• 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.

X1 DDR3 RAM vs PS4 GDDR5 RAM: “Both Are Sufficient for Realistic Lighting”(Geomerics)

Spec-wise the Xbox One's GPU is around HD 7770 and the PS4's GPU is around HD 7850/7870.

And no one of those are hi-end desktop parts, much less by the end of the year. HD7770 was 100 bucks and are already discontinued.

You are attributing to the performance boost in those chips as a biproduct of eDRAM when in reality GT3e simply has more EUs. Do you think that doubling the EU cores has no performance boost? Another thing to note is you are comparing different product stacks, Iris Pro does not compete with Richland just as Bulldozer does not compete with Atom.

Yes, I'm doing that. The only difference between GT3 and GT3e is the embebed eDRAM. Both chips have 40 EUs. And you are the one who introduced Kaveri in this debate.

Also, I never said eSRAM and eDRAM won't have a positive affect on performance, I am saying the reason why eSRAM and eDRAM provide a performance boost is because they help make up for the crappy memory bandwidth of DDR3, not some latency-laden magical pixy dust you and Senjutssage keep talking about.

It's not enough, that's why ESRAM was implemented. Have you learned nothing? APUs are bandwidth starved especially the ones used in the PS4 and Xbox One.

Jesus, stop this already. This is embarrasing.

Btw, when I ask for a diagram, I ask for something like this:

SaturnVA12.png

I'm curious. Did you posted the wrong schematic because you are unable to undestand what it says, or because you wanted to cheat me? A quick look is enough to notice it was totally wrong.
 
Nope, you can't generate anything if you don't have a CPU. Tessellation can generate new geometry based on the already existing input sent by the CPU. But if you don't have the CPU generating the original vertices, you have nothing to start with to generate this new geometry.

The data is already in RAM which a GPU has access to. GPGPU allows GPU to perform transform operations and all the other kinds of OPs that a CPU can. The CPU will only extrapolate the 3D vertex data points based on the cartesian cordinates of the last or last few state(s). A modern GPU is more than capable.

The only thing that was missing was being able to setup a command list of instructions to tell the GPU's wide SIMD arrasy what to do with all the data. And guess what? GPU's on-die command processors can do just that.

I'm not saying a modern GPU won't ever need a CPU to feed it data. But modern GPUs don't necessarily require it to be so anymore.
 
How can something from 2008, as GDDR5 is, be next gen?

The amount of ram and the hUMA maybe?

Most PC's don't have as much ram allocated as the PS4 can. And no, I'm not saying PS4 has 8 gigs for VRAM.

Lets say 1gig for System, leaving 7GB for games. 4-5GB graphics and 2-3 for engine/audio/etc.
 

badb0y

Member
And no one of those are hi-end desktop parts, much less by the end of the year. HD7770 was 100 bucks and are already discontinued.

meet
What nonsenses are you talking now? Both GPU's are unable to compete with discrete mid-end offerings from both AMD/Nvidia. As convenient as they are for consoles, they lack the performance of desktop parts.

You said mid-ranged parts now you are changing them to high-end parts. The GPUs used in PS4 and Xbox One APUs are mid-range offerings of today.
Jesus, stop this already. This is embarrasing.

Btw, when I ask for a diagram, I ask for something like this:



I'm curious. Did you posted the wrong schematic because you are unable to undestand what it says, or because you wanted to cheat me? A quick look is enough to notice it was totally wrong.
The only thing embarrassing is how you cling on to low latency as if it's some sort of game-changer when multiple people have told you otherwise. Whether you want to believe it or not, the ESRAM was added because of the bandwidth contraints. As for the memory diagram I got it from VGLeak which Anand also referenced in his Xbox One write up, both are better and more qualified to speak about the matter than you. If you believe it is wrong I suggest you e-mail the authors of the articles.

In-fact VGLeaks specifically says that ESRAM bandwidth is shared between the GPU and the move engines...not sure why you would think that the CPU would benefit from the ESRAM.
 
meet


You said mid-ranged parts now you are changing them to high-end parts. The GPUs used in PS4 and Xbox One APUs are mid-range offerings of today.

No, they aren't. Those cards are entry level.

AMD's Entry Level Radeon HD 7790 Revealed
http://www.tomshardware.com/news/AMD-Radeon-7790-Saturn-GPU,21447.html

Mid range of today is GTX670~GTX680 (1344/1536 CUDA cores - 256bit bus).
Hi-end is Titan (2688 - 384 bits bus).

