The Official PS4 Thread

[exdeath]

It's all arbitrary software lockout... Do not approve. Wonder if you'd even need custom logic or if a reflash of the firmware/id/etc of the device would work. Should be a serial eeprom in USB devices with the hardware and vendor ids and stuff.

 

[Pheran]

I wrote a piece on PS4 and Xbox One. I tried to stick to the facts, along with as fair an interpretation as I could. Some have said I was too easy on Microsoft, some think I'm too easy on Sony. Whatever, this is a start. Focused mainly on the systems and launch data.

http://losthammer.wordpress.com/201...13-and-beyond-predictions-and-what-to-expect/

Let me know what you think! I'm also looking for other articles and info to add as time goes on.

Nice article, however, this part seems wrong:

at only 32MB, it’s not even large enough to hold a full frame of a 1080p game in cache

I'm not sure what you mean - even at 32 bits/pixel, a 1080p frame is only around 8 MB.

 

[exdeath]

Nice article, however, this part seems wrong:


I'm not sure what you mean - even at 32 bits/pixel, a 1080p frame is only around 8 MB.

Yeah and at the very least you need 24 MB for a 1920x1080x32bpp double buffered + 32bpp depth/stencil buffer. That leaves 8 MB for texture cache and sub/super/multi sampling buffers and render to texture buffers and such needed for AA, DoF, stereoscopic 3D, etc, then vertex streams, etc. Suddenly we can see that 32 MB with 1080p is pushing it.

What makes the embedded SRAM terrible is it doesn't even reap the benefit of embedded on die SRAM: it's bandwidth isn't much more than the external RAM on the XBone and is still almost half as fast as the PS4s external RAM.

Even if the memory controller is split and it can access both the embedded and external ram simultaneously (eg: rendering threads have full access to esram without blocking access to external bus), the total bandwidth with both the esram and external ram together used concurrently is still slightly less than the PS4 external.

That in and of itself doesn't really mean much, but the extra memory management on the programmer of when to use what and what is mapped to where, all for no real benefit over a single easier to manage pool.

Lower latency won't be of any benefit either because the operations involved are deeply pipelined and overlapped, especially graphics.

Heh believe me if anyone can figure out a use to exploit the ESRAM it's me, but I'm having a hard time seeing any advantage to embedded RAM that isn't really much faster (+30%?) than the XBones external RAM let alone the PS4s external RAM. Usually when you think embedded ram to have something on the order of 1000s % faster than main memory, or to decouple from the bus to maximize DMA throughput in a multi processor environment to keep the bus free to maximize concurrency if bus traffic is at a premium (which it's not on PS4). It makes sense with something with slow RAM (Gamecube) or heavily bus driven multiprocessor (PS2 VUs, PS3 SPU scratchpad).

On XBone it will simply require mandatory use on part of the programmer to maximize memory throughput on the XBone in order to stay on equal footing with the PS4 which requires no special treatment of memory address space to achieve the same or better results. And that's ONLY if the memory controller can access the ESRAM and external RAM concurrently.

Really weird design decision. It certainly wasn't for a cost advantage with the $100 difference either.

 

[Pheran]

What makes the embedded SRAM terrible is it doesn't even reap the benefit of embedded on die SRAM: it's bandwidth isn't much more than the external RAM on the XBone and is still almost half as fast as the PS4s external RAM.

Even if the memory controller is split and it can access both the embedded and external ram simultaneously (eg: rendering threads have full access to esram without blocking access to external bus), the total bandwidth with both the esram and external ram together used concurrently is still slightly less than the PS4 external.

That in and of itself doesn't really mean much, but the extra memory management on the programmer of when to use what and what is mapped to where, all for no real benefit over a single easier to manage pool.

Don't get me wrong, I'm not saying the SRAM is great, I'm just objecting to the claim that a whole frame won't fit in it. The PS4 memory subsystem does appear to be much better.

 

[exdeath]

Don't get me wrong, I'm not saying the SRAM is great, I'm just objecting to the claim that a whole frame won't fit in it. The PS4 memory subsystem does appear to be much better.

Made my post a little better with more thought 😛

 

[KaOTiK]

I wrote a piece on PS4 and Xbox One. I tried to stick to the facts, along with as fair an interpretation as I could. Some have said I was too easy on Microsoft, some think I'm too easy on Sony. Whatever, this is a start. Focused mainly on the systems and launch data.

http://losthammer.wordpress.com/201...13-and-beyond-predictions-and-what-to-expect/

Let me know what you think! I'm also looking for other articles and info to add as time goes on.

