Building a gaming PC

[GuyNwah]

It would be best to only test with 64bit games me thinks, Skyrim for example is 32bit so I wouldn't really be surprised if the RAM limits apply to the VRAM as well.

Some have suspected the 32-bit and DirectX 9 legacies in Skyrim, but it's more widely observed than just Skyrim. And Skyrim doesn't have that problem on the 980. A related hypothesis is that it's due to the disabled shader modules on the 970 (which has 13 of 16 enabled; the 980 has all 16 enabled) causing a problem with the addressability of the entire 4GB range of VRAM.

The Evil Within, Far Cry 4, Shadow of Mordor, AC: Unity, and Ryse: Son of Rome are 64-bit titles that are affected. It's also observable in MSI Kombustor. Enough to say it's real, and nVidia knows about it.

What would be interesting to know is whether it appears with a 4GB 970 and 16GB system RAM. All the reports seem to be from units with 8GB system RAM. They seem to have hit a point where the 970 wants to page out, but if there's no free RAM to page to, it goes to the pagefile. The performance hit then comes from having to access the pagefile. That's one hypothesis, anyway.

There's a long and acrimonious thread on the bug at http://forums.guru3d.com/showthread.php?t=396064

 

[sidspyker]

Yeah I saw that thread, I'm a member there hence why it came to my attention.

On an unrelated note I was THINKING(not decided) of buying a new CPU, my eyes are on the 4690(maybe K, maybe not) but my dilemma is, Skylake comes out this year too, but my primary focus is a new CPU by the time Witcher 3 comes out Skylake will come out SOMETIME in the 2nd half of the year or might be possibly delayed, I don't wanna regret my purchase if Skylake turns out to be phenomenal, it's also 14nm so I'm quite curious about it....
(EDIT: Oh wait, that latest lineup is 14nm too isn't it?)

I have no idea what I really wanna say

---------- Updated at 04:37 PM ----------

I guess it's kind of a self defeating question... if you can even call it a question.

There's always something new coming out so wait or not wait is an age old question, bottom line is I'll just have to go for what's available at the time plus with Intel these days the performance decreases are incremental combined with the fact that I have a Phenom II x6 1090T at the moment means it'sll be a 'big' upgrade anyway...

 

[eskiMoe]

Some have suspected the 32-bit and DirectX 9 legacies in Skyrim, but it's more widely observed than just Skyrim. And Skyrim doesn't have that problem on the 980. A related hypothesis is that it's due to the disabled shader modules on the 970 (which has 13 of 16 enabled; the 980 has all 16 enabled) causing a problem with the addressability of the entire 4GB range of VRAM.

The Evil Within, Far Cry 4, Shadow of Mordor, AC: Unity, and Ryse: Son of Rome are 64-bit titles that are affected. It's also observable in MSI Kombustor. Enough to say it's real, and nVidia knows about it.

What would be interesting to know is whether it appears with a 4GB 970 and 16GB system RAM. All the reports seem to be from units with 8GB system RAM. They seem to have hit a point where the 970 wants to page out, but if there's no free RAM to page to, it goes to the pagefile. The performance hit then comes from having to access the pagefile. That's one hypothesis, anyway.

There's a long and acrimonious thread on the bug at http://forums.guru3d.com/showthread.php?t=396064

Well, I have 16GBs of DDR3 and just ran Kombustor's memory burner (3072mb) test:

 

[sidspyker]

@eskimoe, what brand of RAM are you running? (just curious, unrelated to subject)

 

[eskiMoe]

HyperX Fury.

 

[sidspyker]

Hah, I knew it as soon as I saw "1866 Mhz" >_>

@eskimoe white

 

[eskiMoe]

Hah, I knew it as soon as I saw "1866 Mhz" >_>

Can you guess the color too?

 

[tahirahmed]

Yeah I saw that thread, I'm a member there hence why it came to my attention.

On an unrelated note I was THINKING(not decided) of buying a new CPU, my eyes are on the 4690(maybe K, maybe not) but my dilemma is, Skylake comes out this year too, but my primary focus is a new CPU by the time Witcher 3 comes out Skylake will come out SOMETIME in the 2nd half of the year or might be possibly delayed, I don't wanna regret my purchase if Skylake turns out to be phenomenal, it's also 14nm so I'm quite curious about it....
(EDIT: Oh wait, that latest lineup is 14nm too isn't it?)

I have no idea what I really wanna say

---------- Updated at 04:37 PM ----------

I guess it's kind of a self defeating question... if you can even call it a question.