This is an official AMD slide:

amd-vs-nv.jpg


The only thing embarrassing is how you cling on to low latency as if it's some sort of game-changer when multiple people have told you otherwise. Whether you want to believe it or not, the ESRAM was added because of the bandwidth contraints. As for the memory diagram I got it from VGLeak which Anand also referenced in his Xbox One write up, both are better and more qualified to speak about the matter than you. If you believe it is wrong I suggest you e-mail the authors of the articles.

I'm doing what? xD

If you understood from my writings that my central issue is RAM latencies, your fanboyism is much greater than your technical understanding and knowledge.

In-fact VGLeaks specifically says that ESRAM bandwidth is shared between the GPU and the move engines...not sure why you would think that the CPU would benefit from the ESRAM.

Again, SHARED CACHE.
 

RoboPlato

I'd be in the dick
How can something from 2008, as GDDR5 is, be next gen?

8GB of unified high bandwidth memory on an APU set up is clearly a next gen spec. I don't see how you can argue otherwise. Combine that with the fact that the GPU is more powerful than a 7850 and is more efficient in a closed box setting with low level access and it's a huge step up from current gen. I'm not comparing to PC parts, I'm comparing to previous console generation performance jumps.
 

mrklaw

MrArseFace
No, they aren't. Those cards are entry level.


http://www.tomshardware.com/news/AMD-Radeon-7790-Saturn-GPU,21447.html

Mid range of today is GTX670~GTX680 (1344/1536 CUDA cores - 256bit bus).
Hi-end is Titan (2688 - 384 bits bus).

This is an official AMD slide:

amd-vs-nv.jpg




I'm doing what? xD

If you understood from my writings that my central issue is RAM latencies, your fanboyism is much greater than your technical understanding and knowledge.



Again, SHARED CACHE.

I'd say from that diagram that the 7850/7870 is mid-range. And thats roughly what is in PS4.

add in the additional GCN2 fine grained compute for more efficient use of GPU, general efficiency improvements from a closed box and lower level drivers, and I think PS4 should be able to handily outpace a mid-range GPU like that.
 

badb0y

Member
No, they aren't. Those cards are entry level.


http://www.tomshardware.com/news/AMD-Radeon-7790-Saturn-GPU,21447.html

Mid range of today is GTX670~GTX680 (1344/1536 CUDA cores - 256bit bus).
Hi-end is Titan (2688 - 384 bits bus).


This is an official AMD slide:

amd-vs-nv.jpg
Since when do mid-range parts cost $330-$500?

That chart shows that 7850/7870 is midranged....It's also a PR slide that's trying to skew the performance difference between GTX 680 and HD 7970 Ghz because they are within 5%-10% of each other.
I'm doing what? xD

If you understood from my writings that my central issue is RAM latencies, your fanboyism is much greater than your technical understanding and knowledge.



Again, SHARED CACHE.
Your central issue is the same as Senjutsusage's that Xbox One will somehow gain a sizable performance boost from having lower RAM latencies which is as dumb as it sounds. I don't need to be a fanboy to point at the truth and call out wrong posts.
I can't believe we're still arguing raw numbers. Holy fuck.

There is no argument over raw numbers anymore, just some people's hopes and dreams now.
 
The data is already in RAM which a GPU has access to.

GPGPU allows GPU to perform transform operations and all the other kinds of OPs that a CPU can. The CPU will only extrapolate the 3D vertex data points based on the cartesian cordinates of the last or last few state(s). A modern GPU is more than capable.

The only thing that was missing was being able to setup a command list of instructions to tell the GPU's wide SIMD arrasy what to do with all the data. And guess what? GPU's on-die command processors can do just that.

I'm not saying a modern GPU won't ever need a CPU to feed it data. But modern GPUs don't necessarily require it to be so anymore.

Nope. As I said, the GPU can't create the vertices.

http://cggmwww.csie.nctu.edu.tw/~danki/courses_tmp/course_notes/rtr2006/3_shaders.pdf

What vertex processor can not do
–It does not replace graphics operations for assembling the primitives.
•You can not ask vertex processor to change the primitive type or change the order of vertices form the primitives.
–No vertex can be created or deleted.
–Results of a vertex can not be passed to another vertex.


Yes, it is true that many tasks can be done on the GPU, but there are tasks you wound not want on the GPU, such as the ones that are not easily parallelizable. Thus, my point is, even if GPUs now are not as much dependent on the CPU as in the past, games can still benefit a lot from a good CPU.
 

astraycat

Member

Technically, with geometry shaders you can.

