rMBP Thunderbolt GPU Good?

[IceHot]

Hey!

So im thinking about replacing my windows desktop with a rMBP, but the gt650 bothers me a bit, meanwhile iv been looking around for external gpu solutions that uses the thunderbolt standard but havent found any... yet!!?? 🙁 my question is: how much will a thunderbolt gpu improve the already not to shabby gaming performance on a rMBP, and will it be able to use both ports for extra "bandwidth". If it was possible to push the rMBP 3dmark06 score over 22000 id be all over it!! fo sure!

 

[leo.andres.21]

Well, going from desktop to a thin and light laptop, you have to at least make that compromise. Though GT650 is a really good card, it easily outperforms the previous gen of 6750-70 like mine, benchmark was doubled, so it is a very capable card for a notebook this thin and light.
There were some thread that shows overclocking the 650, you should search for it, hopefully it'll help

 

[IceHot]

FYI my pc has 4gb ram amd x4 cpu...2-3 years old, and a 6950 1gb video card... about 17-18k 3dmark 06..

 

[leman]

Hey!

So im thinking about replacing my windows desktop with a rMBP, but the gt650 bothers me a bit, meanwhile iv been looking around for external gpu solutions that uses the thunderbolt standard but havent found any... yet!!?? 🙁 my question is: how much will a thunderbolt gpu improve the already not to shabby gaming performance on a rMBP, and will it be able to use both ports for extra "bandwidth". If it was possible to push the rMBP 3dmark06 score over 22000 id be all over it!! fo sure!

Well, TB is PCI-e 2.0 4x while GPUs are usually installed in a PCI-e 3.0/4.0 16x slot... so high performance gaming is out of question anyway. Besides, the rMBP is able to run games exceptionally well for the laptop its weight. I am getting 40+ fps with Guild Wars 2 on high settings at 1920x1200 and the gameplay is very smooth.

Anyway, using google gave me this: http://www.tomshardware.com/reviews/pci-express-graphics-thunderbolt,3263.html

Ridiculously expensive, but maybe its what you are looking for?

 

[IceHot]

Well, TB is PCI-e 2.0 4x while GPUs are usually installed in a PCI-e 3.0/4.0 16x slot... so high performance gaming is out of question anyway. Besides, the rMBP is able to run games exceptionally well for the laptop its weight. I am getting 40+ fps with Guild Wars 2 on high settings at 1920x1200 and the gameplay is very smooth.

Anyway, using google gave me this: http://www.tomshardware.com/reviews/pci-express-graphics-thunderbolt,3263.html

Ridiculously expensive, but maybe its what you are looking for?

Hmm, my biggest concern in terms of gaming is that i wont be able to play elder scrolls online...But if gw2 runs that great i see no reason why eso shouldnt run at an acceptable framerate.

But if thunderbolt is 4x, and the retina has 2 thunderbolt ports, shouldnt it be possible to run an egpu at 8x if both ports are used? 😛

 

[bogatyr]

Well, TB is PCI-e 2.0 4x while GPUs are usually installed in a PCI-e 3.0/4.0 16x slot... so high performance gaming is out of question anyway. Besides, the rMBP is able to run games exceptionally well for the laptop its weight. I am getting 40+ fps with Guild Wars 2 on high settings at 1920x1200 and the gameplay is very smooth.

Anyway, using google gave me this: http://www.tomshardware.com/reviews/pci-express-graphics-thunderbolt,3263.html

Ridiculously expensive, but maybe its what you are looking for?

You are incorrect in high performance gaming. The TB is faster than an ExpressCard 1.0 which can give high performance gaming in external solutions. The best resource for eGPU may be found at:
http://forum.notebookreview.com/e-gpu-external-graphics-discussion/418851-diy-egpu-experiences.html

Nando4 keeps everything up to date in the first post and helps out hundreds of users in that thread. My eGPU on my X220 performed extremely well using EC 1.0.

 

[Mac32]

TB is definately an useable option, but powerful graphics cards will be bottlenecked to some degree. This obviously means that with coming graphic cards (like Maxwell in 2014) the bottleneck will become greater and greater.

What about using a dual-cabel TB connection for future external graphic cards? That could double the bandwidth maybe? Or does both TB connections in the rMPB use the same TB chip? If not, this could really improve graphics performance in demanding games. AFAIK present external graphics card only use a single TB cabel connection. However what about lag (vs. input lag) when using external graphics solutions?