There's always something new coming out so wait or not wait is an age old question, bottom line is I'll just have to go for what's available at the time plus with Intel these days the performance decreases are incremental combined with the fact that I have a Phenom II x6 1090T at the moment means it'sll be a 'big' upgrade anyway...

I also suggest you go for 4690 (K if possible), if you're coming from Phenom II x6 1090T then it will be a huge difference anyway, also I remember reading somewhere that the K editions of Skylake will appear in 2016 so that's quite a wait.

 

[sidspyker]

Well there's also Broadwell's desktop lineup... whenever that comes, and that's 14nm

 

[tahirahmed]

Yes there are but I am sure if the performance difference between a top tier Haswell i5 and top tier Broadwell is worth the extra cost. Broadwell will be expensive for sure.

 

[Kinley]

So the GTX 960 just came out. Very affordable at a 200$ price tag.

Witcher 3 also getting some publicity on the promo page :

---------- Updated at 04:33 PM ----------

And some impressions:

 

[volsung]

$150 is a wide gap between the 960 and the 970. Is there a similar gap in performance? I'm guessing if there is, an eventual 960 Ti will sit in between the two.

 

[eskiMoe]

A nice budget card. But at this point I probably wouldn't buy a 2GB card anymore.

 

[volsung]

A nice budget card. But at this point I probably wouldn't buy a 2GB card anymore.

That makes me wonder how SLI setups distribute data across separate memories. Technically two 960s would give you 4 GB of distributed memory, enabling computations with up to 4 GB of memory, more processing units but more overhead.

The stupidest thing to do would be to maintain copies of all data on both memories and simulate a shared memory, data parallel approach. I hope games/drivers don't do that but I wouldn't put it past them. Ideally, each device should allocate only the necessary space. If overhead is minimal (which isn't unusual with linear systems involving matrix and vector operations, as they usually scale very well) coupled with the proper operations two 960s should be faster than a 970, for only a bit more money.

Just thinking out loud.

 

[Sana_mia]

That makes me wonder how SLI setups distribute data across separate memories. Technically two 960s would give you 4 GB of distributed memory, enabling computations with up to 4 GB of memory, more processing units but more overhead.

Nope two 2GB cards in SLI gives you 2GB memory, the second card is basically just cloning the memory.

Checked the pricing and honestly for 60-80€ more You'd get a 970GTX so not really an interesting card specially due to the 2GB.

 

[volsung]

Nope two 2GB cards in SLI gives you 2GB memory, the second card is basically just cloning the memory.

Wow, that's such a waste. What the hell is Nvidia doing?

 

[eskiMoe]

It's the same story with Crossfire.

 

[volsung]

It's the same story with Crossfire.

I would imagine. I'm still shocked. So inefficient...

 

[GuyNwah]

It's pretty much what has to be done to make SLI or Crossfire work. In N-way SLI or Crossfire, each GPU is rendering 1 of N frames (alternate frame) or 1/N of each frame (SLI split frame, Crossfire checkerboard) from what is effectively private RAM. You can't do that without equal copies of the data on each GPU, because each one needs the full set of resources, and there's not enough bandwidth between GPUs to use memory on the other.

 

[volsung]

It's pretty much what has to be done to make SLI or Crossfire work. In N-way SLI or Crossfire, each GPU is rendering 1 of N frames (alternate frame) or 1/N of each frame (SLI split frame, Crossfire checkerboard) from what is effectively private RAM. You can't do that without equal copies of the data on each GPU, because each one needs the full set of resources, and there's not enough bandwidth between GPUs to use memory on the other.

which of the two? different frames or different parts of a frame?

Maybe I'm judging based on my lack of graphics knowledge, but in the first case I don't see any data dependencies. Each frame comes with its own dataset and whatever other specifications and it should be possible to render each separately, wthout the entire set of resources. If additional information is needed (border issues, as in previous and next frame) then an appropriate distribution scheme is splitting in overlapping segments of n + a frames for each processing unit, and processing only the n frames. If each frame is split for rendering (which leads to additionaly communication costs for putting together the final result) then use the same scheme mentioned above for data partitioning except within a single frame (e.g. for a large matrix, row splitting with overlap if neighbors are necessary).

I don't know what the input and output representations are and how much bandwidth they need, but I'd think PCI Express is sufficiently fast to do this while keeping the load balancing overhead low.

This is a simple approach used everyday in parallel and high performance computing, for instance applicable to solve large linear systems using (non-contiguous) distributed memory computer architectures (a computing cluster for instance) and LARGE amounts of RAM (more than 32 GB).

I suppose the reasons are pragmatic. Keeping copies of everything is much simpler.