onto the same motherboard and introducing certain hardware optimizations (semi-custom APUs) makes consoles more competitive than equivalent PCs with upgradable sockets and DIMM slots. It's a trade-off, if you will.
Sure, feel free to ask this guy. He knows his job better than all of us.
And this one regarding Horizon:
http://wccftech.com/horizon-zero-dawn-gpu-based-procedural-tech/
http://www.neogaf.com/forum/showthread.php?t=1350648
Negotiator
Member
You're welcome bro.
The other keyword is "potential" which is true it has the potential to boost the performance a lot we already seen the Mass Effect slide that said they go a boost of 30% from using fp16 & we have seen games go from 1080p on PS4 to native 4K on PS4 Pro.
kungfuian
Member
The other keyword is "potential" which is true it has the potential to boost the performance a lot we already seen the Mass Effect slide that said they go a boost of 30% from using fp16 & we have seen games go from 1080p on PS4 to native 4K on PS4 Pro.[/QUOTE]
There are a lot of people in this industry that say a lot of meaningless markety BS but Cerny is not one of them. He's a pretty straight shooter. I'm not sure Cerny could have been more clear on the subject and hopefully over the next year or so RPM will bear fruit.
I 100% agree with a lot of your posts, especially those about perceptible differences between Pro and X!X. For ref I game on both a PRO and a really nice rig with a 1080 in it (need to grab a 1080 TI but waiting for price drop) and I feel like the differences are pretty marginal. I can crank up tomb raider on the PC to max settings and it's hardly different than on PRO. And for VR my PRO running PSVR vs. PC running VIve feels very similar too. Sony just knows how to make really good hardware, plain and simple!
dr_rus
Member
+30% in a relatively simple and short checkerboarding resolve shader which probably takes less than 5% of total frame rendering time. Thus the resulting influence on final performance from this is likely less than 2%.
Sad Affleck
Member
Ok, this is going to be
Using "consoles punch above their weight" as a cost argument is wrong, in my opinion, because it doesn't take into account the different business models that the PC and console industries employ. In my mind it's as weird as saying something like "Clash of Clans punches above its weight compared to Skyrim because Clash of Clans is free and Skyrim is $60." PC hardware manufacturers make their profit from the initial sale, console manufacturers aim to earn most of their profit from game sales and online subscriptions .
Determining if consoles punch above their weight compared to PC requires a comparison based on hardware and relative performance. Determining if a console is better value compared to PC, a different argument than the weight-punching one, requires the study of a whole lot of variables connected to a person's gaming and buying habits so it's a very personal decision.
Now, on the topic of Xbox One X's 6-Teraflop GPU versus PS4 Pro's 4.2-Teraflop-but-with-rapid-packed-math GPU: One could make the argument that PS4 Pro's GPU technology allows it to exceed its on-paper Teraflop rating in games that support Rapid Packed Math, therefore PS4 Pro will be punching above its weight. It is absolutely true that, unless XB1X has GPU customizations that will help it exceeed its teraflop rating and depending on the amount of performance increase that Rapid Packed Math can achieve, the difference in performance between PS4 Pro and XB1X in games that support RPM will be less than the difference suggested by raw teraflop numbers.
That said, in my opinion it's still wrong to say that PS4 Pro will be punching above its weight. PS4 Pro's support for RPM is a hardware feature. The system was designed that way and it will be punching exactly at its weight in games that support RPM. It will actually be punching below its weight in games that won't support RPM because a hardware feature of the GPU will be going unused. Using only the number of teraflops isn't enough to determine if a piece of hardware is punching above its weight because if it was enough then the undisputed king of punching above one's weight is... Nvidia! The RX 480 is 5.8 Teraflops and the GTX 1060 is 4.4 Teraflops yet the Nvidia card matches the Radeon's performance.
This is why there is only one reliable way of measuring and comparing console and PC GPU performance: Real-world testing in actual games. Numbers and hardware specs give us a ballpark figure of performance but almost never tell the whole story. Interestingly enough, they did tell the whole story for the original PS4 and Xbox One because those two systems were so similar in hardware that the power difference on paper ended up being confirmed in the real world.
That's not what they said that's what people came up with in their heads the slide only mentions that it was used in checkerboard rendering resolve because the slide is about checkerboard rendering & when they said 30% performance improvement they was talking about overall. why would they even waste their time talking about a 2% gain?
Sad Affleck
Member
Every bit counts on a fixed spec. 2% here, 3% there, they can add up to a decent performance increase.
That's not what they said that's what people came up with in their heads the slide only mentions that it was used in checkerboard rendering resolve because the slide is about checkerboard rendering & when they said 30% performance improvement they was talking about overall. why would they even waste their time talking about a 2% gain?
I understood that the 30% was savings between 1800p and 1800p-cb. What you imply is that the checkerboarding happened due to fp16. The slides say that cd-resolve was saving 30% which is only a part of the whole pipeline.
dr_rus
Member
That's not what they said that's what people came up with in their heads the slide only mentions that it was used in checkerboard rendering resolve because the slide is about checkerboard rendering & when they said 30% performance improvement they was talking about overall. why would they even waste their time talking about a 2% gain?
This has been discussed to death on places like B3D since March, and yes, this +30% is for the shader which is explicitly mentioned in the slide. What is still unknown is how much frame time this shader even takes but considering that it's a resolve shader it's likely a fairly simple program - which is also why it's even possible to use FP16 in it.
As for why would they mention it - because it's getting harder to extract performance gains from console GCN h/w and even +2% overall is a significant win these days.
