Well the Cell is already GPU like so I don't think they'll have any problems with whatever they plan on doing. The question really is whether it can perform comparably, better, or worse than a DX12 level graphics chip.
Predict: The Next Generation Console Tech
- Thread starter Acert93
- Start date
- Status
Well the Cell is already GPU like so I don't think they'll have any problems with whatever they plan on doing. The question really is whether it can perform comparably, better, or worse than a DX12 level graphics chip.
That really speaks to my question.
Since it is already GPU like in some ways, will it have less of an impact in the future PS4 since, if I understand your post, standalone GPUs will be taking more and more of those tasks of CPU/GPUs.
I guess I'm wondering if it will really be necessary next gen to have an exotic CPU compared to this gen. From what I've been able to garner, it appears that most games are not CPU bound.
Almost all of the AAA PS3 games have the Cell helping RSX in the graphics department. Will that really be necessary next gen?
Well, the real question is can Sony leverage SPUs the way Intel uses x86s for Larrabee? Intel failed to come up with a product thus far. It's not easy to create a competitive GPU out of SPU or x86.
PS4 would be more interesting if Sony go their own way again like PS and PS2 instead of leveraging PC part again, even if the result is something with less peak power. But with no Kutaragi at the helm, I just don't think they will.
It would be suicide, any company outside ATI/Nvidia has no chance even coming close to those two any more.
Look at Intel trying to break into high end GPU's with Larrabee. Nobody has more money or chip experience than Intel, and they fell flat.
ATI and Nvidia have thousands of man years of engineering to build on incrementally, plus just absurdly polished drivers with a similar time investment. Not even to talk of the patent situation. Patents alone might make competing with Nvidia or ATI a non starter.
Even for Cell, Sony needed to partner with IBM to get it done.
Look at Intel trying to break into high end GPU's with Larrabee. Nobody has more money or chip experience than Intel, and they fell flat.
ATI and Nvidia have thousands of man years of engineering to build on incrementally, plus just absurdly polished drivers with a similar time investment. Not even to talk of the patent situation. Patents alone might make competing with Nvidia or ATI a non starter.
Even for Cell, Sony needed to partner with IBM to get it done.
It's the old "make the common case fast". For a graphics part, filling pixels is the common case. With every peripheral function you insert, you reduce the density of the most important feature.
You're still going to have two distinct parts split the overall load between them, if for nothing else, then for manufacturability reasons. If you fill in two specialized systems, you should expect better efficiency than with two general systems that could do either task, but just happen to be cast into something they're not optimal for.
Kinda like people generally don't have four equivalent limbs anymore, but find more success in life using two feet, which are great for moving around, and two hands, which are great for interacting with the world.
aaronspink
Veteran
Which is why all GPUs are split into pixel shaders, geometry shaders, and vertex shaders? What? They merged them all? Damn, there goes that keep everything optimized separately. The primary purpose of GPUs with modern workloads is shading. Filling pixels is a small small fraction of both the die area and workload of modern GPUs.
In light of some of the things Corinne said, the interesting questions IMO are:
Will next gen have one processor (or a group of identical processors) to handle both game code and graphics code?
Does a software renderer require a monolithic processor (or group of identical processors), or can a more traditional CPU/GPU setup handle software rendering just as well?
Is the system's memory architecture more important than the CPU/GPU split?
Is 2012-2013 (assuming the next consoles need to be somewhat forward looking) simply too early for this nonsense?
@Aaron
If you were tasked with designing a console for nextgen, how would you set it up?
Will next gen have one processor (or a group of identical processors) to handle both game code and graphics code?
Does a software renderer require a monolithic processor (or group of identical processors), or can a more traditional CPU/GPU setup handle software rendering just as well?
Is the system's memory architecture more important than the CPU/GPU split?
Is 2012-2013 (assuming the next consoles need to be somewhat forward looking) simply too early for this nonsense?
@Aaron
If you were tasked with designing a console for nextgen, how would you set it up?
You'll note that this grand unification round-about doubled the transistor budget in ATI's case, and much worse for NVIDIA. PC graphics can evolve in weird ways now because Microsoft made it impossible for new players to compete. Other segments though couldn't afford to consume entire process generations just to include the latest checkboxes.
Having silicon idling because there's no work for it do.