 

[IceHot]

TB is definately an useable option, but powerful graphics cards will be bottlenecked to some degree. This obviously means that with coming graphic cards (like Maxwell in 2014) the bottleneck will become greater and greater.

What about using a dual-cabel TB connection for future external graphic cards? That could double the bandwidth maybe? Or does both TB connections in the rMPB use the same TB chip? If not, this could really improve graphics performance in demanding games. AFAIK present external graphics card only use a single TB cabel connection. However what about lag (vs. input lag) when using external graphics solutions?

Well the retina has 2 thunderbolt ports, could they be used to double the bandwidth if the eGPU supports it? :/

 

[bill-p]

Well the retina has 2 thunderbolt ports, could they be used to double the bandwidth if the eGPU supports it? :/

Doesn't work that way. You don't see a video card that plugs into 2x PCIE 8x slots for a reason.

The 2 TB ports are on separate bus, and if the GPU needs to transfer something at 6x speed, then it can't just split the load into 2 3x and transfer it over the 2 ports. It has to transfer all 6x at the same time.

 

[Mac32]

Doesn't work that way. You don't see a video card that plugs into 2x PCIE 8x slots for a reason.

The 2 TB ports are on separate bus, and if the GPU needs to transfer something at 6x speed, then it can't just split the load into 2 3x and transfer it over the 2 ports. It has to transfer all 6x at the same time.

I see (almost figured that), but why can't the signal go at 6x at both ports in parallel? Do you meen that there's only one chip controlling both TB ports, and that only one port can be used at max speed at a time?

 

[DTKblaster]

PCI-E 2.0 x4 isn't all that bad really, when I tested my Radeon HD5870 in it, performance was only reduced by about 5% in benchmarks and game frame rates compared to the 16x slot on my mobo.

 

[JD92]

Doesn't work that way. You don't see a video card that plugs into 2x PCIE 8x slots for a reason.

The 2 TB ports are on separate bus, and if the GPU needs to transfer something at 6x speed, then it can't just split the load into 2 3x and transfer it over the 2 ports. It has to transfer all 6x at the same time.

Why not? Why can something like SLI not be implemented, just using two Thunderbolt ports instead of two PCIe ports?

 

[bill-p]

I see (almost figured that), but why can't the signal go at 6x at both ports in parallel? Do you meen that there's only one chip controlling both TB ports, and that only one port can be used at max speed at a time?

Well, think of it this way:

You can run at 3mph.
Your friend can run at 3mph.

If you and your friend combine, then you have 2 people running at 3mph, not one person running at 6mph.

One person running at 6mph is still faster than both of you guys.

Does that make sense?

It's the same situation here. The problem is that you have two ports running at half speed. They don't just stack up bandwidth like that. Just like you're not guaranteed 2x processing speed by having 2 cores as opposed to 1.

The "ideal" situation is that they stack up to be 8x TB, but it'll be with something like SLI and Crossfire rather than with a single card running at 8x. And since we don't see the "ideal" situation often with SLI and Crossfire, a single card running at 8x is still faster than SLI at 4x.

Why not? Why can something like SLI not be implemented, just using two Thunderbolt ports instead of two PCIe ports?

SLI can definitely be implemented. But it's not going to be as fast as a single 8x bus. It'll be like having 2 4x PCIe cards running in SLI. You'll still have much better performance with a native 8x card.

And the built-in 650M in the rMBP does run at 8x.

 

[Mac32]

Thanks, Bill-p but your analogy doesn't quite make sense to me. 🙂 I'm sure you're right, but the bandwidth would be doubled if you could use both connections at regular speed. To use a car/road analogy: The cars would travel at the same (ie. full) speed, but there would be twice the number of lanes, so more cars could pass by within a specific interval of time.
Again, if each thunderbolt connection had its own chip, why couldn't this be feasible? Unless the mainboard doesn't have the ability to receive and transmit such a large amount of data that is.

 

[lixuelai]

PCI-E bus has rarely been the limitation which is probably why splitting the data has never been explored. That said though a more realistically doable option is getting 2 graphics cards connected via Thunderbolt and use SLI/CF.

