It is confirmed that Scorpio has a Polaris GPU, basically a Radeon r480. While Scorpio GPU has 40 compute units compared to stock r480 36, it operates at lower clock rate giving more or less the same performance. R480 is close to 1060 performance wise, being faster in some cases and slower in others, all depending on the game engine and DX12 optimizations. Still, in the end it is nowhere near 1070, a 4k entry card, performance wise. If MS decides to force "true 4k" requirement on the developers, as opposed to PS4 Pro checkerboard cheating, prepare for either low-to-medium graphics presets or 20-30 fps max.
It is confirmed that Scorpio has a Polaris GPU, basically a Radeon r480. While Scorpio GPU has 40 compute units compared to stock r480 36, it operates at lower clock rate giving more or less the same performance. R480 is close to 1060 performance wise, being faster in some cases and slower in others, all depending on the game engine and DX12 optimizations. Still, in the end it is nowhere near 1070, a 4k entry card, performance wise. If MS decides to force "true 4k" requirement on the developers, as opposed to PS4 Pro checkerboard cheating, prepare for either low-to-medium graphics presets or 20-30 fps max.
To help the DSBR do its thing, AMD is fundamentally altering the availability of Vega's L2 cache to the pixel engine in its shader clusters. In past AMD architectures, memory accesses for textures and pixels were non-coherent operations, requiring lots of data movement for operations like rendering to a texture and then writing that texture out to pixels later in the rendering pipeline. AMD also says this incoherency raised major synchronization and driver-programming challenges.
To cure this headache, Vega's render back-ends now enjoy access to the chip's L2 cache in the same way that earlier stages in the pipeline do. This change allows more data to remain in the chip's L2 cache instead of being flushed out and brought back from main memory when it's needed again, and it's another improvement that can help deferred-rendering techniques.
Phil Spencer correctly identifies AMD GPU issues with it's high TFLOPS i.e. memory bandwidth.
When you talk to me about Scorpio, the term I use about the architecture isn’t the six teraflops which is obviously what we’ve announced, it’s balance. Really what it is, is you want a platform that is balanced between memory bandwidth, GPU power, you know, your ability to move memory and [an] amount of memory around in many ways is more inhibiting to the performance of your game than absolute teraflops on any one of the individual pieces, and when we designed Scorpio we really thought about this balanced rig that could come together at a price-point. Like, I want Scorpio to be at a console price-point, I’m not trying to go and compete with a high-end rig. And because we’re building one spec, we’re able to look at the balance between all the components and make sure that it’s something we really hit that matters to consumers and gamers.
The lesson from Nvidia Maxwell/Pascal designs are increasing memory bandwidth while increasing TFLOPS.
Notice both Vega and Phil Spencer's statement Scorpio's GPU memory handling targets memory movement issues.
RX-480 doesn't fix the old memory bandwidth bottleneck issues.
NVIDIA rubs it in on why AMD GPUs has less consistent performance.
Other factors at work is the "pack math" features.
Relax, Scorpio is not GTX 1080 and GTX 1080 Ti level GPUs.
For Battlefield 1 benchmarks, RX-480 (5.8 TFLOPS) and R9-290 (4.8 TFLOPS) has similar memory bandwidth hence similar frame rate result. This is an example for memory bandwidth bound issues. At least 1 TFLOPS gimped by RX-480's memory bandwidth.
R9-390X doesn't have delta memory compression, hence it's effective memory bandwidth is around 311 GB/s.
I'm recycling Beyond3D's effective memory bandwidth benchmarks to show you the effective memory bandwidth impacts on frame rate.
To workout Hawaii GCN's memory subsystem efficiency: R9-290X's 263 GB/s / 320 GB/s = 82 percent efficient.
R9-390X's 384 GB/s physical memory bandwidth would be around 315.6 GB/s effective memory bandwidth. Hawaii GCN doesn't have any memory delta compression i.e. this GPU design is very old.
Effective memory bandwidth difference between RX-480 and R9-390X is 1.195X
RX-480 reached 35.8 fps
R9-390X reached 42.7 fps
The frame rate between 35.8 fps and 42.7 fps difference is 1.192X
Note how effective memory bandwidth difference between RX-480 and R9-390X has matched the frame rate difference.
Since we establish that effective memory bandwidth impacts BattleField 1, let's predicate Scorpio GPU's frame rate results via 326 GB/s physical memory bandwidth.
Since PS4 Pro is showing RX-470D like results, I'll treat Scorpio the same.
For Scorpio: (326 GB/s x Polaris memory subsystem efficiency 77 percent) x Polaris 1.37X memory delta compression = 343.90 GB/s estimated effective memory bandwdith.
Effective memory bandwidth difference between RX-480 and Scorpio's estimated effective memory bandwidth is ~1.30X.
Apply 1.30X on RX-480's 35.8 fps = 46.6 fps estimate from memory bandwidth gain.
GTX 1070's frame rate is 46.6 fps LOL. This is just RX-480 gaining 326 GB/s memory bandwidth with Polaris delta memory compression bottleneck reduction.
For F1 2016, similar result.
Note why Scorpio's Forza example result landed in GTX 1070 range.
RX-480's 256 bit GDDR5-8000 memory setup is bottlenecking 5.8 TFLOPS.