This wasn't rasied as a discussion for the merits of PS3's hardware efficiencies! The idea was presented by Corinne Yu that programmability is the future. I raise MLAA as an example of where programmability is beneficial. I do not make the claim that SPEs were the best choice for the best graphical capabilities of a console designed to 2005 spec. That said, if those 4 SPEs weren't included and instead their silicon budget had been spent on the GPU or RAM bus or anything else, do you believe GOW3 would look better than it does now? I haven't seen anything anywhere else to suggest that the level of IQ obtained from MLAA isn't the most efficient IQ improver we have, taking 5ms across 4 SPUs instead of 120ms on RSX, and 16xMSAA is beyond the bandwidth of these consoles. If you were designing a console in 2005 to have the best jaggie-free IQ and had the know-how of GOW3, would you not pick PS3 as (currently) the only means to pull it off? Of course it doesn't all come down to IQ, but that's not the point! The point is, programmability gives options that enales slower hardware to outperform faster hardware.
That's all true. However, some titles (okay, maybe one!) are dead-easy for and perfectly suited to a raytracer - a snooker game with true reflections and analytical CSG geometry. Such a game is not a particular good fit for GPUs. You'd need complex meshes to model the perfect forms of a sphere that is a doddle to render with CSG. You'd need to render out all sorts of reflections versus the straight-forward trace of a ray. Because GPUs are designed to be fast at a particular (and popular) style of rendering, they are limited in opportunities to try alternative methods that would be better suited to the job.
Maybe I'm off with my dates, but IIRC Intel's paper on which all this is based appeared in 2009. Unless Santa Monica was ahead of the game and struggling to develop a new PS3-specific AA engine since day one, any and every developer has had as long to work on MLAA implementations. I'm not saying it won't happen on GPUs, just as GPGPU has shown tasks can be reengineered to fit available solutions, but it's a struggle the algorithm doesn't inherently have, a struggle caused by the hardware confines. It's also interesting that this graphics technique was not shown to the world by nVidia or ATi or any other graphics hardware vendor who have given us all sorts of variations on MSAA. What does that tell you about mindsets and out-of-the-box solutions?The current solution
We gave PS3 developers 4+ years to figure this one out, lets give the GPU guys the same time frame to justify their hardware.
That's not what the hardware was designed for, but people being clever have found opportunities to exploit the hardware. However, Perfectly Programmable hardware doesn't have a name like "Vector Rasterizing Unit" that hides the fact it can also correlate your databases if you shuffle your database data up into an abstract format.
I'm pretty sure my initial response was "it won't happen next-gen". There's still room for discrete GPU architectures at the moment. However, the ideal for developers IMO is to do away with these different bits of hardware and just have a single core that does whatever you want it to. That's the ideal. How close the world of technology gets is up in the air, both due to limits in fabrication and competition meaning no single platform.
Interesting analogy. You're right, legs and arms are specialised for different purposes. That works well for people. And if you want to build a robot to do the things people can do, then legs and arms makes sense. But what if you're building an assembly line? The arms and hands are ideal, but the legs serve no purpose. Does it makes sense then to build bipedal automotons to sit at convey belt piecing components together, or would it be better to turn the legs into arms and have more practical functional units for the job at hand?
In the case of GPUs, we had arms and legs - vertex units and pixel units - and yet they've been combined. It's true, given a silicon budget, the specialised hardware is better. 200M transistors of pixel shaders and infrastructure will shade more pixels than 200M transistors of unified shaders and infrastructure. And 100M transistors of vertex shaders will process more vertices than 100M transistors of unified shaders. But which GPU will perform best in real-world cases; the 300M transistor US architecture, or 300M transistor discrete vertex and pixel shaders? The flexibility of US means better overall performance with less idling hardware, and more opportunities to the developers.
