StarDestroyer.Net BBS

Hi,
I am planning to build a new rig and am undecided on several components.
It will be a computer mainly for professional 3D work (3ds Max and Blender) and the computer should be fit for CPU and GPU rendering, even though GPU rendering is more important.
The budget is around 2000€.

One thing I am uncertain about is the CPU.
I think either a i7 5820K Haswell-E (395€) or a i7 6700K Skylake (344€) would be a good choice.
The main board I am not certain about either.
for the 5820K Haswell perhaps the ASUS X99-Pro(355€) and for the Skylake an ASUS Maximus VIII HERO (212€)?

The other components I am more certain about but if they are not good for some reason I'd be happy to exchange them for something else:

The case (Corsair Carbide 200R) comes with two 120mm fans which I am not sure if they are good or if I should invest in two separate case fans.

- EVGA GeForce GTX 970 ACX 2.0 SuperClocked
339€
This will be combined with an old GTX 670 I have in my current computer

- RAM: Kingston ValueRAM DIMM 32 GB DDR4-2133 Kit
123€

- PSU: bequiet! POWER ZONE 750W
130€

-CPU Fan: Alpenfön Brocken Eco
34€

- Case: Corsair Carbide 200R
62€

- SSD: Some 250GB or 500GB drive (70€ to 120€)

- HDD: Some 2TB drive (around 70€)

So, at the moment I am at around 1400€ to 1600€ in total, including taxes. If you think that some other, more expensive components would make more sense there is some leeway upwards.

I see no good reason for an optical drive as I don't think I've used the DVD drive in my current computer in the last 2 years.
For an OS I'll buy a used Windows 7 professional for 30 bucks or so and have a dual boot option with Ubuntu.

So, what do you think? Is this reasonable and which CPU and mainboard should I get?

NOTE: I don't build PCs for rendering,

salm wrote:One thing I am uncertain about is the CPU.
I think either a i7 5820K Haswell-E (395€) or a i7 6700K Skylake (344€) would be a good choice.
The main board I am not certain about either.
for the 5820K Haswell perhaps the ASUS X99-Pro(355€) and for the Skylake an ASUS Maximus VIII HERO (212€)?

I'm going to take a pass on this. If it were gaming, I'd say the 6700 due to the higher clock and newer manufacturing. But more cores for rendering might sell it for the 5820.

The case (Corsair Carbide 200R) comes with two 120mm fans which I am not sure if they are good or if I should invest in two separate case fans.

Depends. With a CPU watercooler, you can push any excess heat off the CPU straight out of the PC. I ran 2 MSI 970s overclocked and never had them get above 80C at the highest load. My new 980ti hit 90C for a short stint because Nvidia won't kick the fans up to high speed automatically. I had to create a custom fan speed curve. With new SSDs, bigger fans on GPUs, and CPUs not running at "catch on fire" temps, airflow isn't what it use to be. The main thing is to make sure fans aren't competing with each other, read: blowing air at each other.

On this note: overclocked a cheap G3258 (3.2Ghz) to 4.0Ghz with stock cooling. Hit 70C on a stress test. Found my old watercooler backplate and installed the cooler. OCed to 4.4Ghz, won't go above 60C. Really does make a difference and all that heat just gets blown out the back.

- EVGA GeForce GTX 970 ACX 2.0 SuperClocked
339€
This will be combined with an old GTX 670 I have in my current computer

970s do have that crummier last 500MBs of VRAM. I don't know if that would be an issue for rendering because it really wasn't for gaming and I forced some games to fill up all my VRAM when possible and noticed no big FPS drops. Well, if I did it didn't matter because I was well above 60FPS and the games it would drop would drop anyway.

How do you plan to run the multiple GPUs? Does 3dMax let you just dump work off on different cards when needed? Because SLIing them would crater the performance of the 970 as it locked to the old GPU speed. You can run two difference chipsets and dump shit like Physix off on the older card. I had plans to test this last weekend with my 980ti and 970, but I got busy.

I'd go with Skylake simply because it's on Intel's newer 1151 pinset. More upgradeable at a later time, where a Haswell-E is pretty much end of the line for that mobo.

TheFeniX wrote:NOTE: I don't build PCs for rendering,

salm wrote:One thing I am uncertain about is the CPU.
I think either a i7 5820K Haswell-E (395€) or a i7 6700K Skylake (344€) would be a good choice.
The main board I am not certain about either.
for the 5820K Haswell perhaps the ASUS X99-Pro(355€) and for the Skylake an ASUS Maximus VIII HERO (212€)?