It is confirmed that Scorpio has a Polaris GPU, basically a Radeon r480. While Scorpio GPU has 40 compute units compared to stock r480 36, it operates at lower clock rate giving more or less the same performance. R480 is close to 1060 performance wise, being faster in some cases and slower in others, all depending on the game engine and DX12 optimizations. Still, in the end it is nowhere near 1070, a 4k entry card, performance wise. If MS decides to force "true 4k" requirement on the developers, as opposed to PS4 Pro checkerboard cheating, prepare for either low-to-medium graphics presets or 20-30 fps max.
I really dont know how to begin to respond to what you said.
You cant judge GPU capabilties by compute units, GPU family, or TFlops alone.
This is the mistake that so many people on this board make.
Think of it like a game engine. If I heavily modify the Unity Engine so that it is capable of open worlds where everything can be interacted with.............its still called Unity. The baseline engine may not be capable of it but a customized version may be.
You seem to be judging the Polaris family on its limitations, ignoring the fact that both MS and Sony took steps to minimize those limitations when they created their refresh.
There is a ton of features that Pro and Scorpio have that arent found in the 480 so comparing them is pointless.
Not to mention Scorpio has more ram, more bandwidth, more customizations, on board DX12, etc.
Its crazy to try to limit it to a 480 just to make yourself feel better.
Who are these people that think console get top of the line GPUs ?
This has never been the case afaik.
Scorpio is in fact AMD's first post-RX-480 with significantly larger memory bandwidth and larger 44 CU scale FinFET GPU design i.e. FinFET era Hawaii XT with bottleneck fixes.
In terms of BOM cost, GTX 1070 is not high end i.e. GTX 1070's PCB 256 bit GDDR5-8000 is in the class as RX-480's PCB 256 bit GDDR5-8000. Note why NVIDIA has large revenue profits. Fools are suckered into thinking that GTX 1070 is high end product when BOM cost is similar to RX-480.
GTX 1070 GP104 chip is only 36 percent larger than RX-480 chip. GTX 1070's 317 mm^2 chip is about 75 percent operational.
I have my own GTX 980 Ti AIB OC and GTX 100 Ti, Scorpio's results are pretty good. AMD needs to be competitive and return NVIDIA's RX-480 class PCB back down to normal pricing.
RX-390X's Hawaii design has existed since year 2013 with NVIDIA Kepler era and it's f.vcking old.
PS4 Pro's 4.2 TFLOPS GPU with 218 GB/s memory bandwidth is bottle-necked by memory bandwidth and bottleneck is worst with RX-480's 5.8 TFLOPS with 256 GB/s memory bandwidth.
In terms of TFLOPS vs memory bandwidth ratio, Polaris 10 is step backwards from R9-390X.
@hrt_rulz01: they made that comparison based on the Forza 6 footage that Turn 10 guys showed them at MS presentation. DF didn't make the actual comparison between F6 Apex ultra settings assets and F6 Scoripio build themselves. Stock F6 runs at 1080p and 60 fps on stock XB1 which really says a lot about how optimized the game is. Apex has more high definition assets so it is hard to compare.
Scorpio is in fact AMD's first post-RX-480 with significantly larger memory bandwidth and larger 44 CU scale FinFET GPU design.
In terms of BOM cost, GTX 1070 is not high end i.e. GTX 1070's PCB 256 bit GDDR5-8000 is in the class as RX-480's PCB 256 bit GDDR5-8000. Note why NVIDIA has large revenue profits. Fools like you are suckered into thinking that GTX 1070 is high end product when BOM cost is similar to RX-480.
GTX 1070 GP104 chip is only 36 percent larger than RX-480 chip. GTX 1070's 317 mm^2 chip is about 75 percent operational.
It's quite simple why Nvidia has large revenue profits:
AMD pretty much dropped from the high end GPU market, effectively creating a monopoly for Nvidia.
Nvidia being the sole vendor for high end GPUs drove up the prices.
A high end GPU used to be 400-500. Now it's around 800-900.
Massive amounts of profit for Nvidia.
Just checked: the GTX1070 is around 400-500.
"Fools like you are suckered into thinking that GTX 1070 is high end product"
@ronvalencia: This only confirms that NVIDIA GPUs has better performance at the same production cost.
Examples for why Pascal GP104 has less bottlenecks
1. Pascal's Pixel Engine (ROPS) is connected to L2 cache instead of AMD's Pixel Engine being connected to memory controller.
On AMD GPUs
AMD GPU's Pixel Engine is connected to memory controller which is a bottleneck. AMD GPU can't do full tiling cache render with Pixel Engine being connected to memory controller.
On XBO, ESRAM enables tile rendering for AMD GPUs but MS compromised compute power with large 32MB ESRAM. Nvidia found a balance between high TFLOPS with large enough cache for tile cache rendering.
Note why AMD using Async compute i.e. Compute Engine is connected to L2 cache which enables AMD GPUs to preform some limited tiling render techniques e.g. AMD's Tiled Forward+ render with compute shaders demo. Hence why some games like Doom Vulkan shows AMD GPUs slotting into NVIDIA's TFLOPS rankings. Note that Pixel Engine path can't be completely be avoided, hence AMD GPUs hits Pixel Engine-to-memory controller bottleneck.
The statement from Digital Foundry was that in Forza it was exhibiting 1070/Fury-class performance. I think you'll see between 390x and Fury X performance in general, or in simple terms better than GTX 980 performance.
P.S.- In Forza, Hitman, and Battlefield 1 you'll see exactly that...1070/Fury-class performance.