You still need the CPU to kick off a command buffer with all the setup for running a geometry shader though.

Also the typical use-case is geometry is generated offline (by artists) and then loaded into RAM at runtime. The CPU may do some SIMD math, but this will be only on a very small subset of data that will be used by the GPU. Uniforms like matrices for the vertex shader to consume.
 

Lord Error

Insane For Sony
No, they aren't. Those cards are entry level.
They're not entry level. I'm sorry, but $400-500 card can't be mid-range no matter which way you slice it. $300 is not mid range either. 7860, which would be the equivalent of what's in PS4 is actual mid-range by today's standards. The gpu that's in new Xbox would probably be mid-low.

Titan and 7990 are equivalent of SLI configuration in it pricing and execution, and they don't belong to a typical low-mid-high classification.
 

ElTorro

I wanted to dominate the living room. Then I took an ESRAM in the knee.
Nope. As I said, the GPU can't create the vertices.

This is a statement about vertex processing units, not about entire GPUs. In addition, the slides are pretty outdated and still assume a graphics pipeline implementation with discrete and independent vertex and pixel shader units.

A more up-to-date overview is this:

IC340510.jpg


In particular, the tesselation stage can take vertices and break them down into a new set of sub-vertices or discard them.

However, that does not mean that the GPU does not depend on the CPU to provide it with the initial list of vertices. After all, the list of relevant vertices must be computed by the CPU and depends on how the game world is geometrically represented, how it changes according to the game world simulation, and how relevant parts of the game world are selected for rendering.
 
Technically, with geometry shaders you can.

You still need the CPU to kick off a command buffer with all the setup for running a geometry shader though.

Also the typical use-case is geometry is generated offline (by artists) and then loaded into RAM at runtime. The CPU may do some SIMD math, but this will be only on a very small subset of data that will be used by the GPU. Uniforms like matrices for the vertex shader to consume.

Yes, it's kinda the same thing you do with tessellation I think; however, you always need the original vertices generated by the CPU.

http://en.wikipedia.org/wiki/Shader

Geometry shaders are a relatively new type of shader, introduced in Direct3D 10 and OpenGL 3.2; formerly available in OpenGL 2.0+ with the use of extensions.[2] This type of shader can generate new graphics primitives, such as points, lines, and triangles, from those primitives that were sent to the beginning of the graphics pipeline.[3]
 
How can something from 2008, as GDDR5 is, be next gen?

The fact that the memory is gddr5 doesnt make it next gen. Its all other aspects that make it forward looking. As one of the posters already mentioned there isn't an APU on the market today (they are coming soon) whereby the CPU and GPU have full access to a unified pool of gddr5 memory... At the same time. a la hUMA. nor is there any graphics card with 8gbs of gddr5 (this will change in the coming year or two). In addition to the amount memory and the ability of the CPU and GPU to access memory at the same time, you have to take into account that this is a closed system with a shader language, pssl, built specifically for the hardware and not for software or many many different combinations o hardware.
 
Since when do mid-range parts cost $330-$500?

Since AMD failed to deliver. GK104 (GTX670 and GTX680) is the Kepler mid end part, substitute of Fermi GF114 (GTX560ti). Titan GK110 is the successor of GTX580 GF110. Price positioning doesn't change that.

Please, stop talking about price when we are talking about technology. Parts prices are just like stock prices, there are a lot of external things influencing them. DDR2 is much more expensive than DDR3, does that make it better?

That chart shows that 7850/7870 is midranged....It's also a PR slide that's trying to skew the performance difference between GTX 680 and HD 7970 Ghz because they are within 5%-10% of each other.

First notice PS4 GPU will run at 1000mhz and 190W TDP. I'm not that interested on how AMD compares themselves to Nvidia, that on how they position every of their own cards. 7700 is the last on that chart, and 7850 is the next one. How can you call that Mid-end? Not even AMD or Tom's Hardware is.

Your central issue is the same as Senjutsusage's that Xbox One will somehow gain a sizable performance boost from having lower RAM latencies which is as dumb as it sounds. I don't need to be a fanboy to point at the truth and call out wrong posts.

argumentum ad logicam.
 

Pimpbaa

Member
Your central issue is the same as Senjutsusage's that Xbox One will somehow gain a sizable performance boost from having lower RAM latencies which is as dumb as it sounds. I don't need to be a fanboy to point at the truth and call out wrong posts.