 

[bill-p]

Thanks, Bill-p but your analogy doesn't quite make sense to me. 🙂 I'm sure you're right, but the bandwidth would be doubled if you could use both connections at regular speed. To use a car/road analogy: The cars would travel at the same (ie. full) speed, but there would be twice the number of lanes, so more cars could pass by within a specific interval of time.
Again, if each thunderbolt connection had its own chip, why couldn't this be feasible? Unless the mainboard doesn't have the ability to receive and transmit such a large amount of data that is.

Well... even if you have two lanes, you don't have a faster car.

Think of the car as the TB port. The passenger sitting in the car is the graphics data. You have 2 cars, so you can accommodate 2 passengers.

The "cars" are limited to 4x speed. If you have a single passenger that wants to get through the road at 8x speed, then neither of the car would be able to do that. The cars can take 2 passengers at 4x speed, though.

Now, the problem is... you can't get any passenger there at 8x speed. That ain't gonna happen even if you split the passenger in half. You can only get 2 passengers there at 4x speed each.

The key point here is "speed".

 

[Mac32]

OK. 🙂 I guess I'm gonna have to buy a dedicated PC gaming rig and a thunderbolt display some day. As it seems thunderbolt just isn't fast enough for hardcore gaming (high gfx settings, big resolution, modern games), and the difference should be especially noticable when Maxwell comes out in 2014. External tb graphics cards are just too expensive and not powerful enough. I loved the idea myself at first, but unfortunately tb has its limitations it seems.

 

[lixuelai]

Doesn't work that way. You don't see a video card that plugs into 2x PCIE 8x slots for a reason.

The 2 TB ports are on separate bus, and if the GPU needs to transfer something at 6x speed, then it can't just split the load into 2 3x and transfer it over the 2 ports. It has to transfer all 6x at the same time.

Actually it does work that way. This applies to your subsequent posts but I am just going to quote this. To use the example you used x6 means 6 PCI-E lanes. So it is akin to a 6 lane highway. Does 2x3 lane highway have the same throughput as a 6 lane highway? Yes. Your error lies in that you didn't realize individual PCI-E lanes all have the same throughput (ex 500MB/s for 2.0) and the numbers associated indicate how many lanes there are, not how fast an individual lane can transfer data.

Anyway as to why has no graphics cards ever used 2x PCI-E slots etc it is likely because PCI-E bus has rarely been a limiting factor in graphics performance.

 

[bill-p]

Actually it does work that way. This applies to your subsequent posts but I am just going to quote this. To use the example you used x6 means 6 PCI-E lanes. So it is akin to a 6 lane highway. Does 2x3 lane highway have the same throughput as a 6 lane highway? Yes. Your error lies in that you didn't realize individual PCI-E lanes all have the same throughput (ex 500MB/s for 2.0) and the numbers associated indicate how many lanes there are, not how fast an individual lane can transfer data.

Anyway as to why has no graphics cards ever used 2x PCI-E slots etc it is likely because PCI-E bus has rarely been a limiting factor in graphics performance.

Well, again, it does not work that way.

I'm well aware that 6x means 6 lanes for a PCIE slot, but in the same sense, you really can't stack 2 3x slots together and call them a x6 slot. It's like saying you can take 2 different train tracks that go to different stations and make it so that all trains going through those tracks end up at the same station.

Yes, it's possible... but the technology to do that is unnecessarily complicated (you need a linker) and it's just to solve a problem that shouldn't be there in the first place (if you have access to a single x6 slot). That's why it doesn't happen.

Similarly, they'd need a dedicated chip and some pretty elaborate hardware-software interplay in order to make 2 x4 slots stack together as a single x8 slot. The whole thing is so hard to achieve (especially since you know how Apple does not play nice with features they don't condone) that I might as well write "it doesn't work that way". Not that it's impossible, but that nobody would bother.

Is it clearer now?

Otherwise you'd think they'd have come out with a Thunderbolt to x8 adapter already:
http://www.sonnettech.com/product/xmacminiserver.html

Edit: or let me put it a different way: since the discussion involves an Apple device, of which the drivers are partially (or totally) controlled by Apple, it might well be "impossible" to achieve unless Apple themselves start supporting the idea.

Even then, it's a big mess to solve a "problem" (that should not exist) for a small niche.

 

[Mr MM]

ok lets be clear

There is a single pcie x16 connection that is shared between the entire mbp

x8 goes to the gpu

x4 goes to the thunderbolt chip, which in turn it can divide or not the signal between the 2 ports

x4 is used for other peripherals.

all of that equals to x16.