Nice article, good job on staying neutral in it I'd say.

 

[Arkaign]

Nice article, however, this part seems wrong:



I'm not sure what you mean - even at 32 bits/pixel, a 1080p frame is only around 8 MB.

Ah, thanks, and thanks to ExDeath as well.

I think I need to amend the article to explain that better. I meant something in the vein of what ExDeath wrote, but wrote that aspect poorly without great clarification. It was something that I read that related to the embedded memory limitations not being enough for useful 1080p abilities.

After all, we don't have 32MB Video Cards anymore for a reason heheh 🙂

Nice article, good job on staying neutral in it I'd say.

Appreciated. I really am pretty happy with the DRM fix on the XB1. I obviously still don't think it's perfect by any means, but now at least I can see myself owning one down the road and not feeling dirty about it haha.

 

[KaOTiK]

Appreciated. I really am pretty happy with the DRM fix on the XB1. I obviously still don't think it's perfect by any means, but now at least I can see myself owning one down the road and not feeling dirty about it haha.

Yeah, I'm not 100% dead set against the console now, but I am far from buying one as well. As I said in another post about the 180 they pulled. They showed their hand on what they want (I went into some +/- with that not going to repeat atm) and with the interview with one of their exec's not being able to say they wouldn't go back to it in the future due to them being able to change their TOS whenever they want leaves me uneasy still. Best case is in a few years if things have truly turned around other stuff is addressed I could possibly buy one.

 

[Arkaign]

My friend Chris Tang just came in and added some more perspectives on it, pretty interesting stuff. He is a really amazing dude, Sega World Champion amongst a huge number of other titles, and has worked in the industry for two decades.

 

[exdeath]

Oh... just occurred to me.

Any memory performance gained by using the ESRAM over the external RAM is going to be negated by the DMA bandwidth used to shuffle the data from external RAM to ESRAM repeatedly in order to use it as high speed cache, vs the PS4 everything just stays and can be accessed "in place" with no penalty and still be faster than the ESRAM.

XBone:

Read from 60 GB/sec RAM -> write to 100 GB/sec ESRAM to load texture into high speed texture cache prior to rendering

Read from 100 GB/sec texture cache on texel lookup -> write to 100 GB/sec ESRAM frame buffer write

PS4:

Read from 176 GB/sec texture cache on texel lookup -> write to 176 GB/sec frame buffer.


The only real benefit of ESRAM is latency but again, pipelining: things like texture prefetching for example, you initiate DMA to load the next texture at the same time you start using the current texture for it's first draw calls, so that doesn't even come into play.

Not making sense at all in light of the PS4 having GDDR5 and being $100 cheaper and still profitable at launch. The only time this makes sense is when your RAM is really slow and there are things you simply cannot do fast enough direct to RAM, but when the ESRAM is barely faster than your RAM and slower than the competition's RAM, I can't find any sense in it.

 

[sze5003]

If they went with ddr 5 the specs would have nearly been exactly the same for both systems. I do believe the ps4 processor is faster. I think it's 2.0ghz vs 1.6?? Or maybe I got it backwards.

My bet is they thought going with their original plan of drm and cloud and security checks that the esram would help since they were expecting to have most of the cloud do some work and have the box take care of the rest.

 

[Arkaign]

True. Add ddr5 and 33% more gpu and they'd be equals not counting os considerations which may or may not be measurably large.

 

[smackababy]

The problem is, theoretical bandwidth means so little. Look at the benchmarks for a graphics card in PCI 16x, 8x, and 4x*. There is such a little difference. We have trouble actually saturating the lanes, DDR3 vs DDR5 won't make a huge impact on the games.

*This was a few years back and cold have possibly changed, but that is doubtful. It is just to highlight max theoretical bandwidth isn't a great metric for gaming power.

 

[exdeath]

The problem is, theoretical bandwidth means so little. Look at the benchmarks for a graphics card in PCI 16x, 8x, and 4x*. There is such a little difference. We have trouble actually saturating the lanes, DDR3 vs DDR5 won't make a huge impact on the games.