Both of them sound more like fanboys than the people they say are fanboys.
 

astraycat

Member
Yes, it's kinda the same thing you do with tessellation I think; however, you always need the original vertices generate by the CPU.

http://en.wikipedia.org/wiki/Shader

You just need the initial kick.

For example, you could have a vertex shader that does nothing. Then in the geometry shader stage it generates 4 vertices for a full-screen quad.

On the CPU side you'd essentially say "draw me one point" and give it no data, then kick that off to the GPU. No initial vertex data needed. Using a geometry shader in this case is kind of wasteful though. If you needed a full-screen quad you could say instead "draw me a quad" with no data then switch based on the Vertex ID to position to fill in the actual data.

But a again, it's very rare that the CPU generates a significant amount of data for a frame. The data tends to already be available, the CPU just loads it and sets the corresponding configuration data to draw.
 

Lord Error

Insane For Sony
First notice PS4 GPU will run at 1000mhz and 190W TDP. I'm not that interested on how AMD compares themselves to Nvidia, that on how they position every of their own cards. 7700 is the last on that chart, and 7850 is the next one. How can you call that Mid-end? Not even AMD or Tom's Hardware is.
Are you looking at the same chart? 7750 is last, then 7770, then 7850, and PS4 GPU is somewhat better than that, it's like imaginary 7860.
Also, noone sane is going to just ignore prices of these things when determining what's mid-end. I'm sure Nvidia and AMD would love that though, so if they release some quad GPU board for $2000, then that's new 'high end'.
 

badb0y

Member
Since AMD failed to deliver. GK104 (GTX670 and GTX680) is the Kepler mid end part, substitute of Fermi GF114 (GTX560ti). Titan GK110 is the successor of GTX580 GF110. Price positioning doesn't change that.

Please, stop talking about price when we are talking about technology. Parts prices are just like stock prices, there are a lot of external things influencing them. DDR2 is much more expensive than DDR3, does that make it better?



First notice PS4 GPU will run at 1000mhz and 190W TDP. I'm not that interested on how AMD compares themselves to Nvidia, that on how they position every of their own cards. 7700 is the last on that chart, and 7850 is the next one. How can you call that Mid-end? Not even AMD or Tom's Hardware is.



argumentum ad logicam.
I am now typing from my phone so please excuse any typos.

As far as code names are concerned you are absolutely right but that doesn't change the fact that AMD and nVidia positioned the GTX 680 and HD 7970 GHz as their high end. The GTX Titan is fairly new and at $1000+ it's a halo part that only hardcore enthusiasts will buy.

As for HD 7790 it's a mainstream part but the HD 7800 based parts are definitely mid-range.

You still have yet to post any information on how lower latency will improve GPU performance.
 
Both of them sound more like fanboys than the people they say are fanboys.

Oh, thanks for your input. Glad to know I sound as a Fanboy to people without the basic technichal knowledge.

Once of the first things I said here, is that I'm not concerned about DDR3 being a bandwidth bottleneck to One GPU. In fact, I said the same for WiiU several months ago. People here is like OMG DDR3 TOO SLOW without understanding how memory works. Yes, DDR3 at 1600mhz may mildly hurt APU performance, and many people overcome that overclocking DDR3:

3DMark11Entry.png

AvePerfDiff.png


As we can see, some performance is gained from massive memory overclocking. So it's safe to assume that desktop APUs are, to some extent, bandwidth starved.

But now move to the quid:

Those AMD desktop parts have a 128 bits shared BUS, coming from usual DDR3 dual channel setup from PC.
Xbox One have a 256 bits bus, not counting aggregated busses.

Those APUs have around 29.8GB/s with DDR3@1866mhz.
Xbox One DDR3@2133 256 bit provides 68Gb/s.

Answer me, please, I beg you, how can be more than double bandwidth from regular desktop APUs a problem for One's GPU? Why GAF can't understand that one thing is memory chips and other thing is memory bus or interfaces? Why everyone compare DDR3 performance without taking into account BUS width and memory setup? Why am I trying to talk about buses and memory subsystems and in return I have a retarded guy nagging about chips latencies and quoting me to reply to another poster (Sage)?

Geez!
 

Biker19

Banned
It's simply not enough especially when your primary competitor has more. It would be extremely silly for you to base your statement on the scope of current games today. The scope of games will exponentially increase with this new generation.

This. 7 GB's of GDDR5 RAM used for games, along with a better GPU will do far more damage to a console with 5 GB's of DDR3 RAM being used for games along with 32 MB's of eSRAM & with a weaker GPU than the GPU inside PS4 when it comes to future multiplat games.

