torok said:
What I'm actually saying is that a PS4 is better than your PC and better than every other home PC around there. Simply because it was architectured to have a lot of memory bandwidth planning about living in a world where it's limited.
|
I had to laugh.
if the PS4 is really better than my PC...
Then why Can't it:
Transcode a movie to another device, whilst playing THREE copies of Battlefield 4 at 1440P and having plenty of spare processing power to record television, while decoding music off my NAS while doing xSplit and having a dozen background applications all at the same time?
I'll tell you why. The Playstation 4 isn't fast enough and it will never be faster than my PC.
You can keep dreaming in your fantasy land.
The Playstation 4 is a cost sensitive device, it has a low-end processor a mid-range graphics card and only 8Gb of Ram.
I think you would have to be the only person on this forum/planet who actually thinks the PS4 is better than my PC.
torok said:
Now about physics. 4 GPUs, that's nice! Now split work between them. Pass data between them, loading everything on CPU RAM and passing to other GPU. 4 GPUs will give you a speedup of, in a good case, 2 X, unless you run a optimized and heavily parallel algorithm. All of that just shows that you don't understand GPU programming. First, raw teraflops aren't a good measure. The entire architecture must allow all that power to be used. Take Seti@Home as an example. It has a high count of raw floating point calculation power, but it will lost for much smaller clusters using Infiniband because of comunication overhead. And that is the problem with parallel computing, comunication costs a lot. And a GPU can be treated as a single cluster (except for being a SIMD machine while regular clusters are MIMD). And please, don't say your GPU is faster than 2 PS4 GPUs and a CPU. You can't compare a CPU and a GPU like that. MIMD x SIMD. A MIMD CPU can execute different instructions at the same time while a SIMD GPU only executes the same instruction on all cores with different data. A MIMD machine can do everything a SIMD machine does, while the opposite isn't true. If I run an algorithm with a lot of conditional statements (AI for example) or a recursive one, your GPU won't beat even the Pentium 4 of the article I passed. |
Completely utterly wrong.
Data doesn't have to go through System Ram or the CPU in order for the GPU's to communicate with each other, anyone who has written compute jobs for GPU's knows this.
If you have bandwidth sensitive data that needs to be sent you can use some of the spare compute resources on the GPU to compress the data before sending it off to the next GPU via the PCI-E Bus. (Which is almost 16GB/s for PCI-E 3.0 16x.)
Data-sets can also be duplicated in each of the GPU's memories, in-fact this is the default behaviour, so the GPU's only need to poll each other to see what work has been and needs to be done, it's not like you have to transfer gigabytes of data all at the same time.
http://research.microsoft.com/apps/pubs/default.aspx?id=192884
https://developer.nvidia.com/gpudirect
As for Teraflops not being a good measure of performance, you're partly right.
If you are doing a compute job that only deals with floating point mathematics, it is essentially the only measure of performance that you need to worry about.
Game engines are a little differen't however they don't just deal with floating point and if they do they can have varying levels of precision depending on need.
And I'm well aware you can't compare a CPU and GPU due to the vastly different ways of how they process information, CPU's excel at Integer math, whilst GPU's excel at floating point.
CPU's are serialistic in how they process data, the best example is to use a book.
A CPU will read a book from the first page to the last, page by page.
A GPU will read every single page of a book at the same time.
Sometimes applications/games have dependencies in the processing chain for timing purposes, so they need to be processed in a serial manner.
It's a good thing I have the best of both worlds, GPU's significantly faster the Playstation 4 (And more of them) and the fastest consumer CPU money can buy.
torok said:
You still refuses to see the memory wall. I sent you a link for a paper of computing conferences. I sent you a link to a paper wrote by John Von Neumann itself. If he can't convince you, no one more can. I will say again, do you really think that your non-scientific observations of your own computer usage are a better source of info than a research from Von Neumann? I'm not talking about running simpe tasks. I'm assuming that a group of developers will sit down and say "let's use every single core, every drop of power of the GPU, offload tasks as possible, optimize access to minimize cache misses, use well the bandwidth and extract every single drop of power here". Thats what gives you power. Balancing load to find a point where you maximize the usage of everything on that environment. And yes, it means even using cache correctly to avoid misses. And for that, you need to know exactly what is under the hood. Every spec. That's why supercomputers are homogeneous clusters, because knowing exactly the balance of power between nodes allows you to make the right decisions. |
I'm not refusing to see it. - You're just missing the point completely, but that's not my fault.
It's all logical and you're blind to it.
torok said:
Just to end, real time raytracing is planned by NVidia in around 5 years, after 2 new iterations of GPUs. 3 or 4 console generations are 20 years, that's crazy. They have a double effort running in parallel, their GPU and tools (OptiX) improving and newer and more optimized algorithms do do faster ray tracing (using more optimized tree data structures, etc). Actually, when Sony and MS asked developer of what they wanted for a new console, they were asked to not use a ray tracing capable GPU, because all engines would have to be redone from scratch to use it. Of course, that was a little exagerations since not even a Titan could do it at realtime, but current GPUs are pretty close of doing that, so I think they just want to be sure of it. |
Ray Tracing will be brought in before 20 years, but it will be an additive technique there are some things that Ray Tracing isn't as efficient/good at compared to some of the things like Photon Mapping.
You won't get a full-on Ray-Tracing only implementation for a long time yet.
This current generation is mostly going to be focused on geometry and lighting, they are the two largest gains consoles will see over the prior generations.
czecherychestnut said:
I'll only make two points. 1) Regarding GDDR5 vs DDR3 latency, yes GDDR5 has higher latency (clock cycles) than DDR3, however what this analysis fails to take into account is the difference in clock speed between the two. The PS4 uses GDDR5 running at 5.5GT/s, which gives it a master clock speed of 1.375GHz (GDDR does 4 transfers per cycle). DDR3-1600 runs at 800 MHz (DDR3 does 2 transfers per cycle), so if the latency is measured in ns (which matters more than clock cycles as its the actual time the CPU will be waiting for data) then GDDR5 latency can be 70% higher when measured in clock cycles and still be faster than DDR3-1600. I'm currently on my tablet which sucks but later i'll link a Samsung GDDR5 data sheet that shows the latency is less than 70% higher in clock cycles |
You take the CAS Latency which is latency per clock cycle, and using some math you can get the real-world ns latency.
GDDR5 is roughly 20% higher than DDR3 in ns.
If you were comparing CAS latency, then it's significantly higher for GDDR5 because by it's very nature runs at a higher clock.
I remember seeing allot of people complaining about higher CAS latencies on DDR3 when it was first introduced, but that was because of it's higher clocks, the actuall ns latency remained the same though.
You also have other "latencies" in DRAM too, but they aren't important for the topic at hand. (And I would assume the differences would be about the same anyhow.)
) 