I'm going to take a pass on this. If it were gaming, I'd say the 6700 due to the higher clock and newer manufacturing. But more cores for rendering might sell it for the 5820.

That is what I am not sure about, either. I am tending towards the 6700 because it is a bit cheaper and because I mainly do GPU rendering at home these days while creating scenes. After that I do a test rendering on CPU and then send the whole works to the cloud where the final animation gets rendered on CPU.
So the GPU renderer is a constantly used tool and very important to be fast. The CPU render is used in one or several isolated instances. It is nice if it is fast but by far not as important as the GPU.
But then, the 5820 isn't that much more expensive, so, hmm...
Is there no such thing as a 6820 Skylake?

Depends. With a CPU watercooler, you can push any excess heat off the CPU straight out of the PC. I ran 2 MSI 970s overclocked and never had them get above 80C at the highest load. My new 980ti hit 90C for a short stint because Nvidia won't kick the fans up to high speed automatically. I had to create a custom fan speed curve. With new SSDs, bigger fans on GPUs, and CPUs not running at "catch on fire" temps, airflow isn't what it use to be. The main thing is to make sure fans aren't competing with each other, read: blowing air at each other.

On this note: overclocked a cheap G3258 (3.2Ghz) to 4.0Ghz with stock cooling. Hit 70C on a stress test. Found my old watercooler backplate and installed the cooler. OCed to 4.4Ghz, won't go above 60C. Really does make a difference and all that heat just gets blown out the back.

So, are you suggesting a water cooler or simply the stock coolers because these days the components do not create as much heat as in the past?

970s do have that crummier last 500MBs of VRAM. I don't know if that would be an issue for rendering because it really wasn't for gaming and I forced some games to fill up all my VRAM when possible and noticed no big FPS drops. Well, if I did it didn't matter because I was well above 60FPS and the games it would drop would drop anyway.

I know about the last 500MB being useless. It is not that much of an issue, though. My current 670 has only 2GB and the scenes I work on on my home PC have never exceeded even that.
I thought about getting a 680 ti but the price increase from 340€ to 680€ doesn't seem worth it.

How do you plan to run the multiple GPUs? Does 3dMax let you just dump work off on different cards when needed? Because SLIing them would crater the performance of the 970 as it locked to the old GPU speed. You can run two difference chipsets and dump shit like Physix off on the older card. I had plans to test this last weekend with my 980ti and 970, but I got busy.

3ds Max has no GPU renderer. You either use a third party renderer such as Octane or, like in my case, you use Blenders Cycles renderer.
The renderer splits the image into small parts called "buckets". Depending on several circumstances, one bucket is usually 128X128 or 256X256 pixels. Each card renders one bucket at a time and the final image is built up step by step. So if you are unlucky the slower card will grab the last bucket and the faster card will idle in that time. Hence it is not recommended to use two extremely assymetrically powered cards. But the slower card will never slow the faster card down. using a 970 and 670 in combination is fine.
SLI is not recommended as most GPU renderers can not deal (or at least only very poorly) with SLI at this time.

I have also been looking Kingston HyperX SHPM2280P2 SSD. It costs twice as much as a "normal" SSD of that size but writes at 600MB/s and reads with 1400MB/s. That would be interesting for streaming mesh caches and image sequences in post production.
Perhaps it would also make sense to get 64GB of ram. 32 aren't that much these days.

Terralthra wrote:I'd go with Skylake simply because it's on Intel's newer 1151 pinset. More upgradeable at a later time, where a Haswell-E is pretty much end of the line for that mobo.

That's a good point.
To be honest I have never upgraded a CPU but perhaps in this case it would make sense for when the Cannonlakes will be released.

I believe Skylake-E was expected around Q'4 2016 last I heard. Though I am not sure if that was rumored/leaked info was for production or availability.

salm wrote:So, are you suggesting a water cooler or simply the stock coolers because these days the components do not create as much heat as in the past?

Yea, sorry. I tend to get off on rants and don't finish my thought: the advantages to a water-block is the lower operating temp during utilization (usually 10C at a minimum), but also that they tend to connect to your case backplate fans and just jet all that hot air right out the back. So, everything runs cooler to help extend the life of all the parts. I also found mine for $35 and since I never had the stock cooler, it wasn't anymore expensive.

Now, gaming doesn't exactly beat up all your cores. But I assume rendering programs are better at that kind of utilization, so the extra cooling would be even more beneficial. I'd recommend one over the stock cooler..

I thought about getting a 680 ti but the price increase from 340€ to 680€ doesn't seem worth it.