I think that'll matter in the long run when it comes down to the multiplats with better performance &/or better graphics. Maybe not now, but it'll definitely happen later on down the road.
 
a
Oh, thanks for your input. Glad to know I sound as a Fanboy to people without the basic technichal knowledge.

Once of the first things I said here, is that I'm not concerned about DDR3 being a bandwidth bottleneck to One GPU. In fact, I said the same for WiiU several months ago. People here is like OMG DDR3 TOO SLOW without understanding how memory works. Yes, DDR3 at 1600mhz may mildly hurt APU performance, and many people overcome that overclocking DDR3:

3DMark11Entry.png

AvePerfDiff.png


As we can see, some performance is gained from massive memory overclocking. So it's safe to assume that desktop APUs are, to some extent, bandwidth starved.

But now move to the quid:

Those AMD desktop parts have a 128 bits shared BUS, coming from usual DDR3 dual channel setup from PC.
Xbox One have a 256 bits bus, not counting aggregated busses.

Those APUs have around 29.8GB/s with DDR3@1866mhz.
Xbox One DDR3@2133 256 bit provides 68Gb/s.

Answer me, please, I beg you, how can be more than double bandwidth from regular desktop APUs a problem for One's GPU? Why GAF can't understand that one thing is memory chips and other thing is memory bus or interfaces? Why everyone compare DDR3 performance without taking into account BUS width and memory setup? Why am I trying to talk about buses and memory subsystems and in return I have a retarded guy nagging about chips latencies and quoting me to reply to another poster (Sage)?

Geez!
Because the Xbox One GPU has far more CUs which require significantly more bandwidth to be feed to them???

Edit: What APU is that?

This. 7 GB's of GDDR5 RAM used for games, along with a better GPU will do far more damage to a console with 5 GB's of DDR3 RAM being used for games along with 32 MB's of eSRAM & with a weaker GPU than the GPU inside PS4 when it comes to future multiplat games.

I think that'll matter in the long run when it comes down to the multiplats with better performance &/or better graphics. Maybe not now, but it'll definitely happen later on down the road.
I can't see Sony partitioning a large amount for the OS, but until we have a more concrete rumor/leak about the amount they're using for the OS, I think its best not to assume for now.
 
Are you looking at the same chart? 7750 is last, then 7770, then 7850, and PS4 GPU is somewhat better than that, it's like imaginary 7860.
Also, noone sane is going to just ignore prices of these things when determining what's mid-end. I'm sure Nvidia and AMD would love that though, so if they release some quad GPU board for $2000, then that's new 'high end'.

Ok, lets rewrite market positioning:

Über ethusiast:

Titan Quad Sli

enthusiast

Titan

Hi-end

GTX680
GTX670
HD7970

Mid-range

HD7850
GTX660
HD7700

Low-end

Radeon HD 7750
Radeon HD 7670
Radeon HD 7570
Radeon HD 7470
Radeon HD 7450
Radeon HD 7350
GeForce GTX 650 Ti Boost
GeForce GTX 650 Ti
GeForce GTX 650
GeForce GTX 645
GeForce GT 640 Rev. 2
GeForce GT 640 (GDDR5)
GeForce GT 640 (DDR3)
GeForce GT 635
GeForce GT 630 (GDDR5)
GeForce GT 630 Rev. 2
GeForce GT 630
GeForce GT 625
GeForce GT 620
GeForce GT 610
GeForce 605

Better this way to accomodate HD7770 into mid-tier?
 

badb0y

Member
Oh, thanks for your input. Glad to know I sound as a Fanboy to people without the basic technichal knowledge.

Once of the first things I said here, is that I'm not concerned about DDR3 being a bandwidth bottleneck to One GPU. In fact, I said the same for WiiU several months ago. People here is like OMG DDR3 TOO SLOW without understanding how memory works. Yes, DDR3 at 1600mhz may mildly hurt APU performance, and many people overcome that overclocking DDR3:

3DMark11Entry.png

AvePerfDiff.png


As we can see, some performance is gained from massive memory overclocking. So it's safe to assume that desktop APUs are, to some extent, bandwidth starved.

But now move to the quid:

Those AMD desktop parts have a 128 bits shared BUS, coming from usual DDR3 dual channel setup from PC.
Xbox One have a 256 bits bus, not counting aggregated busses.

Those APUs have around 29.8GB/s with DDR3@1866mhz.
Xbox One DDR3@2133 256 bit provides 68Gb/s.