Scorpio is in fact AMD's first post-RX-480 with significantly larger memory bandwidth and larger 44 CU scale FinFET GPU design.
In terms of BOM cost, GTX 1070 is not high end i.e. GTX 1070's PCB 256 bit GDDR5-8000 is in the class as RX-480's PCB 256 bit GDDR5-8000. Note why NVIDIA has large revenue profits. Fools like you are suckered into thinking that GTX 1070 is high end product when BOM cost is similar to RX-480.
GTX 1070 GP104 chip is only 36 percent larger than RX-480 chip. GTX 1070's 317 mm^2 chip is about 75 percent operational.
It's quite simple why Nvidia has large revenue profits:
AMD pretty much dropped from the high end GPU market, effectively creating a monopoly for Nvidia.
Nvidia being the sole vendor for high end GPUs drove up the prices.
A high end GPU used to be 400-500. Now it's around 800-900.
Massive amounts of profit for Nvidia.
Just checked: the GTX1070 is around 400-500.
"Fools like you are suckered into thinking that GTX 1070 is high end product"
HAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHahahahahahahaha xD
AMD is busy with Scorpio R&D which is effectively the missing "supercharged" 44 CU larger chip RX-490 PC SKU.
During Xbox 360 R&D, ATI was late with thier DX10 Radeon HD XT SKU i.e. ATI was focusing on Xbox 360 R&D over PC's high end market. This is not the first time that ATI/AMD was missing in PC's high end market
@hrt_rulz01: they made that comparison based on the Forza 6 footage that Turn 10 guys showed them at MS presentation. DF didn't make the actual comparison between F6 Apex ultra settings assets and F6 Scoripio build themselves. Stock F6 runs at 1080p and 60 fps on stock XB1 which really says a lot about how optimized the game is. Apex has more high definition assets so it is hard to compare.
Scorpio's first 4K demo, the following screenshot has XBO level graphics settings with PC's 4K assets in 4K resolution.
Scorpio's later demo was with PC's ultra max settings and PC 4K assets.
The track was Forza 6's Nürburgring wet track which incidentally, it's PC's paid version of Forza 6. DF's tech demo reveal may have cause surged purchases for Forza 6's Nürburgring DLC. MS is hitting two birds with one stone.
I tested Nürburgring wet rack with Ultra max setting, MSAA 4X and 4K resolution on my GTX 980 Ti AIB OC. It was defeated by Scorpio by sustained 4K/60 fps at ~88 percent GPU usage.
Scorpio wasn't designed to compete against the larger scale 64 CU Vega 10.
Scorpio's 44 CU GPU slots between 36 CU Polaris 10 and 64 CU Vega 10
It's 40 CU and it IS Polaris, not something in between Polaris and Vega, with reduced clock speed to boot. The only redeeming feature is the increased memory bandwidth.
@ronvalencia: there's no proof that Forza 6 Scorpio build has the same 4k assets as Apex. Of course MS will say they are the same but until DF will get Scorpio themselves and check them side by side we can only speculate.
@ronvalencia: there's no proof that Forza 6 Scorpio build has the same 4k assets as Apex. Of course MS will say they are the same but until DF will get Scorpio themselves and check them side by side we can only speculate.
LOL, you didn't read the article.
"This is us. This is ForzaTech running 60 frames a second, 4K," Tector says proudly. "We're still running with settings that we would have used in Forza 6. Since it's Xbox, we're using EQAA, so it's like a 4:2 EQAA. That's the actual GP utilisation so we're only using 60 per cent of the compute to get to this. Importantly, I know I've just said it's like a Forza 6 set-up, but this is also including 4K content, so all of our build system - we've got authored assets for this set of the models, cars, tracks everything. We pushed it through and made sure the 4K textures were flowing through."
Also, NPC cars' LOD was disabled for Scorpio i.e. showing the new GPU's the new geometry power.
"The crazy story here is that we've gone over our PC ultra settings and for everything that's GPU-related, we've been able to max it - and that's what we're running at, 88 per cent," says Tector, pointing to the utilisation data at the top of the screen. Right beneath it is the anti-aliasing setting - 4x, or rather 8:4x using the Radeon EQAA hardware AA.
Scorpio's max setting exceeded PC's ultra settings with MSAA 4X instead of EQAA.
Wet track has plenty of semi-transparent/alpha and blend effects which hammers ROPS (Pixel Engine path).
It's 40 CU and it IS Polaris, not something in between Polaris and Vega, with reduced clock speed to boot. The only redeeming feature is the increased memory bandwidth.
there's no proof that Forza 6 Scorpio build has the same 4k assets as Apex. Of course MS will say they are the same but until DF will get Scorpio themselves and check them side by side we can only speculate.
There is no 40-44CU Polaris on PC. Scorpio is the missing 490/490x. RX 480 is crippled by lack of memory bandwidth to the point that the 5.1TFLOPS R9 390 outperforms the 5.8TFLOPS RX 480 in 4k by 12% across over a dozen titles. RX 480 cannot match Scorpio's bandwidth even with max OC. Period.
Scorpio is at least 25% more powerful than a RX 480 oc'd to 1305MHz core, 2250MHz mem in Forza. It's the fucking RX 490, man.
---
What Nvidia GPU would the RX 490/490x slot against? Just answer, no spin.