Edit: Actually, I've looked up the presentation and they do in fact state the frame times of "CB resolve + temporal AA" stage which varies between titles by almost 100% but remains on the lower side of total frame time:
So yeah, I was actually correct about this stage being at less than 5% of total frame time as it is close to 7% in BF1 when coupled with TAA shader.
There's also this slide where they specifically call their checkerboarding "packed" hinting that this is the only part of the engine using RPM right now:
Negotiator
Member
Let's discuss some potential RPM applications with "radical" performance increases:
http://cdn.wccftech.com/wp-content/...RX-Vega-GPU-Architecture_Tress-FX-NCU-RPM.png
Who says that ND cannot do this in TLOU2 and Ellie's hair? Wouldn't the PS4 Pro be able to get a nice performance boost compared to OG PS4? And they can spend the rest of the power however they see fit.
https://www.youtube.com/watch?v=qPNiIeKMHyg
I remember TressFX being very computationally intensive in Tomb Raider 2013, so RPM should be very helpful in hair rendering/animation/physics.
Can you buy a $199 PC that runs Uncharted Lost Legacy/Horizon Zero Dawn? How can the PS4 do that while being so cheap? How many nVidia teraflops would you need to run these games on PC? Would you surprised if they needed a minimum of GTX 970 (2x flops) to cover various overheads? And no, contrary to the popular belief, 750 Ti wouldn't suffice.
https://twitter.com/id_aa_carmack/status/436012673243693056
https://twitter.com/id_aa_carmack/status/50277106856370176
A pet peeve of mine is when people claim that nVidia flops are always so much better, without using the proper context. Have you thought that perhaps nVidia GPUs are tuned for pure graphics (pixel/vertex shaders) workloads and legacy APIs (DX11/OpenGL), while on the other hand Radeon GPUs are tuned for mixed (graphics/compute) workloads running concurrently on the same CUs and low-level APIs (DX12/Vulkan)?
Let's say that GCN achieves 70% ALU occupancy in DX11 and a Maxwell reaches 90%. The latter seems more "efficient", right? But what if GCN achieves 98-99% ALU occupancy in Vulkan? That's exactly what happens in games like Doom and that's why you see cards like RX 480 "punching above their weight" and surpassing the more "efficient" GTX 1060.
https://twitter.com/SebAaltonen/status/888875323961593856
No offense, but I'll take a (skilled) programmer's word over anyone else's.
http://cdn.wccftech.com/wp-content/...RX-Vega-GPU-Architecture_Tress-FX-NCU-RPM.png
Who says that ND cannot do this in TLOU2 and Ellie's hair? Wouldn't the PS4 Pro be able to get a nice performance boost compared to OG PS4? And they can spend the rest of the power however they see fit.
https://www.youtube.com/watch?v=qPNiIeKMHyg
I remember TressFX being very computationally intensive in Tomb Raider 2013, so RPM should be very helpful in hair rendering/animation/physics.
I don't understand the disagreement here.
Can you buy a $199 PC that runs Uncharted Lost Legacy/Horizon Zero Dawn? How can the PS4 do that while being so cheap? How many nVidia teraflops would you need to run these games on PC? Would you surprised if they needed a minimum of GTX 970 (2x flops) to cover various overheads? And no, contrary to the popular belief, 750 Ti wouldn't suffice.
https://twitter.com/id_aa_carmack/status/436012673243693056
https://twitter.com/id_aa_carmack/status/50277106856370176
A pet peeve of mine is when people claim that nVidia flops are always so much better, without using the proper context. Have you thought that perhaps nVidia GPUs are tuned for pure graphics (pixel/vertex shaders) workloads and legacy APIs (DX11/OpenGL), while on the other hand Radeon GPUs are tuned for mixed (graphics/compute) workloads running concurrently on the same CUs and low-level APIs (DX12/Vulkan)?
Let's say that GCN achieves 70% ALU occupancy in DX11 and a Maxwell reaches 90%. The latter seems more "efficient", right? But what if GCN achieves 98-99% ALU occupancy in Vulkan? That's exactly what happens in games like Doom and that's why you see cards like RX 480 "punching above their weight" and surpassing the more "efficient" GTX 1060.
https://twitter.com/SebAaltonen/status/888875323961593856
No offense, but I'll take a (skilled) programmer's word over anyone else's.
Sad Affleck
Member
You can't determine if a piece of hardware is punching above or below its weight based on price. You measure real-world performance to determine that.
Well, regarding your first question, even a $5999 PC won't run these games because they are exclusive, therefore using them in a PC/console comparison is utterly pointless. As for the second question, it's basic console market economics. PS4 costs $199 because it is 2012 technology still being sold today as an entry point to an ecosystem where the real money is made.
Assuming that developers have access to the same budget and time as the PS4 version? If I had to guess, a graphics card with 10% more horsepower than the PS4.
Surprised? That's an understatement. I would be completely shocked if you needed a GTX 970 to match PS4's GPU. After four years of multiplatform releases I don't know of a single game that requires a GTX 970 for PS4-like performance. Do you?
I think you misunderstood my post in this part because I 100% agree. I never said that Nvidia flops are better, in fact my main point of contention with OnQ123 is that he is quoting specs without the proper context. So just to be clear, the only thing that ultimately matters is real-world performance. Everything else is just speculation and theories until real benchmarks appear.
None taken
Do you wake up with me on your mind or something? I always have context with my post you just seem to have a hard time understanding for some reason.
Sad Affleck
Member
My reference was about previous comments and threads. Don't worry, I said I won't be criticizing you on a personal level anymore.
Doom coming to Switch this holiday, Wolfenstein II: The New Colossus in 2018
Makes sense that they would put extra work into running fp16 code
Makes sense that they would put extra work into running fp16 code