The PCIE Lane Misconception?
JCThomas
Posts: 254
Howdy all.
I recently broke down my main workstation, an X99 workstation with 3 Titan X (Maxwell) and 64 GB RAM. I upgraded (downgraded?) my rig to a new Ryzen rig: Ryzen 7 1700, 32GB RAM, a single Titan Black, and a single Titan X Maxwell. I’m running this on a B350 motherboard from gigabyte.
My motherboard has 2 PCIE slots, an X16, and an X4. I have the titan black on the top x16 slot, driving the monitors, and the Titan X on the bottom, x4 slot, dedicated to Octane. I’m getting phenomenal speeds from the Titan X. The PCIE lane breakdown and speed has no effect on render time. My Titan X is a champ, it’s got a 79% ASIC quality, so I’ve got it overclocked 300 MHz, and I’m scoring 147 in Octane Benhmark.
I bring this up only to point out that a motherboard, CPU, and chipset, with dozens and dozens of lanes, may not be worth the cost. I’m fortunate to have a Microcenter nearby, and I got this CPU and motherboard together for $263 after rebate.
Thoughts?
Daz 3D is part of 
Connect
DAZ Productions, Inc.
224 S 200 W, Suite #250
Salt Lake City, UT 84101
Licensing Agreement | Terms of Service | Privacy Policy | EULA
© 2024 Daz Productions Inc. All Rights Reserved.
Comments
Yeah, add that to the long list of MANY misconceptions in the tech community. A lot of people think "more is better", without really knowing or understanding the details.
Other misconceptions:
They're not.
At least not from any price/performance comparison vs. a desktop.
Price/performance is a separate issue, and applies to every single piece of equipment you buy.
The misconception regarding laptops is that they always get too hot to do rendering because they don't have enough space for cooling. As a general statement that's not true. While you may find a piece of junk laptop that overheats during heavy use, the same applies to desktops. Decent equipment is designed for continuous service throughout its design life.
By the way, as an example look at the Founders Edition GPU's, with the GPU inside a box with a single small fan. Not much different than a GTX 1060 inside a laptop. And those work fine. Only if you operate it outside its design parameters will you have a problem.
Yes, i believe it has been proven that PCI slots do not have a major effect on rendering. Or at least, from what i have read, no one has been able to determine much difference between x8 and x16
that does not mean that x4 will not suffer somewhat.
i am curious though, if you have an x16 slot and an x4 slot, why not put the card doing all the work in the x16 slot? seems silly to have the card only being used to drive your displays in the x16 slot and the card doing all the heavy lifting in the x4 slot.
If nothing else, switching them around would be a good test to see if the slot makes a difference in your benchmarks
Yeah, that will be a good experiment. I'll update once I do it. As for having the stronger card in the x4 slot...it's so that the impact rendering has on the rest of my system is minimal. Also, I don't have to use any of the render cards VRAM to drive the monitors. And even though it's getting old, the titan black is no slough. I used to get a lot of video stuttering if I was watching something, even just navigating chrome or edge really, while I was rendering something on the GPUs. Anyway, I'll let everyone know if there's any difference when I switch slots.
Puget has a great article about the difference between x8 and 16, and in their tests the x8 render was actually neglibly faster. Even without switching slots to know for sure, it's still significant to know that the average Titan X Maxwell scores 127 on Octane Bench, and I've got mine to 147 on an x4 slot.
Actually, RAM speed does matter a LOT on Ryzen machines. The CCX interconnect called Infinity Fabric runs at the same speed as your RAM. So your machine will actually get faster overall with faster RAM. It's quite significant, even to the average user.
As for water cooling, there is indeed no difference anymore between air and water cooling for the most part depending on the quality of the components you get. Where water cooling is better is when overclocking. So there are still cases where water cooling is the way to go.
As for PCIe lanes, it can make a difference if you do SLI gaming. Some games will run faster with full x16 slots. I doubt you could really notice it on your own though.
What you are seeing with speed is not PCIe lane dependent. If you where too move the Titan X over to the 16x slot, your initial render start time would be faster. The render itself is handled on the GPU, so there is not much traffic going over the PCIe bus.
Now on the flip side: If you are playing a current AAA game at 8K resolution with dual Titan XP GPUs running SLI, you will probably see a noticable FPS drop if you go from 2 16x slots down to 2 4x slots.
My point is that MANY in the tech community and online computer forums think that if a particular technology is newer, and/or the spec numbers are bigger, then automatically it's better. So obviously x16 is better and faster than x4 because obviously 16 is bigger than 4. And a 4Ghz CPU is much better than a 3 Ghz CPU because 4 is bigger.
Stuff like "does my particular computer and software take advantage of that" become irrelevant complications. And they frantically search for specific exceptions where in fact it may matter, or maybe it results in a tiny 2 point increase in performance, and use that to "prove" their point. I see that in some tech forums all the time. And then they'll run off and tell everyone they need faster RAM or a 16 core CPU solely because they saw the first 4 minutes of a tech video that showed a graph of slightly higher performance in an irrelevant benchmark.
