PlayStation 4 (codename Orbis) technical hardware investigation (news and rumours)

[patsu]

Hmm... I suppose we will need to know what goes into that 1GB. e.g., If a game use up 7GB, can we watch NetFlix and RemotePlay the game at the same time ? I am assuming non-games will run outside the OS reserved RAM. The OS will suspend the game and apps when they are swapped out of view.

Authorized background jobs can probably be dispatched/delegated to the secondary CPU.

that website is well known for posting fake rumors... one of their articles even says that PS3 GPU pushes 400Gflops...

Thanks for the heads up !

 

[MrFox]

If devkits cannot physically have more than 8GB, should we assume the games won't be able to use the full memory until new devkits are available?
How did it work for the 360 when both devkits and retail had the same amount of memory?

 

[dagamer]

If devkits cannot physically have more than 8GB, should we assume the games won't be able to use the full memory until new devkits are available?
How did it work for the 360 when both devkits and retail had the same amount of memory?

I don't think launch games used it to the fullest extent.

 

[patsu]

I thought there were rumors about 16GB devkits ? It was our first indication that PS4 may have more than 4GB RAM at that time (although no one took it seriously). Perhaps it was configured differently.

As for using up memory, it may be quite easy to use up all the available memory in a sloppy manner. That's probably not what we want.

 

[marcberry]

This is from the HD7790 but the math concept will work on all AMD GCN chip sets or GPUs of a APU and yes XBOX AND PS4.
NOTE: Do not use Line(3) for XBOX AND PS4 not the same SIMDS.

1.Line(2) Divide 896 by 14 = 64 Threads
2.Line(2) 14 Compute Units x 64 Threads = 896 Stream Processors
3.Line(2) 896 Stream Processors x Line(1)1000MHz = 896000
4.Use 896000 x line(3) 2 = 1.79TFsp
http://www.amd.com/us/products/desktop/graphics/7000/7790/Pages/radeon-7790.aspx#3

(1)1000MHz Engine Clock
1GB GDDR5 Memory
1500MHz Memory Clock (6.0 Gbps GDDR5)
96GB/s memory bandwidth (maximum)
1.79 TFLOPS Single Precision compute power
GCN Architecture

(2) 14 Compute Units (896 Stream Processors) or SIMDs
56 Texture Units
64 Z/Stencil ROP Units
16 Color ROP Units
(3) Dual Geometry Engines

Last one, the PS4: Vgleaks "-ALU(32 64-bit operations per cycle)"

4 SIMDs per SC or CU x 32 Threads = 128 Threads
18 SC or CU x 128 Threads = 2304 (Stream Processors) x
800MHz = 1.8TFsp same as SONY spec.


The MATH is undeniable !!!!!!

 

[milk]

As for using up memory, it may be quite easy to use up all the available memory in a sloppy manner. That's probably not what we want.

That will make development easier and more eficient during the beggining of the gen, leading to better games thanks to less dev time spent on getting things up and running, and more time on tweaking and polish. A couple years in and devs will already be optmising their memory usage and be wishing they had more.

 

[Quaz51]

If devkits cannot physically have more than 8GB, should we assume the games won't be able to use the full memory until new devkits are available?
How did it work for the 360 when both devkits and retail had the same amount of memory?

devtool can use OS footprint, OS footprint is lot of empty buffer or devkit can use special OS

 

[ERP]

If devkits cannot physically have more than 8GB, should we assume the games won't be able to use the full memory until new devkits are available?
How did it work for the 360 when both devkits and retail had the same amount of memory?

The extra memory in a devkit is a convenience in no way does it prevent you from using all of the memory. Just means you have to be more active in monitoring it. As opposed to trying to reduce the memory footprint after the fact.
360 devkits were stuck with 512MB until the retail board was redesigned over a year after launch.

Its also a little different this time with the "OS" rumored to be reserving such large amounts of memory, there is little reason why for development that overhead couldn't be reduced and given back to the game.

 

[patsu]

That will make development easier and more eficient during the beggining of the gen, leading to better games thanks to less dev time spent on getting things up and running, and more time on tweaking and polish. A couple years in and devs will already be optmising their memory usage and be wishing they had more.

I hope so ! In general, "better games" is more a question of budget, management, time, marketing, experiences, talent, so on and so forth. Technology can empower people, but it seldom guarantees (human) results. ^_^

I half expect companies that released buggy games last few gens to continue to release buggy games nextgen if all else stay the same.