This is so stupid. Consoles are not PCs. They will never perform like a modern PC. Therefore, the strength of consoles is exclusives. Therefore, Sony knows what they are doing.
This is so stupid. Consoles are not PCs. They will never perform like a modern PC. Therefore, the strength of consoles is exclusives. Therefore, Sony knows what they are doing.
End of story.
"Sony has exclusives, therefore Scorpio cannot perform like a PC." ???
Name a GPU that can match Scorpio's performance in Forza @ 4k/60fps, Ultra settings. Then quote it's price.
AMD is busy with Scorpio R&D which is effectively the missing "supercharged" 44 CU larger chip RX-490 PC SKU.
During Xbox 360 R&D, ATI was late with thier DX10 Radeon HD XT SKU i.e. ATI was focusing on Xbox 360 R&D over PC's high end market. This is not the first time that ATI/AMD was missing in PC's high end market
I gotta say I own a GTX970 and it's great bang for the buck. But all the Nvidia graphics cards I see now are just not good price/performance wise. Especially Titan graphics cards are very bad in that regard.
I can't wait to see AMD's Vega offerings, because I'd like to buy a graphics card at an acceptable price point again, which I haven't seen since the GTX970.
Scorpio wasn't designed to compete against the larger scale 64 CU Vega 10.
Scorpio's 44 CU GPU slots between 36 CU Polaris 10 and 64 CU Vega 10
It's 40 CU and it IS Polaris, not something in between Polaris and Vega, with reduced clock speed to boot. The only redeeming feature is the increased memory bandwidth.
@ronvalencia: there's no proof that Forza 6 Scorpio build has the same 4k assets as Apex. Of course MS will say they are the same but until DF will get Scorpio themselves and check them side by side we can only speculate.
Wrong, Scorpio has a total 44 CU with 40 enabled and 4 disabled yield issues just like R9-390 Pro.
An AMD Polaris 10 "Ellesmere" based graphics card (RX 470 or RX 480) was taken apart down to its die, for science. Close up die-shots of the silicon reveal that 36 GCN compute units is all that the silicon has, and that the RX 480 indeed maxes out this stream processor count, with 2,304 stream processors at its disposal.
The Scorpio Engine processor measures 360mm2 and features seven billion transistors. We got to see the chip plan, with the four shader engines occupying the majority of the die, skewed towards the left of the layout. Each SE actually has 11 compute units, with one disabled per block to increase chip yield on the production line. To the right of the GPU sit the two clusters of custom CPU cores, while the memory interfaces skirt the edges of the chip.
Four Shader Engine x 11 Compute Units = 44 Compute Units
Hawaii XT's block diagram with four Shader Engine x 11 Compute Units = 44 Compute Units
According to Digital Foundry, 60 deep graphics pipeline changes was applied to an existing design.
For RX 4XX series, SKU with greater than 256 bit bus has 4K gaming guidance i.e. RX-490 SKU.
To help the DSBR do its thing, AMD is fundamentally altering the availability of Vega's L2 cache to the pixel engine in its shader clusters. In past AMD architectures, memory accesses for textures and pixels were non-coherent operations, requiring lots of data movement for operations like rendering to a texture and then writing that texture out to pixels later in the rendering pipeline. AMD also says this incoherency raised major synchronization and driver-programming challenges.
To cure this headache, Vega's render back-ends now enjoy access to the chip's L2 cache in the same way that earlier stages in the pipeline do. This change allows more data to remain in the chip's L2 cache instead of being flushed out and brought back from main memory when it's needed again, and it's another improvement that can help deferred-rendering techniques.
Phil Spencer correctly identifies AMD GPU issues with it's high TFLOPS i.e. memory bandwidth.
When you talk to me about Scorpio, the term I use about the architecture isn’t the six teraflops which is obviously what we’ve announced, it’s balance. Really what it is, is you want a platform that is balanced between memory bandwidth, GPU power, you know, your ability to move memory and [an] amount of memory around in many ways is more inhibiting to the performance of your game than absolute teraflops on any one of the individual pieces, and when we designed Scorpio we really thought about this balanced rig that could come together at a price-point. Like, I want Scorpio to be at a console price-point, I’m not trying to go and compete with a high-end rig. And because we’re building one spec, we’re able to look at the balance between all the components and make sure that it’s something we really hit that matters to consumers and gamers.
The lesson from Nvidia Maxwell/Pascal designs are increasing memory bandwidth while increasing TFLOPS.
Notice both Vega and Phil Spencer's statement Scorpio's GPU memory handling targets memory movement issues.
RX-480 doesn't fix the old memory bandwidth bottleneck issues.
PC version would have full 44 CU enabled with higher clock speed.
It is confirmed that Scorpio has a Polaris GPU, basically a Radeon r480. While Scorpio GPU has 40 compute units compared to stock r480 36, it operates at lower clock rate giving more or less the same performance. R480 is close to 1060 performance wise, being faster in some cases and slower in others, all depending on the game engine and DX12 optimizations. Still, in the end it is nowhere near 1070, a 4k entry card, performance wise. If MS decides to force "true 4k" requirement on the developers, as opposed to PS4 Pro checkerboard cheating, prepare for either low-to-medium graphics presets or 20-30 fps max.
Except you can't directly compare the specs of PC and a Console. That's not how it works. They have different approach using resources as Digital Foundry said...But just to be safe..just wait for the release for actual confirmation when Scorpio launches.