As noted in the hardware thread, my 980ti was a rough buy even at $130 off. Back in the early 2000s, I never paid more than $200 for a card, then slapped a waterblock on it an OCed it until my screen would turn to mush, then dial it back a bit.

But the 980ti crushes the 970 in all areas. It's a beast of a card and really makes the 980 look like a wash comparatively (real expensive, not really that much more powerful). Not exactly surprising though.

Once again, as a guy who has rendered a few pieces for Jedi Academy and Skyrim only, I have no idea where your bottle-neck would even be on something like that. A 980ti might be overkill. But looking at some benchmarks (warning: tomshardware link, but first I found): it's blowing out the 980 and keeping up with the Titan, so I assume the 970 is fairing much worse.

Ah, sorry, I somehow wrote 680 instead of 980 in my last post.

Ok, thank you for the watercooling hint. I will look into that. Extending life of the parts is a good idea and worth spending a few extra bucks as I intend to keep this rig for the next 5 or so years.

I might have to revise my view won the 980ti as well. The benchmarks look very intriguing and I have a couple of unused € in my budget.

Doing some resarch on 3d rendering forums it appears that the 5820 is regarded as better than thte 6700 for 3d purposes because 6 slowwer cores are better than 4 faster cores for 3D. Also the 5280 has quad channel RAM while the 6700 only has dual channel RAM.

salm wrote:Ah, sorry, I somehow wrote 680 instead of 980 in my last post.

You mean the 980ti is 980 eurobucks vs 680? Your price on the 970 is right around what we pay here (~$350) and the 980ti are running around $650. Sure you aren't looking at the Titan? That's still about a $1000 around these parts. At $650, the 980ti is a tough buy as it's about 2x the performance of the 970. But at 3x the cost? Nah.

Doing some resarch on 3d rendering forums it appears that the 5820 is regarded as better than thte 6700 for 3d purposes because 6 slowwer cores are better than 4 faster cores for 3D. Also the 5280 has quad channel RAM while the 6700 only has dual channel RAM.

From reading, Quad channel support is expected in Q3 this year, but not confirmed. Though, I assume that would mean a new board. You seem to be in a rough spot timing-wise.

The Broadwell-E processors are just about to launch (as in, the next few days) so don't even consider the 5820K. Personally I'd go for the 8 or 10 core version if at all possible, with current DRAM speeds having four memory channels for only 6 cores is pretty much a waste, although the Broadwell-E six core will still have over twice the L3 cache of the Skylake which may be quite noticeable in some apps.

Skylake-E has already slipped from Q3 to Q4 and will almost certainly slip to next year, to give Broadwell-E a reasonable sales window if nothing else. This is not a huge loss as the Broadwell to Skylake 'tock' was rather underwhelming in terms of performance gains for popular apps. Really we're waiting for Cannonlake for the next major core count and cache bump (and maybe some clock bumps if we're lucky) and that's probably going to slip to early 2018.

TheFeniX wrote:

salm wrote:Ah, sorry, I somehow wrote 680 instead of 980 in my last post.

You mean the 980ti is 980 eurobucks vs 680? Your price on the 970 is right around what we pay here (~$350) and the 980ti are running around $650. Sure you aren't looking at the Titan? That's still about a $1000 around these parts. At $650, the 980ti is a tough buy as it's about 2x the performance of the 970. But at 3x the cost? Nah.

From reading, Quad channel support is expected in Q3 this year, but not confirmed. Though, I assume that would mean a new board. You seem to be in a rough spot timing-wise.

[/quote]
Hehe, lol. Sorry for causing the confusing, I'm a bit overworked with the most boring contract at the moment.
The 980Ti costs around 680€ (taxes included).

Starglider wrote:The Broadwell-E processors are just about to launch (as in, the next few days) so don't even consider the 5820K. Personally I'd go for the 8 or 10 core version if at all possible, with current DRAM speeds having four memory channels for only 6 cores is pretty much a waste, although the Broadwell-E six core will still have over twice the L3 cache of the Skylake which may be quite noticeable in some apps.

Skylake-E has already slipped from Q3 to Q4 and will almost certainly slip to next year, to give Broadwell-E a reasonable sales window if nothing else. This is not a huge loss as the Broadwell to Skylake 'tock' was rather underwhelming in terms of performance gains for popular apps. Really we're waiting for Cannonlake for the next major core count and cache bump (and maybe some clock bumps if we're lucky) and that's probably going to slip to early 2018.