 

[Pete]

This is from the HD7790 but the math concept will work on all AMD GCN chip sets or GPUs of a APU and yes XBOX AND PS4.
NOTE: Do not use Line(3) for XBOX AND PS4 not the same SIMDS.
[...]
2.Line(2) 14 Compute Units x 64 Threads = 896 Stream Processors
[...]
One more, the XBOX NEXT: Vgleaks " A SIMD executes a vector instruction on 64 threads at once in lockstep "

4 SIMDs per SC or CU x 64 Threads = 256 Threads
[...]
The MATH is undeniable !!!!!!

Welcome to B3D. Let me be the nth to deny your math (as I'm pretty sure it's already been discussed here).

First, PS4 and Xbox Next both use GCN architecture GPUs, thus both will likely have the GCN characteristic 1 CU = 4 SIMD-16.

Second, your math is faulty because your labels are faulty.
1 CU * (4 SIMD-16 / CU) * (16 MADDs / SIMD-16) * (2 FLOP / MADD) = 128 FLOP / CU.

You start with the correct terms ("14 CUs * 64 threads"/CU), then stumble once you introduce VGLeaks' mistake (a CU, not a SIMD, works on 64 threads--you just said so in one of your many "Line(2)"s!). If you'd actually labelled your work, you'd see where you went wrong:

(4 SIMD / CU) * (64 threads / CU) = 256 SIMD threads / CU^2 = = nonsense.

What you're saying is that PS4 doubles and XBN quadruples GCN's SIMD width, which probably isn't the case.

If you're trying to make a saving throw to show that XBN is more powerful than PS4, please note that "versus" posts are verboten in B3D unless (1) the comparison is technically relevant and not an attempt at "my fantasy number is bigger than yours," and (2) you've got the technical knowledge to back up your numbers. And save some exclamation points for the rest of us, please.

 

[tunafish]

I thought there were rumors about 16GB devkits ? It was our first indication that PS4 may have more than 4GB RAM at that time (although no one took it seriously). Perhaps it was configured differently.

It's not possible to provide more than 8GB of ram on 256-bit GDDR5 at the moment.

 

[MrFox]

It's not possible to provide more than 8GB of ram on 256-bit GDDR5 at the moment.

I remember the rumors started with an 8GB standard PC plus a 2GB modified AMD card (then upgraded to 4GB). Much later, a single leak said devkits were "somewhere between 8GB and 16GB".
I'm thinking that the leak's miscommunication was because they were counting the amount of memory in the whole devkit. So, including the host PC's main memory adds up to between 8 and 16.

 

[patsu]

Yap ! I thought it could be a Gaikai dev server (for developing/testing streaming PS4 demoes/games), assuming rumor was true.

 

[Grall]

It's not possible to provide more than 8GB of ram on 256-bit GDDR5 at the moment.

Can't GDDR5 support more than one memory device per channel? ...Excluding clamshell mode of course, as that's two devices sharing half a channel each. Some graphics cards in the past with extra-large onboard framebuffers have had two RAM devices per channel, but I'm unaware if any of those ever used GDDR5 RAM...

 

[tunafish]

Can't GDDR5 support more than one memory device per channel? ...Excluding clamshell mode of course, as that's two devices sharing half a channel each. Some graphics cards in the past with extra-large onboard framebuffers have had two RAM devices per channel, but I'm unaware if any of those ever used GDDR5 RAM...

No. Most of the reason why GDDR5 can reach such high interface clocks is precisely that the data interface is point-to-point.

 

[Ethatron]

The likely need to bin the Orbis APU may be a reason to have a dedicated low-power core.

Who would they sell the weaker chips to I wonder?

 

[MrFox]

Who would they sell the weaker chips to I wonder?

The lowest bins could be overvolted and used to fill up the server racks for Gaikai, everything becomes superbly easy to cool when noise isn't an issue
But isn't 800MHz already a very low frequency for GCN cards? I would think they chose this frequency specifically to avoid the big loss from binning.

 

[marcberry]

Welcome to B3D. Let me be the nth to deny your math (as I'm pretty sure it's already been discussed here).

First, PS4 and Xbox Next both use GCN architecture GPUs, thus both will likely have the GCN characteristic 1 CU = 4 SIMD-16.