@navyguy21 said:
@Orchid87 said:
It is confirmed that Scorpio has a Polaris GPU, basically a Radeon r480. While Scorpio GPU has 40 compute units compared to stock r480 36, it operates at lower clock rate giving more or less the same performance. R480 is close to 1060 performance wise, being faster in some cases and slower in others, all depending on the game engine and DX12 optimizations. Still, in the end it is nowhere near 1070, a 4k entry card, performance wise. If MS decides to force "true 4k" requirement on the developers, as opposed to PS4 Pro checkerboard cheating, prepare for either low-to-medium graphics presets or 20-30 fps max.
I really dont know how to begin to respond to what you said.
You cant judge GPU capabilties by compute units, GPU family, or TFlops alone.
This is the mistake that so many people on this board make.
Think of it like a game engine. If I heavily modify the Unity Engine so that it is capable of open worlds where everything can be interacted with.............its still called Unity. The baseline engine may not be capable of it but a customized version may be.
You seem to be judging the Polaris family on its limitations, ignoring the fact that both MS and Sony took steps to minimize those limitations when they created their refresh.
There is a ton of features that Pro and Scorpio have that arent found in the 480 so comparing them is pointless.
Not to mention Scorpio has more ram, more bandwidth, more customizations, on board DX12, etc.
Its crazy to try to limit it to a 480 just to make yourself feel better.
Its a game console not an ex girlfriend.
Everyone knows this. (well at least I hope so) They could release better improvements every year, but not 300x better. It's a business prospective, and everyone has to understand that.
As for the rest of your post. I believe it. The single biggest factor between PC and Consoles is that Consoles are optimized much more heavily. So even if specs don't seem groundbreaking, it can still do some heavy lifting if you know what I mean.
@Orchid87: Your dealing with the biggest clowns on the internet. Let's apply generality from one tech demo towards all future games. Third party will be a shitty compromised 4K/sub 30fps.
To be clear, then: Project Scorpio doesn't feature Ryzen cores, but the Xbox team are not so concerned about this. "On the CPU side of things, we could still meet our design goals with the custom changes we made," Kevin Gammill points out. "At the end of the day we are still a consumer product. We want to hit the price-points where consumers want to purchase this. It's about balancing the two."
The new x86 cores in Scorpio are 31 per cent faster than Xbox One's, with extensive customisation to reduce latency in order to keep the processor occupied more fully, while CPU/GPU coherency also gets a performance uplift. There's significant hardware offloading too - some of which is inherited from Xbox One, some of which is radically new
It is confirmed that Scorpio has a Polaris GPU, basically a Radeon r480. While Scorpio GPU has 40 compute units compared to stock r480 36, it operates at lower clock rate giving more or less the same performance. R480 is close to 1060 performance wise, being faster in some cases and slower in others, all depending on the game engine and DX12 optimizations. Still, in the end it is nowhere near 1070, a 4k entry card, performance wise. If MS decides to force "true 4k" requirement on the developers, as opposed to PS4 Pro checkerboard cheating, prepare for either low-to-medium graphics presets or 20-30 fps max.
Don't hate MS for releasing a console that can do 4k better than a 1070.
To help the DSBR do its thing, AMD is fundamentally altering the availability of Vega's L2 cache to the pixel engine in its shader clusters. In past AMD architectures, memory accesses for textures and pixels were non-coherent operations, requiring lots of data movement for operations like rendering to a texture and then writing that texture out to pixels later in the rendering pipeline. AMD also says this incoherency raised major synchronization and driver-programming challenges.
To cure this headache, Vega's render back-ends now enjoy access to the chip's L2 cache in the same way that earlier stages in the pipeline do. This change allows more data to remain in the chip's L2 cache instead of being flushed out and brought back from main memory when it's needed again, and it's another improvement that can help deferred-rendering techniques.
Phil Spencer correctly identifies AMD GPU issues with it's high TFLOPS i.e. memory bandwidth.
When you talk to me about Scorpio, the term I use about the architecture isn’t the six teraflops which is obviously what we’ve announced, it’s balance. Really what it is, is you want a platform that is balanced between memory bandwidth, GPU power, you know, your ability to move memory and [an] amount of memory around in many ways is more inhibiting to the performance of your game than absolute teraflops on any one of the individual pieces, and when we designed Scorpio we really thought about this balanced rig that could come together at a price-point. Like, I want Scorpio to be at a console price-point, I’m not trying to go and compete with a high-end rig. And because we’re building one spec, we’re able to look at the balance between all the components and make sure that it’s something we really hit that matters to consumers and gamers.
The lesson from Nvidia Maxwell/Pascal designs are increasing memory bandwidth while increasing TFLOPS.
Notice both Vega and Phil Spencer's statement Scorpio's GPU memory handling targets memory movement issues.
RX-480 doesn't fix the old memory bandwidth bottleneck issues.
NVIDIA rubs it in on why AMD GPUs has less consistent performance.
Other factors at work is the "pack math" features.
Relax, Scorpio is not GTX 1080 and GTX 1080 Ti level GPUs.
For Battlefield 1 benchmarks, RX-480 (5.8 TFLOPS) and R9-290 (4.8 TFLOPS) has similar memory bandwidth hence similar frame rate result. This is an example for memory bandwidth bound issues. At least1 TFLOPS gimped by RX-480's memory bandwidth.