 

[Saladinos]

Doesn't work that way. You don't see a video card that plugs into 2x PCIE 8x slots for a reason.

The 2 TB ports are on separate bus, and if the GPU needs to transfer something at 6x speed, then it can't just split the load into 2 3x and transfer it over the 2 ports. It has to transfer all 6x at the same time.

SLI uses two GPUs in different PCI slots. Theoretically you could have two lower-powered cards in SLI and see a benefit.

Yes, SLI is a bit of a hack though.

 

[derbothaus]

3DMark06 is a stupid test for todays graphics cards. It has completely plateaued. A Radeon 9800 will score similar to a GTX 280 that will score similar to a GTX680. Now run the tests in Vantage or 3dMark11 and the 9800 will stagger and the GTX680 run run over everything.
The 650m in the retina is shown to be faster than a GTX660 as it is heavily overclocked to 900MHz on the core. The problem is in the massive pixel needs. If you don't mind scaling it will fit most all your needs in a laptop compromise as Razer, Alienware and other "Gamer" laptops offer the same spec in their portables. Of course heat and extended playing may be a consideration.

 

[lixuelai]

Well, again, it does not work that way.

I'm well aware that 6x means 6 lanes for a PCIE slot, but in the same sense, you really can't stack 2 3x slots together and call them a x6 slot. It's like saying you can take 2 different train tracks that go to different stations and make it so that all trains going through those tracks end up at the same station.

Yes, it's possible... but the technology to do that is unnecessarily complicated (you need a linker) and it's just to solve a problem that shouldn't be there in the first place (if you have access to a single x6 slot). That's why it doesn't happen.

Similarly, they'd need a dedicated chip and some pretty elaborate hardware-software interplay in order to make 2 x4 slots stack together as a single x8 slot. The whole thing is so hard to achieve (especially since you know how Apple does not play nice with features they don't condone) that I might as well write "it doesn't work that way". Not that it's impossible, but that nobody would bother.

Is it clearer now?

Otherwise you'd think they'd have come out with a Thunderbolt to x8 adapter already:
http://www.sonnettech.com/product/xmacminiserver.html

Edit: or let me put it a different way: since the discussion involves an Apple device, of which the drivers are partially (or totally) controlled by Apple, it might well be "impossible" to achieve unless Apple themselves start supporting the idea.

Even then, it's a big mess to solve a "problem" (that should not exist) for a small niche.

How would they go to different stations when PCI-E is an interconnect.

 

[IceHot]

So if i hooked up a 680gtx to it, how much would be bandwidth limit the performance? 25%? ??

 

[bill-p]

How would they go to different stations when PCI-E is an interconnect.

"Interconnect" means the ability to form an end-to-end connection between ports.

So if I have two computers that I want to chain together to do parallel processing, I can connect their two PCIE ports together, and allow them to communicate back and forth simultaneously. That means communication goes both ways. The "interconnect" part is actually right in its name... and dates back all the way to PCI, which stands for Peripheral Component Interconnect.

The different stations is due to the fact that each PCIE port/slot (collection of lanes) runs on its own dedicated bus. As opposed to the old PCI standard, which shared a single bus with all slots.

You can connect different buses together, yes, but then you'll need a master controller to regulate data traffic in both buses and synchronize them so that data received and transmitted in both slots are ordered correctly... because different buses have different latencies. It's not impossible, but it's hard to do. That's why we have x16 slots rather than 2 x8 slots even though very few peripheral can actually max out an x16 slot.

On a motherboard with PCI Express slots, each PCI Express slot is connected to the motherboard chipset using a dedicated lane, not sharing this lane (data path) with other PCI Express slots.

Source: http://www.hardwaresecrets.com/article/190

So if i hooked up a 680gtx to it, how much would be bandwidth limit the performance? 25%? ??

Depends on the load. But you are looking at 4x bandwidth either way.

http://www.tomshardware.com/reviews/thunderbolt-performance-z77a-gd80,3205-4.html

I think I beat around the bush too much, but the short of it is: the chipset only supports a max of 4x link, and if you connect using 2 ports, then you get 2 separate 2x links, not 2 separate 4x links. Even if you could group the 2 ports together, you're still limited to 4x.

And while current graphics processors can't max out x8 regularly, they'd definitely max out x4. You're looking at almost 50% drop in some cases.