Second, your math is faulty because your labels are faulty.
1 CU * (4 SIMD-16 / CU) * (16 MADDs / SIMD-16) * (2 FLOP / MADD) = 128 FLOP / CU.

You start with the correct terms ("14 CUs * 64 threads"/CU), then stumble once you introduce VGLeaks' mistake (a CU, not a SIMD, works on 64 threads--you just said so in one of your many "Line(2)"s!). If you'd actually labelled your work, you'd see where you went wrong:

(4 SIMD / CU) * (64 threads / CU) = 256 SIMD threads / CU^2 = = nonsense.

What you're saying is that PS4 doubles and XBN quadruples GCN's SIMD width, which probably isn't the case.

If you're trying to make a saving throw to show that XBN is more powerful than PS4, please note that "versus" posts are verboten in B3D unless (1) the comparison is technically relevant and not an attempt at "my fantasy number is bigger than yours," and (2) you've got the technical knowledge to back up your numbers. And save some exclamation points for the rest of us, please.

No!! Vgleaks: " A SIMD executes a vector instruction on 64 threads at once in lockstep " they use SC but it is the same as a CU.
No!! Vgleaks "-ALU(32 64-bit operations per cycle)" Here just for you
Try this WITH ANY AMD GCN chip set you wll get the same TF.

hd 7970
2048 / 32 = 64 Threads
32 Compute Units * 64 Threads = 2048 Stream Processors
2048 Stream Processors * 1050MHz =
2150400 * 2 = 4.3TFsp I can go on and on and on.

HD 7790
896 / 14 = 64 Threads
14 Compute Units * 64 Threads = 896 Stream Processors
896 Stream Processors * 1000MHz = 896000
Use 896000 * 2 = 1.79TFsp Hey that worked too.

Here is one .
4 SIMDs per SC or CU * 32 Threads = 128 Threads
18 SC or CU *128 Threads = 2304 (Stream Processors) *
800MHz = 1.8TFsp same as SONY spec.

4 SIMDs per SC or CU * 16 Threads (32 from VG) = 64 Threads
18 SC or CU * 64 Threads = 1152 (Stream Processors) *
800MHz * 2 = 1.843200TFsp same as SONY spec. You pick one!

How about this one.

4 SIMDs per SC or CU * 16 (not the pages & pages of data about SIMD-64wide) = 64 Threads
12 SC or CU * 64 Threads = 768 (Stream Processors) *
800MHz * 2 = 1.228800TFsp I think that worked too.

4 SIMDs per SC or CU * 16 (not the pages & pages of data about SIMD-64wide) = 64 Threads
12 SC or CU * 64 Threads = 768 (Stream Processors) *
800MHz * 2 = 1.228800TFsp I think that worked too.

It's not how you get their it's how you finish. It's finished!!!

 

[BRiT]

Your math is still faulty.

 

[ultragpu]

http://www.ps4blog.net/2013/04/12/ps4s-not-so-secret-weapons-for-longevity-ps4-apu/
So according to Mark Cerny in an interview all the launch titles are not even utilizing any of the APU architecture advantages nor are they writing the HSA API but will be fully tapped into later on.
I wonder how big of a difference will they make?

On PS4′s APU where the data in memory can be manipulated either by the GPU or CPU or in sequence without any performance hits due to data transfer, CPU can pre-process data using branch-heavy algorithms and hand it off to the GPU, or the GPU could pre-process data for the CPU. It’s like a super-efficient tag team. Mark Cerny, in an interview with the Japanese press, mentions that these features aren’t put to use in the launch titles, but will be something that opens up new possibilities later on. It is easy to understand, since current devkits have APU’s that do not have the latest features provided by the PS4 APU.

Whereas PS4 can be treated and coded for like an ordinary PC which will result in a performance to that of mid-to-high end PC’s, in the long run, developers will untap these resources allowed by the unique features of hardware, which will provide extra performance that will allow PS4 to prosper in the later years of its life-cycle. PS4 is not designed to win the brute force race (a race that is already lost to the PC), but it will offer so much more from its fixed set of hardware along the course of its life.

In the same interview mentioned earlier, Mark Cerny lays out a road-map for PS4′s longevity, along with meaty details about PS4′s architecture. The mid-term goals for increasing performance is opening the shader API’s for deeper access to hardware. So you won’t get these in the launch titles. The long term prospect is mature HSA API’s to get the most out the architecture. Mr. Cerny states that [implications] will be “huge” and that these will happen in the long run.