R9-390X doesn't have delta memory compression, hence it's effective memory bandwidth is around 311 GB/s.
I'm recycling Beyond3D's effective memory bandwidth benchmarks to show you the effective memory bandwidth impacts on frame rate.
To workout Hawaii GCN's memory subsystem efficiency: R9-290X's 263 GB/s / 320 GB/s = 82 percent efficient.
R9-390X's 384 GB/s physical memory bandwidth would be around 315.6 GB/s effective memory bandwidth. Hawaii GCN doesn't have any memory delta compression i.e. this GPU design is very old.
Effective memory bandwidth difference between RX-480 and R9-390X is 1.195X
RX-480 reached 35.8 fps
R9-390X reached 42.7 fps
The frame rate between 35.8 fps and 42.7 fps difference is 1.192X
Note how effective memory bandwidth difference between RX-480 and R9-390X has matched the frame rate difference.
Since we establish that effective memory bandwidth impacts BattleField 1, let's predicate Scorpio GPU's frame rate results via 326 GB/s physical memory bandwidth.
Since PS4 Pro is showing RX-470D like results, I'll treat Scorpio the same.
For Scorpio: (326 GB/s x Polaris memory subsystem efficiency 77 percent) x Polaris 1.37X memory delta compression = 343.90 GB/s estimated effective memory bandwdith.
Effective memory bandwidth difference between RX-480 and Scorpio's estimated effective memory bandwidth is ~1.30X.
Apply 1.30X on RX-480's 35.8 fps = 46.6 fps estimate from memory bandwidth gain.
GTX 1070's frame rate is 46.6 fps LOL. This is just RX-480 gaining 326 GB/s memory bandwidth with Polaris delta memory compression bottleneck reduction.
For F1 2016, similar result.
Note why Scorpio's Forza example result landed in GTX 1070 range.
RX-480's 256 bit GDDR5-8000 memory setup is bottlenecking 5.8 TFLOPS.
You are underestimating memory bandwidth related bottlenecks. At 4K, R9-390X can still beat RX-480 OC and that's without any delta compression updates.
Math calculations
For Scorpio
100 percent / 66 percent of GPU usage for XBO graphics details +4K textures and 4K resolution = 1.51X head room
---
90 percent of XBO's 1.31 TFLOPS = 1.179 TFLOPS for 1920x1080p Forza 6 XBO result
For 4K resolution, 1.179 TFLOPS X 4X = 4.716 TFLOPS (GCN 1.1)
1.51X headroom x 4.716 TFLOPS (GCN 1.1) = 7.12116 TFLOPS scaled from XBO's TFLOPS (GCN 1.1).
IPC gain factor
7.12116 TFLOPS / 6 TFLOPS = 1.187X
RX-480 has 15 percent CU IPC gains which is nearly pointless since it's bottlenecked by 256 GB/s memory bandwidth.
ROPS's read and write is gimped by memory controller bottleneck i.e. it's not connected to L2 cache.
Polaris 10 is a half assed upgrade i.e. it's incomplete.
PS4 Pro's 4.2 TFLOPS GPU is bottlenecked by memory bandwidth which is worst for RX-480. Fp16 shaders are the alternative method to reduce memory bandwidth usage and switching to Fp16 shader code path departs from speeding up existing XBO 3D game engines.
Scorpio is like R9-390X at 7.12 TFLOPS with reduced memory bandwidth bottlenecks and designed to speed up existing XBO 3D game engines. Again, memory bandwidth is very important.
Notice how the same bum who predicted that this PoS was going to have a Ryzen cpu is immediately up in this thread pulling irrelevant charts and graphs out his ass.
Notice how the same bum who predicted that this PoS was going to have a Ryzen cpu is immediately up in this thread pulling irrelevant charts and graphs out his ass.
You are supporting shit arguments without any information backing. This argument "Get this shit out of here, maybe a 4 GB RX-480 or 3 GB GTX 1060 but that's about as far as it's going." comes out of his ass hole.
I wasn't fully committed with 8 core Ryzen due to the size of CPU cores e.g. I even spec'ed out quad core Ryzen and quad core Puma in the same style as ARM style Big.small 8 core setup. I even stressed the importance for CPU's rendering threads to be covered with sufficient processing resource and MS has alternative designs to reach similar goals e.g. new GCP and lower latency Puma.
I based my arguments from R9-390X (5.9 TFLOPS) results which is already far superior to RX 480 4GB at 4K.
R9-390X needs to combine with RX-480's improvements to obtain better estimation.
And Scorpio is GIMPED by CPU, put a Jaguar 8 cores paired with a 390X and tell me that you will get exactly the same performance as with $1,000 i7 or a $400 dollar one.
LINK me to where Scorpio is confirmed to have an i7 use on that benchmark and i admit scorpio will perform exactly as a 390X.
It has the same crappy Jaguar at 2.3ghz if the PS4 Pro was bottle neck by a 2.1 jaguar and the Pro is 4.2TF what the hell you think will happen when you use the same Jaguar 200mhz faster but this time with a 6TF GPU.?
Your common sense is horrible and you want to picture something here that will not happen just so you can put MS ahead.
Pascal doesn't have the same performance per Gflop as Polaris so comparing flops is useless.
The 1060GTX 3GB is 4.0 TF yet beat the 5.83TF RX480 in many games,so comparing flops between Nvidia and AMD is a joke we know this for generations and you will not rewrite history just to suit your biased arguments.