*This was a few years back and cold have possibly changed, but that is doubtful. It is just to highlight max theoretical bandwidth isn't a great metric for gaming power.

It means a lot when CPU and GPU share the same memory and you are rendering multi sampled depth buffered 1080p with depth of field, etc.

This isn't a PCIe link between two independent subsystems with their own dedicated memory like a PC. This is shared memory.

There is little bus traffic on PCIe because there are gobs of DEDICATED RAM on the graphics card. The GPU though I gaurantee is chewing all your onboard graphics ram bandwidth. Now share it with 8 CPU codes for both program and data at the same time.
Th

 

[Arkaign]

The problem is, theoretical bandwidth means so little. Look at the benchmarks for a graphics card in PCI 16x, 8x, and 4x*. There is such a little difference. We have trouble actually saturating the lanes, DDR3 vs DDR5 won't make a huge impact on the games.

*This was a few years back and cold have possibly changed, but that is doubtful. It is just to highlight max theoretical bandwidth isn't a great metric for gaming power.

Think about what you're saying for a second.

Yes, in the PC world, that is mostly true, particularly with lower-end and midrange GPU speeds. For the sole purposes of connecting the GPU to the rest of the system.

There is a MASSIVE caveat to that however. And that is that the GPUs you're talking about have their own memory that the graphical/modeling data is loaded into during the game. And that memory is GDDR5 on any decent card these days. The memory that the GPU is directly using in-game needs to be FAST in order for framerate/detail to be good. PCI Express is only used to get the data TO the card during the loading. This is why when you turn the settings up on PC beyond what the card can keep entirely on the card's memory, you get poor results.

So yes, the GDDR5 is going to make a tremendous difference in-game for cross-platform ports between the consoles.

Basically you can get an idea of how big the gap is by comparing cards that come in both DDR3 and GDDR5 flavors.

http://tech2.in.com/reviews/graphics-cards/his-hd-6670-1gb-ddr3-graphics-card/231692#show

That's a lowly 6670, which is less than half the power of the ~7780 in the PS4. Even it is severely hamstrung by the DDR3 vs. the GDDR5 variant. Add GCN and more compute units, and you've got a lot more data that needs to be moving around quickly.

I hope that clears things up for you 🙂

 

[smackababy]

It means a lot when CPU and GPU share the same memory and you are rendering multi sampled depth buffered 1080p with depth of field, etc.

I am not experienced coding to the metal, as DOS was dead by the time I got into development and have stuck with PC. I still can see it being better by having faster memory, but the amount it actually translates to better graphics or increased FPS? I don't think so. Especially with games being locked at 60FPS, I don't think the PS4 will have that big of an advantage. Possibly at the end of it's lifecycle, but for most of it the games should be close enough people won't notice.

 

[Arkaign]

I am not experienced coding to the metal, as DOS was dead by the time I got into development and have stuck with PC. I still can see it being better by having faster memory, but the amount it actually translates to better graphics or increased FPS? I don't think so. Especially with games being locked at 60FPS, I don't think the PS4 will have that big of an advantage. Possibly at the end of it's lifecycle, but for most of it the games should be close enough people won't notice.

It's okay not to understand at first. But think of it this way :

Game is loading, so the game data is loaded into system memory and graphics memory (in the case of XB1/PS4, this is shared). The PCI Express bus you talked about earlier is just the road to get the graphics modeling/texture/etc onto the GPU in order to play it.

After the data is loaded and you're ready to play, you need the GPU to be able to draw from that memory very quickly in order to draw frames. The more detailed the game data, the faster you need to run through it to draw those frames. Have you seen the difference in real-world detail and performance when you're comparing cards with a greater than 2X difference in memory speed? It's huge.

So what that means is that PS4 variants will have either smoother framerate with identical content (by a large factor), or they can simply turn up AA, texture details, modeling details, draw distance, fog/FX, shaders, shadows, etc, without dipping below the target framerate.

Facts are facts, and specs are specs. This isn't PS3 vs. 360 anymore, when each was so different that it was impossible to compare, and with odd/terrible limitations to deal with. The PS4 and XB1 are nearly identical in basic design and coding base, but the actual horsepower between them is a fairly sizable gap.

This doesn't mean that first-party titles on XB1 won't be great. Graphics aren't everything by any stretch. This does mean however, that unless a studio/dev intentionally makes a horribad PS4 version, that they will invariably look and/or play better every single time.