Answer me, please, I beg you, how can be more than double bandwidth from regular desktop APUs a problem for One's GPU? Why GAF can't understand that one thing is memory chips and other thing is memory bus or interfaces? Why everyone compare DDR3 performance without taking into account BUS width and memory setup? Why am I trying to talk about buses and memory subsystems and in return I have a retarded guy nagging about chips latencies and quoting me to reply to another poster (Sage)?

Geez!
First, the iGPUs used are much better than what is available today. No desktop APU right now has anything close to a HD 7850 in it. So obviously a better iGPU requires a more bandwidth as demonstrated by your charts from Tom's Hardware.

Second, If I understand you correctly you are saying that Xbox One has enough bandwidth to support both the GPU and the CPU so why would Microsoft spend die space on ESRAM instead of more CUs or a better CPU? It makes no sense.
 

AgentP

Thinks mods influence posters politics. Promoted to QAnon Editor.
Answer me, please, I beg you, how can be more than double bandwidth from regular desktop APUs a problem for One's GPU? Why GAF can't understand that one thing is memory chips and other thing is memory bus or interfaces? Why everyone compare DDR3 performance without taking into account BUS width and memory setup? Why am I trying to talk about buses and memory subsystems and in return I have a retarded guy nagging about chips latencies and quoting me to reply to another poster (Sage)?

Geez!

The bandwidth a GPU needs is a function oif how it much it can process. The highest end video cards have 386-bit 200GB/s+ bandwidth just for the GPU. The XB1 has 68GB/s to share with the CPU. The PS4 has 18 CUs instead of 12 and hence it has 176GB/s bandwidth to share.

Using a low end laptop APU as a model is not very convincing. Those parts are very low end. But you are right, saying DDR3 is not important, it is the actual bandwidth that is, but even quad channel 256-bit DDR-2133 is not enough to serve both CPU and 12CU GPU, hence the 32MB eSRAM.

Erlier people were discussing Piledriver vs Jaguar. I just wanted to note that Tom's Hardware has benchmarks for both parts in two different reviews, but the same tests. Per Watt, the Jaguar did indeed seeem to be much better, depending on the benchmark it was 1.5x-4x better than the Piledriver score. Even using per clock cycle it was still better. I assume this is wh Sony and MS used the part, so they can spend their TDP allowance where it is needed, on graphics.
 

Nachtmaer

Member
I can't see Sony partitioning a large amount for the OS, but until we have a more concrete rumor/leak about the amount they're using for the OS, I think its best not to assume for now.

Didn't the 1GB come from the original 512MB OS footprint rumor back when the PS4 was going to have 4GB RAM in total and people just assumed it doubled as well when they switched to 8GB? At least that's how I remember it.
 
The bandwidth a GPU needs is a function oif how it much it can process. The highest end video cards have 386-bit 200GB/s+ bandwidth just for the GPU. The XB1 has 68GB/s to share with the CPU. The PS4 has 18 CUs instead of 12 and hence it has 176GB/s bandwidth to share.

Highest end video cards are ready to output über-resolutions such as 5760x1080p in multiple GPU's configurations. I don't believe consoles will go over 1080p.

Using a low end laptop APU as a model is not very convincing. Those parts are very low end. But you are right, saying DDR3 is not important, it is the actual bandwidth that is, but even quad channel 256-bit DDR-2133 is not enough to serve both CPU and 12CU GPU, hence the 32MB eSRAM.

Wich low end laptop APU? I was doing calcs using max rated DDR3 speed for current FM2 desktop APU, HD8670D. You can slightly increase that OCing the memory to 2133mhz, but wont magically double throughput.

Erlier people were discussing Piledriver vs Jaguar. I just wanted to note that Tom's Hardware has benchmarks for both parts in two different reviews, but the same tests. Per Watt, the Jaguar did indeed seeem to be much better, depending on the benchmark it was 1.5x-4x better than the Piledriver score. Even using per clock cycle it was still better. I assume this is wh Sony and MS used the part, so they can spend their TDP allowance where it is needed, on graphics.

You are confusing raw power with power efficiency. Obviously, Jaguar cores are more efficient thermal wise than Piledriver, but later can reach much higher frecuencies and are able to scale better adding more cores. They also include more functions and instructions to increase performance. Problem is Piledriver is too power hungry for consoles thermal goals.
 

LukeTim

Member
Not a driver. Console has very thin APIs (application programming interfaces). The purpose is to allow programmers to write software that focusses on what the software wants to do and not have to worry about micromanaging the hw, i.e. moving data in and out of the processor registers etc. APIs provide a layer of abstraction to make programming on a given piece of hw easier.