Oh, thank you, that is extremely good to know.
8 or 10 cores sounds nice. I guess they are going to be rather expensive in the beginning but eh, if nothing else the existance of 8 and 10 cores should bring down the cost of 6 cores.

If I were to build a new PC right now, I'd go with an i5-6600K Skylake just to replace it with the Cannonlake-E whenever that comes out and I had the cash to upgrade. No sense locking into a finished socket.

Terralthra wrote:If I were to build a new PC right now, I'd go with an i5-6600K Skylake just to replace it with the Cannonlake-E whenever that comes out and I had the cash to upgrade. No sense locking into a finished socket.

The -E versions of the processors are on a completely different socket/motherboard to the basic desktop chips. They're essentially cut-down server parts with lots of cores, lots of cache and no integrated graphics. You can't upgrade a current Skylake chip to anything-E without buying a new motherboard and probably memory as well. Unless you meant replacing the whole PC.

The socket for Skylake-E (LGA 2017-A) is supposedly already available on the Asus X-99E, which is also compatible with Haswell-E & Broadwell-E, although I haven't done detailed research on this as I usually stick to Xeons.

Mea culpa. I thought the -Es were on the same LGA (1151).

Terralthra wrote:If I were to build a new PC right now, I'd go with an i5-6600K Skylake just to replace it with the Cannonlake-E whenever that comes out and I had the cash to upgrade. No sense locking into a finished socket.

The i5s don't do hyperthreading so they are not as useful to me as i7s.

salm wrote:

Terralthra wrote:If I were to build a new PC right now, I'd go with an i5-6600K Skylake just to replace it with the Cannonlake-E whenever that comes out and I had the cash to upgrade. No sense locking into a finished socket.

The i5s don't do hyperthreading so they are not as useful to me as i7s.

Do you get any meaningful benefit out of hyperthreading? It might be useful on 2 or 4 core machines where there aren't a lot of execution resources, but I find the benefit to be marginal or even negative on 8 core plus machines.

Tests show that you get around 5% to 30% percent performance increase with hyperthreading when rendering scenes in raytracers such as VRay or Mental Ray on quad cores. In some cases it can be slower due to special circumstances but usually it is faster.
I have no idea if this is any different for 6 or 8 core machines.

salm wrote:Tests show that you get around 5% to 30% percent performance increase with hyperthreading when rendering scenes in raytracers such as VRay or Mental Ray on quad cores. In some cases it can be slower due to special circumstances but usually it is faster. I have no idea if this is any different for 6 or 8 core machines.

Classic ray-tracing is branchy float code with a unpredictable memory access pattern so that makes sense. Everyone else writing graphics and physics engines has restructured the algorithms for maximum vector parallelism and pipelined memory access, but if you're going to do the classic single ray at a time btree intersect test then the (per-thread) IPC is inherently going to be poor.

Starglider wrote:

salm wrote:Tests show that you get around 5% to 30% percent performance increase with hyperthreading when rendering scenes in raytracers such as VRay or Mental Ray on quad cores. In some cases it can be slower due to special circumstances but usually it is faster. I have no idea if this is any different for 6 or 8 core machines.

Classic ray-tracing is branchy float code with a unpredictable memory access pattern so that makes sense. Everyone else writing graphics and physics engines has restructured the algorithms for maximum vector parallelism and pipelined memory access, but if you're going to do the classic single ray at a time btree intersect test then the (per-thread) IPC is inherently going to be poor.

I think that is just what current non-realtime renderers used in film production do. There's no practical alternative besides GPU rendering.

So some new components have come (i7 6800K and GTX 1080) out and I put together a rig I´m thinking of buying. I am not sure yet on the RAM. Are the timings, CL14 16-16-31, a good combination with the mainboard and CPU in question?

Does this make sense? Perhaps I could wait some more and get the GTX 1070 instead of the 1080 which is supposed to cost only around 420€. Or I could get two 1070s. for around the same price as on 1080.
I guess the PSU is a bit overkill but I can get it from a friend for 100€ cheaper than retail price, so, why not.

All prices inlcude 19% VAT.