Unfortunately, stuff like this is very complicated, and it depends.
Interesting stuff. What I'd like to know, however, is whether faster pcie lanes make a difference when you're rendering on mulitple GPUs. I'm planning on replacing my render box in just over a year's time with one using my existing 1080, plus a new 1080ti.
Cheers,
Alex.
Probably not too much, it's changed I think with the increasing speeds of the PCI spec, now at 3.0. Not an apples to apples comparison, but my i7 6700K with two 1080 Ti's would start rendering a scene faster than my Threadripper 1950x with two Titan X Pascals. The i7 6700K has a faster core clock, but PCI 3.0 8 lanes for each card vs 16. But a few other things that are probably a factor - the memory speed and GPU was clock on the Titans, the mem speed is like 1K slower and the core clock lower. Also the Threadripper had either a failing SSD in it, or loading Daz content from a 7200 RPM hard drive.
Once it loads the scene into the card's memory and starts rendering I doubt there's much difference, but I could be wrong.
Thanks for your response but I think the issue is the communication between the two parts of the iray render, running on seperate GPUs, rather than the load speed. I should have made this a little clearer in my original post.
Cheers,
Alex.
So you're thinking that in a scene that takes, say, 10 minutes to render because the GPU's are doing all the zillions of calculations of light paths, that there's enough discussion between the GPU's on the PCI bus to make a significant difference? I'm not sure why you think that's an issue. Even if it was, I doubt you'll find an all-in-one answer. I suppose you could find a PCI performance monitoring app and see if there's a lot of talk between GPU's during a render, but I'm guessing if there's little or no performance difference due to PCI speeds with transfer to/from the CPU, the same applies for between GPU's.
Well in the beginning stages of benching and stabilizing my overclock, I accidentally left my Titan Black running for a run of Octane Bench, when I was only boosting the Titan X by 50mhz or so. It scored 220, which is exactly in line with the score of a single titan black and a single titan x.
You're talking about absolutes, with the exception of water being better; it can be cooler, it can also leak; air leaks don't usually cause issues. :) Cooler is better, but it doesn't mean you need water, or even Nitrogen if you want to take it to extremes. But return for cooler is the clasical diminishing-returns.
"Oh and laptops aren't good for rendering"; yes they can render, but that doesn't make them a good choice. Folks that buy a laptop then use if for rendering are doing so becuase the advantages of a laptop v desktop are more important than the negatives for rendering... Or they don't know the benefits of a desktop for rendering, and associated benefits.
It is about choice... As well as a saleman's advice (just as suspect as random friend, forum poster, dude in the pub).
And that's one of the misconceptions. Below a certain point, cooler is irrelevant. But people think that if 120C is damaging to your CPU, then of course you should keep your CPU at 10C all the time or else it will die next week. That's the misconception.
The truth is, if you keep your components within their continuous design rating they will last for their design life. Unless they are junk.
If its continuous rating is 80C, then adding a refrigerator compressor and 55 gallon liquid nitrogen tank to keep it at -30C below ambient isn't going to make much difference. Your computer will likely be obsolete before your components die. And you will have wasted a ton of money.
I think your test is not relevant and you actually have no numbers to backup your assumption.
You need to do the benchmark on a big scene that involves a lot of transfert between your CPU's memory and GPU's memory.
I think a real time viewport benchmark at a high resolution (4K) with real time modifications would be more significative. It also doesn't have to be iray but could be an OpenGL viewport test provided there are enough data going back and forth so that the transfert time would be significant (thus the big scene in high res). From there you could test the influence of the Pcie speed and/or your memory speed
If your point was to show that PCIe speed is not that much important for a one shot rendering (without realtime feedback) then it's OK. Once textures and geometries are in the VRAM, the PCIe doesn't play a big role
The only thing I'll add is that only the best actual Gfx card can saturate the PCIe in 8x (I don't think your card fit in that category). Have to wait the next gen to see if the 16x can be fully used.