Drivers on the other hand are a greater layer of abstraction that lets software be intepreted to run on many different hw configurations., e.g. Nvidia drivers allows games to work many different GPUs across the range of products in a series.

Consoles would use drivers for some thing like peripherals I would imagine. But those would be far more simple than same GPU drivers.

As a side note however it has been confirmed that the XB1 runs all its software in an effective virtual machine. Basically meaning the hw the developers see when they're coding their games is an entirely virtualised environment. As far as I'm aware they have no direct access to the hw, however on the other hand the game VM layer has been optimised for maximum performance, with all if not most GPU features exposed in the VM. I imagine MS did this for security, given how easy x86 hw would be to hack (i assume).

That's not strictly true...

A device driver is any piece of software which drives hardware directly... makes it work, makes it do something. By which definition, a console necessarily requires drivers to function.

A driver is a driver regardless of whether it needs to support many variants or one single variant of a piece of hardware.
 
The amount of ram and the hUMA maybe?

Most PC's don't have as much ram allocated as the PS4 can. And no, I'm not saying PS4 has 8 gigs for VRAM.

Lets say 1gig for System, leaving 7GB for games. 4-5GB graphics and 2-3 for engine/audio/etc.

The amount yeah dont know what so special about it.

How can hUMA be specials when console through the years already used an unified memory system. You know like the 360 did and probably other before it.
 
That's not strictly true...

A device driver is any piece of software which drives hardware directly... makes it work, makes it do something. By which definition, a console necessarily requires drivers to function.

A driver is a driver regardless of whether it needs to support many variants or one single variant of a piece of hardware.

No I agree with you. But I was really speaking of drivers in the context I believed the poster was intending, I.e. GPU drivers.
 

LukeTim

Member
No I agree with you. But I was really speaking of drivers in the context I believed the poster was intending, I.e. GPU drivers.

Okay, I see what you mean.

One would hope that the OS architecture on consoles is so specifically and well designed that drivers would be less bloated and much more efficient than on PCs.
 
Okay, I see what you mean.

One would hope that the OS architecture on consoles is so specifically and well designed that drivers would be less bloated and much more efficient than on PCs.

The only difference is that console "drivers" have less error checking because they run on a fixed platform. Nothing to do with the OS.

With the "new" consoles being basically PCs in architecture, I don't think we can hope for amazing performance compared to the bigger PCs. Also those CPUs... wow...
 
People mix this up very easily:

UMA can mean Unified Memory Architecture as well as Uniform Memory Access. Xbox360 only had the former, PS4 will have both.

Can you clarify it a bit more or post an link i could have read it wrong on google.
Xbox 360 uses unified memory and uniform memory access is an method too implement unified memory architecture. hUMA as far as i just read via google means that its also cache coherent so data is always up to date in memory.

This should simplify usage of pointers in cpu code and gpgpu kernel/shader code and concurrent programming if im not mistaken .
 
Unified Memory Architecture (or Shared Memory Architecture) just means that you have a single RAM pool for two different processors, just like in the Xbox360 which has 512MB GDDR3 for both CPU and GPU.

To understand AMD's (heterogeneous) Uniform Memory Access I recommend this Ars Technica article:

http://arstechnica.com/information-...orm-memory-access-coming-this-year-in-kaveri/

It's a great read!

Read it and tbh its sounds exactly the same as unified memory architecture only AMD is throwing marketing terms against it to make it sounds more for their APU then it really is.
My internship boss(dont know the word) was kinda hyped for hUMA and APU in general because it could speed up image processing in camera which had an APU built in.

Maybe i miss the complete picture but the wikipedia pages clearly states uniform memory access is one of the ways to implement an uniform memory access.
Nvidia is coming with an virtual solution where both memory pools can be seen as one pool but you still have the pcie bus as an potential bottleneck.
 
Isn't the difference between UMA and hUMA just simple cache-coherency? So instead of just reading/writing to the same memory pool you can actually let the CPU/GPU work together because both now "know" what the other is doing? I imagine this will take a while to take off and with UMA the biggest resource hog - the copying between CPU/GPU is gone I am not even sure that it is as revolutionary as AMD makes us think. However I am indeed curious how things will turn out.
 

AgentP

Thinks mods influence posters politics. Promoted to QAnon Editor.
Highest end video cards are ready to output über-resolutions such as 5760x1080p in multiple GPU's configurations. I don't believe consoles will go over 1080p.