 

[exdeath]

I am not experienced coding to the metal, as DOS was dead by the time I got into development and have stuck with PC. I still can see it being better by having faster memory, but the amount it actually translates to better graphics or increased FPS? I don't think so. Especially with games being locked at 60FPS, I don't think the PS4 will have that big of an advantage. Possibly at the end of it's lifecycle, but for most of it the games should be close enough people won't notice.

Look at it like this.

Max out the graphics on you PC to the point you can barely maintain 30 fps and utilizing all available GPU memory bandwidth, Eg texture/shader/frame buffer/fillrate limited not compute limited. You are now using max memory throughput of your graphics memory even if you don't touch the PCIe bus again.

Now remove your main ram, remove PCIe, connect your CPU to video ram directly, and put all your non graphics code, data, OS, applications, threads, etc that was in main ram and put in graphics RAM. Multiply by 8 cores.

Remember, you were maxed out already at barely 30 fps with just dedicated graphics, now you have to constantly feed instructions and data to 8 CPUs too. What happens when a CPU wants to do something like sort a large data set in RAM, zero an array, process a list of objects, etc?

 

[exdeath]

This is one area consoles are more efficient. On a PC the graphics resources are duplicated in main ram and video ram.Either the API gaurantees availability of the texture once loaded and keeps a private copy to reupload to the graphics card or the programmer must keep a copy ready in the event of "surface lost" as a result of textures being evicted from graphics ram and needing reloading. In a unified memory model this doesn't happen because the graphics RAM is the main ram and the programmer has more direct control of it.

 

[Arkaign]

Yeah go check this out :

http://anandtech.com/bench/Product/537?vs=549

The Xbox One GPU is fairly close to that 5750 when you consider it hobbled by the DDR3 memory speed. The PS4 GPU is fairly close to that 7850 (actually it's slightly more powerful than 7850, but close enough for comparisons sake).

 

[Arkaign]

This is one area consoles are more efficient. On a PC the graphics resources are duplicated in main ram and video ram.Either the API gaurantees availability of the texture once loaded and keeps a private copy to reupload to the graphics card or the programmer must keep a copy ready in the event of "surface lost" as a result of textures being evicted from graphics ram and needing reloading. In a unified memory model this doesn't happen because the graphics RAM is the main ram and the programmer has more direct control of it.

Exactly. Console architecture now basically makes something like the PCI Express bus irrelevant as a means of getting data back and forth between system ram and video memory.

 

[Arkaign]

If we clocked a 7790 and 7850 at 800Mhz Core each, and then matched the memory speed for the consoles, we could actually make a viable test case to show performance gap.

That doesn't mean that the consoles won't perform better than the PC (when comparing equal spec console vs PC, or close to it), all things being equal due to API considerations and OS overhead. At the same time, the gap between low and high video memory bandwidth doesn't go away at all.

 

[KaOTiK]

True. Add ddr5 and 33% more gpu and they'd be equals not counting os considerations which may or may not be measurably large.

MS is supposedly (iirc this is all rumored by multiple trusted sources) holding back 3GB, 2 cores, and around 10% of the GPU (I know it isn't a GPU, just woke up 😛 ) for the OS's. Sony is supposedly (again rumored but these are unconfirmed rumor, but they seem to fall in line) holding back 1-1.5GB, 2 cores and 10% of the GPU.

If that is true it will widen the gap a lot more.

 

[Aikouka]

The problem is, theoretical bandwidth means so little. Look at the benchmarks for a graphics card in PCI 16x, 8x, and 4x*. There is such a little difference. We have trouble actually saturating the lanes, DDR3 vs DDR5 won't make a huge impact on the games.

*This was a few years back and cold have possibly changed, but that is doubtful. It is just to highlight max theoretical bandwidth isn't a great metric for gaming power.

PCI-E lane speed isn't comparable though. The lane speed only affects the transfer from devices that are off the video card itself such as the hard drive, system memory, etc. The data rates that are being discussed are more comparable to the speeds on the video card itself between the RAM and the GPU. To give you an idea, high-end AMD cards have over 50% more bandwidth than even the PS4.

If you want to see how much this speed affects things, reduce the memory clock speed on your graphics card, which should reduce its throughput (memory bits * memory speed / 8 = bandwidth).

 

[Durvelle27]

i wonder if this will be available

Looks awesome