These are not misconceptions but rather good general advice. It's good to have room rather than to be limited just because of one factor. Your POV is about building a Rig that just fit a particular need at a point in time. You don't take in account that software, hardware or needs may change. We're not talking about the ice cream you'll be eating within 5 min and forget about it
Software and Hardware are not always on par. You may have a 8 core processor but most applications are only optimized up to 4 cores. That is today's state but may not be the case in two years. With your reasonning, we also don't need bigger disk, but with the cost per GB going down we've seen that softwares are taking more and more space and new use like home archiving, NAS,etc have appeared
If you were right then you could go with an i3 with just one PCIe slot and live with that forever instead of buying a Ryzen 8 core. The reality is that more is better because you have enough power so that your hardware doesn't get in your way to achieve what you want to do and whichever software you use, and/or in case of a need of an upgrade, you don't have to buy a new PSU just because you want to buy the brand new GFX card for ex
We've seen quantity of people who bought prebuilt HP/Dell/whatever PC that couldn't be upgraded. Evolutivity is a key point you seem to ignore
You're also ignoring that all these advices are for people who want to get the most performance from their Hardware. In that optic, Watercooling is the way to go. Even for the average user, the advantage is less noise and a better heat dissipation, even without going to extreme. In my POV it's about bringing the best tech to the mass. Watercooling was reserved to the most techie some years ago. Not the case today with the AIO solutions. The same goes with other techs like SSD
More is better because it brings more power and technologies for the mass at a reasonnable price. When 8 core CPU becomes the norm, next step will be 16 cores. Finding a use to theses cores is an other debate but you have to wait for softwares to catch up
Going back to the PCIe thema, today, most users really only need at best two PCIe at 8x for Gfx cards. But you don't know how the future is going to be. If the future of 3D is real time render (blender Evee for ex) the PCIe and RAM speed may have a bigger impact.
Takeo.Kensei, I'm only suggesting that people find out what users really need before recommending without understanding those needs or the technology. Anyone can say to get the best and fastest CPU and GPU and liquid cooler and RAM and power supply. That's easy. But it doesn't help the people who have to come up with money to spend, and make an intelligent choice based on their particular needs.
I bought an 8 core, 16 thread Ryzen that I'll probably never need, because the technology is going toward GPU's, and 90% of my apps don't care about all those threads. Sure, something may happen and suddenly an app will come by that makes use of it. But that's up to me to decide if I want to take that risk or not. And if I'm going to recommend something to someone, I want to find out what they will need, and give them the options and some sort of cost/benefit so they can decide. But first I need to know their needs, and know the technology well enough to tell them the cost/benefit.
Unless you want to spend time analyzing each software for each user in search of advice with every particular use case, general advice apply. At some point the user has to make some research and decide for himself/herself. From what I've seen, advice are usually within the requester's budget and fit his/her need and that is what counts. No need to go for an in depth analysis
So I guess the best response to anyone buying a new computer is buy a Threadripper with 128GB RAM and a Titan V and 4 PCI x16 slots and a 1250 watt power supply and 6 SSD's. That should cover it
No, you forgot the water cooling. Also, definitely go dual on the video card.
Yeah good point. Definitely dual cards for future-proofing, and to get that extra 6.7% improvement in render times. And two Titan V's means we're talking liquid nitrogen cooling and at least a 2,000 watt power supply. That will assure the GPU's run at -20C below ambient when rendering. And of course we don't stress the power supply. Sorry, I meant dual power supplies for a beast like this. Who knows when we'll get 2 more Titan V's.
I agree. The analysis should be nothing more than "Um, what software do you plan to use". That's it. Assuming that anyone asking about what computer to buy isn't going to be overclocking (or else they'd probably already know), it pretty much comes down to what they plan to use it for.
And unless they plan to use some of the handful of apps that presently rely heavily on CPU's (some video editing stuff, engineering/science apps, ZBrush I suppose, some video games, and a few others), then they should focus on GPU's since this is a DAZ Studio forum and they'll probably be doing rendering.
They should ignore PCI buses, except to make sure they have enough slots if they want to double up on GPU's or something. They should ignore RAM speeds (except to make sure it's on the qualified vendor list for the board), and probably get 8-16 GB of RAM, depending.
They can probably get a mid- or low-range CPU (see above) and probably don't need a high end motherboard. The power supply should probably be in the 500-650 watt range, unless they might double up on GPU's in the future. And even then that might be fine (my GTX 1080ti plus GTX 1070 can barely hit 420 watts at full, and I'd have to buy another 1080ti to get up even close to 650 watts).
SSD's are nice and fast when booting up and loading and other OS stuff, if that's important.
They should forget about water cooling since they won't be overclocking.
What GPU should they buy? Well take a look at the attached spreadsheet that summarizes the benchmark Iray results and price/performance. I used the GTX 1060 as a base for comparison. Basically, if your main task is Iray rendering, get as much GPU as you can afford.
Personally, for most users here I think that's a reasonable starting point, and will change based on their particular software requirements. But I think it removes some of the many misconceptions about what most users here really need based on their workload. And allowing for the fact that not only might users need to upgrade in the future, they also might give up this hobby altogether next year and move on to something else that has nothing to do with computers and GPU's.
Yes I said that. Agreed, keep it below its rated max, including allowing for brief spikes, then its probably epeen, or one just enjoys the processes.
Personally, I like silent (not quit), so I'd rather sacrifice some coolness if required; as long as my components are opperating within their designed tollerances. I don't tend to overclock - way too much noise, or it requires more cooling to counteract it, and for rendering, when there could be hours of running, then there is definitely the risk of shortening the life of components.
... As ever, it is about choices.