Oh did i mention the 1060GTX has 192GB/s bandwidth.? That is less than the RX480 256GB/s,again you are comparing different architectures and Scorpio is NOT VEGA is polaris,hell was FP16 confirmed for Scorpio because i didn't read it on DF article.?
Just because sony added FP16 doesn't mean MS also did,the PS4 has 8 Aces the xbox one doesn't,the xbox one has more move engines than the PS4.
They fallow different paths by the way.
Outside of fanboysm Ronvalencia i think you should rethink your arguments here,sure you can argue the RX480 has gimped bandwidth,but the same can be say about Scorpio CPU,and we know the same modification done to Scorpio CPU was done before to the XBO as well and even the 360 by MS own words the customization they did to the command processor is nothing new,so this is not some kind of secret sauce that would turn that Jaguar into a i5 or i7 by any means.
So comparing the 390X vs Scorpio using PC benchmarks using an i7 is totally wrong,if you at least had Ryzen you may argue something,but 31% improvement over the xbox one and even less over the Pro CPU wise prove it it is the same Jaguar with a stronger GPU.
Now you have to face the arguments you your self have done claiming the jaguar inside the Pro was a bottle neck 200mhz will produce on Pro probably 4 or 6 frames more,but it will not turn the Pro into something that is not,the same with scorpio i think when other multiplatforms games start hitting scorpio is where we will see some downgrade vs PC even compare to equal GPU,like it was the case with the XBO and PS4.
Scorpio is another unbalance piece of hardware,they increase GPU power more than 400% but increase CPU 31% is a joke that Phil Or any one call Scorpio a balance system,they have the memory,the bandwidth the GPU power but lack CPU power so is not balance.
And Scorpio is GIMPED by CPU, put a Jaguar 8 cores paired with a 390X and tell me that you will get exactly the same performance as with $1,000 i7 or a $400 dollar one.
LINK me to where Scorpio is confirmed to have an i7 use on that benchmark and i admit scorpio will perform exactly as a 390X.
It has the same crappy Jaguar at 2.3ghz if the PS4 Pro was bottle neck by a 2.1 jaguar and the Pro is 4.2TF what the hell you think will happen when you use the same Jaguar 200mhz faster but this time with a 6TF GPU.?
Your common sense is horrible and you want to picture something here that will not happen just so you can put MS ahead.
Pascal doesn't have the same performance per Gflop as Polaris so comparing flops is useless.
The 1060GTX 3GB is 4.0 TF yet beat the 5.83TF RX480 in many games,so comparing flops between Nvidia and AMD is a joke we know this for generations and you will not rewrite history just to suit your biased arguments.
Oh did i mention the 1060GTX has 192GB/s bandwidth.? That is less than the RX480 256GB/s,again you are comparing different architectures and Scorpio is NOT VEGA is polaris,hell was FP16 confirmed for Scorpio because i didn't read it on DF article.?
Just because sony added FP16 doesn't mean MS also did,the PS4 has 8 Aces the xbox one doesn't,the xbox one has more move engines than the PS4.
They fallow different paths by the way.
Outside of fanboysm Ronvalencia i think you should rethink your arguments here,sure you can argue the RX480 has gimped bandwidth,but the same can be say about Scorpio CPU,and we know the same modification done to Scorpio CPU was done before to the XBO as well and even the 360 by MS own words the customization they did to the command processor is nothing new,so this is not some kind of secret sauce that would turn that Jaguar into a i5 or i7 by any means.
So comparing the 390X vs Scorpio using PC benchmarks using an i7 is totally wrong,if you at least had Ryzen you may argue something,but 31% improvement over the xbox one and even less over the Pro CPU wise prove it it is the same Jaguar with a stronger GPU.
Now you have to face the arguments you your self have done claiming the jaguar inside the Pro was a bottle neck 200mhz will produce on Pro probably 4 or 6 frames more,but it will not turn the Pro into something that is not,the same with scorpio i think when other multiplatforms games start hitting scorpio is where we will see some downgrade vs PC even compare to equal GPU,like it was the case with the XBO and PS4.
Scorpio is another unbalance piece of hardware,they increase GPU power more than 400% but increase CPU 31% is a joke that Phil Or any one call Scorpio a balance system,they have the memory,the bandwidth the GPU power but lack CPU power so is not balance.
It is confirmed that Scorpio has a Polaris GPU, basically a Radeon r480. While Scorpio GPU has 40 compute units compared to stock r480 36, it operates at lower clock rate giving more or less the same performance. R480 is close to 1060 performance wise, being faster in some cases and slower in others, all depending on the game engine and DX12 optimizations. Still, in the end it is nowhere near 1070, a 4k entry card, performance wise. If MS decides to force "true 4k" requirement on the developers, as opposed to PS4 Pro checkerboard cheating, prepare for either low-to-medium graphics presets or 20-30 fps max.
To help the DSBR do its thing, AMD is fundamentally altering the availability of Vega's L2 cache to the pixel engine in its shader clusters. In past AMD architectures, memory accesses for textures and pixels were non-coherent operations, requiring lots of data movement for operations like rendering to a texture and then writing that texture out to pixels later in the rendering pipeline. AMD also says this incoherency raised major synchronization and driver-programming challenges.
To cure this headache, Vega's render back-ends now enjoy access to the chip's L2 cache in the same way that earlier stages in the pipeline do. This change allows more data to remain in the chip's L2 cache instead of being flushed out and brought back from main memory when it's needed again, and it's another improvement that can help deferred-rendering techniques.