They all support the same output resolutions, just the low end cards can't pump enough pixels to get high frame rates. My GTX 670 is driving my 1080P TV, the horsepower goes into high resolution textures, more AA, etc. Not just output resolution. The same is true for the XB1 and PS4, the more power they can provides (assuming the memory can feed it), the better the image will look or the better the frame rate will be.

You can find discrete video cards that have ~50GB/s bandwidth, you will see that the GPU paired with that bandwidth is low end and the card cannot keep up with the big boys even at 1080P. The same is true for the XB1, without the 32MB eSRAM the GPU would starve since it will probably see 40-50GB/s max left over from IO and CPU.
 
Well let's try it this way:

Your CPU speaks language A and your GPU speaks language B. That means even if you have a shared memory architecture like in Xbox360 a GPU does not understand data with language A. Even if these RAM chips are physically the same pool, the system still distinguishes between data pool A with CPU stuff and data pool B with GPU stuff. If the GPU needs some data from pool A then you have to copy it into pool B (even if it's the same RAM chips like on Xbox!).

The advantage of an Unified Memory Architecture is that you don't have to send your data over a PCIe-bus (like in a PC), but you still need to do the copy work which still will give you a penalty on latency. That's not bad for rendering since GPUs are designed to hide latency (CPU -> GPU -> monitor = good!) but for computing you'll have to send that data back to CPU and a CPU is extremely latency-sensitive (CPU -> GPU -> CPU = bad!).

AMD's heterogeneous Unified Memory Access tries to eliminate this copy-overhead by teaching the processors to speak both language A and B. That means both CPU and GPU can work on every piece of data in the RAM without having to copy it into a dedicated pool for language A or B. This reduces the latency and eventually CPU and GPU can work in concert on tasks.

Sorry but it makes absolutely no sense too me there is a reasons IEEE has standards for Floating point and integer formats. Memory has nothing to do with language A and B it needs to be read from and written too.

If what you say is true then this 360 bandwidth picture on wikipedia makes absolutely no sense.
Hell even the 360 has something like onion bus the ps4 has (the 10.8GB/s bus)?

File:X360bandwidthdiagram.jpg
 
Sorry but it makes absolutely no sense too me there is a reasons IEEE has standards for Floating point and integer formats. Memory has nothing to do with language A and B it needs to be read from and written too.

If what you say is true then this 360 bandwidth picture on wikipedia makes absolutely no sense.
Hell even the 360 has something like onion bus the ps4 has (the 10.8GB/s bus)?

File:X360bandwidthdiagram.jpg

The subject is that with hUMA there is a unified memory address, so that both GPU and CPU know in which part of the memory are stored the data of each other and can access them directly.
 

tfur

Member
Never forget dr. apocalipsis in the 8 GB GDDR5 thread....

atPhDee.jpg



A lot of people will be dissapointed once the box arrive and won't be able to match a GTX560ti with 1GB GDDR5.

You need processing power to do Volumetric shadows, not infinite memory pool. What is a '3D texture' to begin with? Developers will be happy once they wont need to optimize their bloated stuff to fit in, gamers will have shorter load times, because there wont be any better use for so much RAM in a so weak system.

In fact, it's most than probably than raw performance will be lowered down a bit, since they will have to loose timmings on the IMC for it to be able to hold double RAM.

There is a lot of people making laugh about the 8GB GDDR5 Meme, and people actually believing this is some sort of advanced stuff. Going from 4GB to 8GB in the same PCB is as easy as double sided PCB for memory chips. Later on, as chips will double the density, Sony will be able to halve the chip count for the cheaper revision of the board.

8GB is not good for gamers. Memory is boring stuff you need to feed the beasts. We, graphic whores, need more shader units, more ROP'S and TMU's, more speed and fillrate, and a better core CPU, not that crappy Jaguar stuff that our mommies use in their shitty Acer netbooks.

8GB GDDR5 is Sony being Sony after pay for preexisting tech instead of engineer it.
 
And I still believing the same.

Give me a Core i5 with a beefier GPU and 4GB of GDDR5 over crappy cpu and mid range GPU with 8GB GGDR5 any day of the day.

With those consoles you will fall short of power much earlier than run out of bandwidth.
 

Nafai1123

Banned
And I still believing the same.

Give me a Core i5 with a beefier GPU and 4GB of GDDR5 over crappy cpu and mid range GPU with 8GB GGDR5 any day of the day.

With those consoles you will fall short of power much earlier than run out of bandwidth.

Please explain how you would design a better system with the same TDP.
 
Top Bottom