- Intel Core i7-6800K €450,-
- EVGA GeForce GTX 1080 Founders Edition €789,-
- RAM: 'Corsair DIMM 64GB DDR4 2400 Quad-Kit Vengeance LPX, CL14 16-16-31 €269,-
- SSD: Kingston HyperX SHPM2280P2H/240GB €184,-
- HardDisk: Toshiba DT01ACA200 2TB €67,-
- CPU Cooler: Alpenföhn Brocken Eco €33,-
- Case: Corsair Carbide 200R €56,-
- Case Fan: be quiet! Silent Wings 2 140mm €18,-
- Mainboard: Asus X99-PRO/U3.1 €323,-
- PSU: Thermaltake Toughpower DPS G Platinum 1200W ATX23 €180,-

Total: €1918,- + €450,- VAT = 2368,-

What resolution are you planning to game at? The 1080 is significantly more powerful than the 1070, but both are good enough for 60 FPS at 1440P and neither will run 4K well without SLI. Personally, I am waiting for the new $199 AMD card but my budget is a lot smaller than yours.

Arthur_Tuxedo wrote:What resolution are you planning to game at? The 1080 is significantly more powerful than the 1070, but both are good enough for 60 FPS at 1440P and neither will run 4K well without SLI. Personally, I am waiting for the new $199 AMD card but my budget is a lot smaller than yours.

From the OP:

Hi,
I am planning to build a new rig and am undecided on several components.
It will be a computer mainly for professional 3D work (3ds Max and Blender) and the computer should be fit for CPU and GPU rendering, even though GPU rendering is more important.

Arthur_Tuxedo wrote:What resolution are you planning to game at? The 1080 is significantly more powerful than the 1070, but both are good enough for 60 FPS at 1440P and neither will run 4K well without SLI. Personally, I am waiting for the new $199 AMD card but my budget is a lot smaller than yours.

Gaming is not really a deciding factor.
Even my current system (i7 2600k, GTX 670) runs every game I am interested just fine for my needs.

I'll be getting a set of VR goggles when prices drop, probably the Vive, which I'd like the system to be able to handle but that shouldnt be problem for neither the 1080 nor the 1070.
AMD cards are unfortunately not an option because GPU renderers require or at lest work significantly better with CUDA cores.

salm wrote:Does this make sense? Perhaps I could wait some more and get the GTX 1070 instead of the 1080 which is supposed to cost only around 420€. Or I could get two 1070s. for around the same price as on 1080.

I still haven't found any good information on benchmarks for something like "2 970s vs 980ti for blender." I always find single-card comparisons for rendering since everyone jumps to "NO SLI FOR RENDERING" which is not what I'm looking for at all. In a gaming setup (SLI): the ONE more powerful card will nearly always outperform SLI.

However with split rendering, you would think you'd double (or halve) your rendering speed. Even more so considering the CUDA count of the 1070 vs the 1080. The 980ti has more CUDA cores at a much lower clock. But, and once again not a rendering guy, CUDA counts are pretty important for rendering from what I've been reading. How much more than Mhz, I do not know.

I would assume that, with proper setup (from reading on the blender forums, updates will need to be made) that two 1070s with correctly setup split rendering (I don't even think that's the right term as I believe it applies to graphics rendering, such as for video/games, but anyways) should outperform a 1080 by a large margin. But man, this shit gives me a headache.

I guess the PSU is a bit overkill but I can get it from a friend for 100€ cheaper than retail price, so, why not.

Way overkill, Even running SLI, you could get away with a 600W. 500W seems to be the bare minimum, 600W would give you room to breath. But at $100 off retail, it would be a good buy.

I still recommend a watercooling block with detached fan, such as this type considering the type of CPU utilization you are shooting for. However, I got mine on sale for $35~, so it was a given. The cooler you have specced out should have no issues if you setup good airflow.

GPGPU scaling including rendering codes is usually close to linear for two or three cards, assuming a top-of-the-range CPU (and why wouldn't you if you're spending that much money). More gets problematic for various reasons (CPU saturation, bus saturation, context switch overheads in various different layers) and requires expert system-level optimisation to get good scaling. I can say from experience if you want to do anything useful with 8 GPGPU cards (in the same machine) you pretty much need a dual xeon server board.

As for SLI, we're on the cusp of GPU manufacturers finally fixing their crappy multi-card graphics scaling solutions so that they actually work properly with a single large pool of compute cores and memory, i.e. the way CPUs do, rather than the shared nothing architecture we currently have. Unfortunately this is not the generation in which that happens; Nvidia actually backed away from SLI a little on the 1080/1070 by deprecating (although not completely disabling) three and four card link-ups. Maybe AMD will step up and fix this. This is made even worse by everyone cheerleading alternate frame rendering and neglecting split frame rendering just because it posts higher headline framerate numbers. AFR is awful for VR, both the latency and the microstutter; until we get true resource pooling, SFR is really the only sane solution for multi-card VR output. Having one card handle each eye is of course a special case of SFR with a static split (that should work pretty well).