Phil Spencer correctly identifies AMD GPU issues with it's high TFLOPS i.e. memory bandwidth.
When you talk to me about Scorpio, the term I use about the architecture isn’t the six teraflops which is obviously what we’ve announced, it’s balance. Really what it is, is you want a platform that is balanced between memory bandwidth, GPU power, you know, your ability to move memory and [an] amount of memory around in many ways is more inhibiting to the performance of your game than absolute teraflops on any one of the individual pieces, and when we designed Scorpio we really thought about this balanced rig that could come together at a price-point. Like, I want Scorpio to be at a console price-point, I’m not trying to go and compete with a high-end rig. And because we’re building one spec, we’re able to look at the balance between all the components and make sure that it’s something we really hit that matters to consumers and gamers.
The lesson from Nvidia Maxwell/Pascal designs are increasing memory bandwidth while increasing TFLOPS.
Notice both Vega and Phil Spencer's statement Scorpio's GPU memory handling targets memory movement issues.
RX-480 doesn't fix the old memory bandwidth bottleneck issues.
NVIDIA rubs it in on why AMD GPUs has less consistent performance.
Other factors at work is the "pack math" features.
Relax, Scorpio is not GTX 1080 and GTX 1080 Ti level GPUs.
For Battlefield 1 benchmarks, RX-480 (5.8 TFLOPS) and R9-290 (4.8 TFLOPS) has similar memory bandwidth hence similar frame rate result. This is an example for memory bandwidth bound issues. At least 1 TFLOPS gimped by RX-480's memory bandwidth.
R9-390X doesn't have delta memory compression, hence it's effective memory bandwidth is around 311 GB/s.
I'm recycling Beyond3D's effective memory bandwidth benchmarks to show you the effective memory bandwidth impacts on frame rate.
To workout Hawaii GCN's memory subsystem efficiency: R9-290X's 263 GB/s / 320 GB/s = 82 percent efficient.
R9-390X's 384 GB/s physical memory bandwidth would be around 315.6 GB/s effective memory bandwidth. Hawaii GCN doesn't have any memory delta compression i.e. this GPU design is very old.
Effective memory bandwidth difference between RX-480 and R9-390X is 1.195X
RX-480 reached 35.8 fps
R9-390X reached 42.7 fps
The frame rate between 35.8 fps and 42.7 fps difference is 1.192X
Note how effective memory bandwidth difference between RX-480 and R9-390X has matched the frame rate difference.
Since we establish that effective memory bandwidth impacts BattleField 1, let's predicate Scorpio GPU's frame rate results via 326 GB/s physical memory bandwidth.
Since PS4 Pro is showing RX-470D like results, I'll treat Scorpio the same.
For Scorpio: (326 GB/s x Polaris memory subsystem efficiency 77 percent) x Polaris 1.37X memory delta compression = 343.90 GB/s estimated effective memory bandwdith.
Effective memory bandwidth difference between RX-480 and Scorpio's estimated effective memory bandwidth is ~1.30X.
Apply 1.30X on RX-480's 35.8 fps = 46.6 fps estimate from memory bandwidth gain.
GTX 1070's frame rate is 46.6 fps LOL. This is just RX-480 gaining 326 GB/s memory bandwidth with Polaris delta memory compression bottleneck reduction.
For F1 2016, similar result.
Note why Scorpio's Forza example result landed in GTX 1070 range.
RX-480's 256 bit GDDR5-8000 memory setup is bottlenecking 5.8 TFLOPS.
If the cpu in the ps4 and xbone is a bottleneck, why would they use it in the first place knowing this information???
Because its not the same CPU.
Its Jaquar BASED, sure, but tons of modifications have been made, as well as half the workload removed (due to on board DX12 processing).
People are ignoring the facts because it fits their agenda.
PS4 Pro used the same CPU just overclocked, there is a difference.
Dont get me wrong, i dont think that Scorpio will be able to compete with a 1070 on a regular basis with every game, but the fact that it outperformed it PERIOD should be noteworthy.
"The crazy story here is that we've gone over our PC ultra settings and for everything that's GPU-related, we've been able to max it - and that's what we're running at, 88 per cent," says Tector, pointing to the utilisation data at the top of the screen. Right beneath it is the anti-aliasing setting - 4x, or rather 8:4x using the Radeon EQAA hardware AA."
Scorpio is locked at 60fps and still has about 22% left in the tank.
"It's impossible to keep all cars snapped into place as per the Scorpio demo, but the evidence seems to suggest we're looking at performance in the same ballpark as an Nvidia GTX 1070-class GPU here and even thatcan drop frames at ultra settings when wet weather hits."
If the cpu in the ps4 and xbone is a bottleneck, why would they use it in the first place knowing this information???
Because its not the same CPU.
Its Jaquar BASED, sure, but tons of modifications have been made, as well as half the workload removed (due to on board DX12 processing).
People are ignoring the facts because it fits their agenda.
PS4 Pro used the same CPU just overclocked, there is a difference.
Dont get me wrong, i dont think that Scorpio will be able to compete with a 1070 on a regular basis with every game, but the fact that it outperformed it PERIOD should be noteworthy.
No, it did not outperform it. We can't say that unless we have the same build of Forza running on a 1070.
Log in to comment