I've been pondering the effect the new cell chip would have on gaming and vid cards.
I don't see any immediate impact on the graphic card market, but if these processors are as powerful as IBM, Sony and Toshiba claim they are ("a super-computer on a chip") then maybe in the future they'll be used for graphic cards too.
The big issue I see is if they are so good, it'll cause a drastic change in the way games are written and games wouldn't be compatible with DX or GL.
Course, this is all speculation. These chips could suck balls!
And, I remember reading a while back that the graphic card chips are much much much more powerful than the fastest Athlons and P4s, so what diff would it make anyways.
Well that's enough of my rambling, what do u guys think?
CELL is a good chip. But it is not quite a "Super Computer on a Chip". At least not yet. If they are able to put 4 CELL processors on a single die (~1TFLOPs) then we are looking at Supercomputer type performance.
CELL will not replace GPUs. Period. CELL is highly parallel. But GPUs are also high parallel. But GPUs have a ton of useful features *specific* to 3D rendering that would take magnitudes of more power to do at equal performance with a general processing type chip like CELL. CELL has a lot of power but GPUs have a lot of special pathways designed to speed up the rendering process and also is designed to mask latency.
The big deal with CELL is that CPUs are just transitioning to parallelism and CELL is making a big splash as it is not only parallel (1:8 PPC:APU... basically the PPC core delegates tasks to the auxiliary processing units), but CELL is designed to work in tandem with other CELLS. The goal is to put 2,4,8,16, etc... of these things on a single die as the manufacturing processes allow. e.g. STI (Son-Toshiba-IBM) are aiming to put out a 32chip Rack that performs 16TFLOPs. In the NYT article I read this in this seems to be a "future" project, which makes sense (current 90nm CELL is 256GFLOPs, the 32chip rack would need 512GFLOP chips which indicates this setup will most likely be a 65nm chip with 2 CELLs on each chip).
This bodes well for the PS4 also. If PS3 gets 1 or 2 CELLs in the Broadband Engine (I doubt 4, but who knows), that means in 5-6 years PS4 could incorperate dozens upon dozens of CELLs. This would be great for backwards compatibility and scales on and on and on... Of course some tweaking of CELL will happen over time (new features, more cache, etc...) but the concept is simple: Make a chip that is powerful and scales very well. STI has done this with CELL.
Now the down side: the PPC core is said to be a little less powerful than the G5. Makes you wonder how the Xenon with 2-6 PPC cores will compare... with the Unified Shaders in the ATI R500 and more PPC cores it looks like the Xenon will be better for general computing type tasks whereas the CELL, which might be doing the Vertex Shading, with the nVidia chip in the PS3 may be a floating point monster. This leads to the next point: The CELL is a great chip, but is not the "end all-be-all" chip. Even the very Pro-CELL people I have talked to at B3D are pretty clear about that. It will do geometry, AI, Physics, etc... type stuff VERY well. But it is not a GPU replacement and it is not a desktop replacement (yet). Just like you would not want a GPU for your CPU, CELL is not quite ready to replace our desktop chips. It may try to do that one day but it is still in its infancy. Many have tried, all have failed.
CELL... Good tech, will make for one heck of a console, but don't let the hype get to you too much.
EDIT: Just to bring this back around to the GPU issue. CELL may make some inroads in CPUs because CELL is a mutli-core processor from scratch whereas x86 chips are just now getting to the mutli-core/thread stage of development. But both are faced with a software library not designed for taking advantage of multiple-cores. CELL has an adavantage of being designed from the ground up as a multicore unit (a single CELL is really 9 processors, a PPC delegating and 8 APUs) and will scale well. But I am not gonna dump $10k in software for CELL anytime soon! On the GPU front the CELL does not pose much threat because parallelism is part of the design and does not face the software issues. A R9700 basically has 2 "cores" and a 6800GT has 4 "cores". Of course I use cores here lightly because they are more parallel pipelines--GPUs do not duplicate redundant parts of the chip. This is why we will continue to see GPUs with more and more pipelines and not a lot of dual core and quad core GPUs (unless some big change happens) because the redundancy of the chips.
Now we *may* see a CELL Vertex Shader some day. I think nVidia was aiming for this and may be why they are dragging their feet on unified shaders. They are all for unified code, but would prefer the units remain separate. Having a modified CELL doing all the Vertex Shading would free a nVidia GPU to dedicate more realestate to Pixel Shading. And since nVidia is helping with the PS3 (they are making the GPU that is linked to the CELL) it would make sense for nVidia to leverage this research just like they leveraged the Xbox research for the nForce. But with ATI in bed with Microsoft now and the fact nVidia sucked MS into a bad chip deal on the Xbox I doubt MS is really excited about allowing nVidia leverage the CELL on the PC in GPUs. This is all speculation of course But from the Xenon patents it seems the XeCPU and R500 are more general purpose. The R500 has unified shaders which means it can load balance. So if they had 48 Shader units, a geometry heavy game may use 16 VS and 32 PS on average, whereas a pixel shader heavy game would use something like 4 VS and 44 PS on average. Of course a unified model allows for a lot of balancing and therefor I would guess that minimum framerates would be pretty good--a high geometry area is not going to hit you as hard because the GPU can dynamically adjust. Another aspect of the Xenon is that the PPC have "vector" units that can take some of the geometry work (MS requested a very high speed link where the CPUs and GPU can read from the same cache). If this means a designer could have the CPUs do all the vertex shading and let the GPU do all the PS we do not know but it does open some questions. Maybe the biggest difference we know so far is that it looks like PS3 may have 1 PPC core, whereas the Xenon may have 2-6 PPC cores. This would indicate the Xenon would be better at general processing stuff and the PS3 at math intensive tasks. Both sound great and both seem to have different emphasi. The PS3 will be at least 256GFLOPs whereas the Xenon CPUs are estimated in the 80GFLOPs range (to compare, a top of the line P4 is in the high 20s). The Xenon CPUs though will be very powerful for certain tasks though. All in all the new consoles are looking to be excellent devices. Some of the tech invariably will reach the desktop, but again, just as you would not want a GPU being your CPU, you would not want CELL as a GPU right now. Maybe if PS4 is about to have 32 CELL cores and 16TFLOPs of performance we can look at this, but not now.<P ID="edit"><FONT SIZE=-1><EM>Edited by Acert93 on 02/08/05 01:10 AM.</EM></FONT></P>
This discussion reminds me of our Power5 discussion of a while back, and we came to the same conslusion, as much as GPUs are becoming CPU equivalents, they will never be fully interchangeable and especially GPUs will get more specific as time goes on. Sure their resultant power could crunch some numbers as well, but really the combo of both for the foreseeable future will always trump a small collection of only one of the other.
- You need a licence to buy a gun, but they'll sell anyone a stamp <i>(or internet account)</i> ! - <font color=green>RED </font color=green> <font color=red> GREEN</font color=red> GA to SK