This is the ideal behind programmable graphics hardware, with the added benefit of being simpler to design for as you won't have to worry about specific workloads, much as US has liberated developers from worrying about vertex to pixel ratios. It won't be as fast as the current specialist rasterising hardware, but it will enable more approaches, and in my opinion, as much faith as anything, developers will find such amazing ways of rendering stuff if freed from the triangle-mesh, fragment shaded model of the past two decades
Its hard to really say at this point. Whilst the SPUs weren't good enough for the typical workload in 2004 when they were developing the PS3 due to programming constraints and the nature of the GPU workload at the time, they may fit the future model of GPU workloads even better given the programmable nature of the GPU architectures moving beyond the first generation compute shaders and the reduction of relative importance in terms of proportional die area of the ROPs and texture units. Developers are already working on an SPU model for graphics so its not much of a stretch to consider them taking it further or even to the point where they use one main CPU/GPU die with a secondary die consisting only of fixed function units + massive cache given that both Tex and ROP units are heavy consumers of bandwidth.
I.E. Have one large 200mm^2 die for the CPU with an array of SPUs + a few specialised CPU cores and one smaller 100mm^2 die for the GPU with say 30MB of on die cache for both the texture and framebuffers. Essentially make it a dumb daughter die which can be integrated into the main chip at some future point.
See the above answer + like always it really depends. CPUs and GPUs are going to meet in the middle at some point and it really does depend on what classification you give the SPUs in the Cell almost as much as it depends on whether a GPU can perform GPU + SPU class tasks more efficiently than a SPU can perform most GPU + CPU tasks more efficiently given transistor and power efficiency overall. Its this balance which will determine the balance of their next system.
Never say never in the technology business. Time and again a company has come from nowhere and gained significant market share. There was once a time when Nvidia stood alone as the top of the GPU field and they were considered untouchable and almost immediately afterwards ATI came onto the scene.
I don't think the absurdly polished drivers are going to help significantly in terms of developing for console hardware as they are mainly aimed at extracting performance from an open platform and are an unnecessary factor for a closed platform. PC GPUs are built to run legacy programs as well as future programs whereas a closed platform can take a blind leap towards the future without care for its ability to perform on legacy software.
aaronspink
Veteran
And increased performance at the same budgets in the majority of games. The transistor budgets double roughly ever gen just from process changes but the switch to unified by both Nvidia and ATI resulted in both better overall efficiency and better overall performance. The reason why is that the perf/cost ratio of dedicated vertex/pixel shaders wasn't high enough to overcome the benefit of the combined pool which could be dynamically switched between the tasks.
Quite honestly, tessellation will go the same way as separate vertex/geo/pixel shaders. Likely next gen or the one after, tessellation will just be another program run on the shader infrastructure. The only thing likely not to transition to being just another program is the texture samplers. Everything else on a GPU really cannot justify via its perf/cost being unique. I expect dedicated ROPs will disappear relatively soon.
Yup. I have to think Sony will go with a next-gen Cell and an up to date NVidia or ATI part, if they have any interest in keeping the ability to sell their current software base (and allow upgraders to keep their purchased software on their shiny new hard drives).
I know the speculation is that they're avoiding that path, though.
EE-based would be an insane nonsense nowadays. Pixel/vertex/geometry-shaders are the requirement. A future in-house design would surely be Cell-based, with the possibility of using in other devices which would be the only point to an in-house component. EE based?...not on your nelly!
It makes sense only from the perspective of licensing costs. I suspect that the royalties on various PS3 technologies have cost them an arm and a leg!
But EE is only the CPU, and it's only a MIPS CPU with 2 additional Vectorunits. So it's not so different to the CELL Design.
@ Squilliam
What royalities ?
Hence my WTF moment
I know actually few about the PS2 hardware and how devs used it. I read quickly on wiki and EE philosophy seems pretty close to the Cell: a CPU core + vector units. I see nothing that would prevent vector units to handle what ever shaders depending on how they evolve. Why would Sony go back to EE instead of growing/reworking the cell, I don't know. Maybe they feel more confortable working on a MIPS CPU? Or they think that EE is a better base than the cell to design a graphic oriented multi-core CPU? Licensing issues as Squilliam says? Sony supposedly can change the Cell but to witch extend, can they do whatever they want to the PPU, I don't know? Do they have the competences to do so? Same answer.
I don't know much either about PSP internal and nobody knows about the PSP2. Is PSP close to PS2 from a software POV? Could they plan have the psp2 and the ps4 pretty close from a development pov?
I don't give much credit to Groo for that one, still Charlie has been pretty good sources for quiet some stuffs even if his early papers on a matter can prove semi-accurate
Last edited by a moderator:
- Status
Similar threads
- Locked
