eGPU Performance Loss - PCI Express vs. Thunderbolt
 
Notifications
Clear all

eGPU Performance Loss - PCI Express vs. Thunderbolt  

 of  11
  RSS

ataylor
(@ataylor)
Eminent Member
Joined: 4 years ago
 

I as more wondering how NOT having the monitors connected to the egpu's but directly internal via tb3 adaptor will effect things. i.e will i not get full speed out of the gtx cards when no monitor is connected to them? I assume it won't make a difference for "compute" tasks.

This post was modified 3 years ago

2018 MacBookPro 15 in. /w Touchbar - i9 - 32gig ram - 1080 Aorus 1080 Gaming Box - Sonnet-Breakaway550 with GTX 1080 - Still need to get both of the GTX 980TI working for 4 GTX total cards


ReplyQuote
(@martin_poon)
New Member
Joined: 3 years ago
 

planned to get the Razer core x with RTX 2080ti, but hesitated after reading an article which said there is much 31% performance drop than those desktop setups

is it a hardware problem or software or alleviated any in recently?

IMG 8808

To do: Create my signature with system and expected eGPU configuration information to give context to my posts. I have no builds.

.

ReplyQuote
(@rydell_hall)
New Member
Joined: 3 years ago
 

Am I understanding this right, if you have a Thunderbolt 3 laptop (we'll say the latest Razer 15" so everything is up to spec) with an eGPU connected to an external 4k monitor, you will only see a ~5% reduction in performance vs a similarly spec'd desktop? I started considering getting a laptop+eGPU to replace my desktop but saw specs that you lose up to 30% of the GPUs performance.  So now I'm reading this saying that's only the case for resolutions lower than 4k. I have a 4k external monitor that I would be using at home for gaming, then when I'm gaming away from home I'd just use my laptop by itself. This seems like the perfect solution if that's the case. I don't care about a 5% performance loss, but a 30% is not worth it.

To do: Create my signature with system and expected eGPU configuration information to give context to my posts. I have no builds.

.

ReplyQuote
Mini i5
(@mini-i5)
Prominent Member
Joined: 3 years ago
 

@rydell_hall

Until you personally test a specific game on a specific laptop+eGPU - and are able to adjust the relevant power and performance settings - go ahead and assume a 30% performance loss. 👨‍💻

Sometimes the loss will be greater, depending on the game, and the laptop’s thermal management (aka thermal throttling).

This thread is a good resource. It explains the various factors involved. There are many.

2018 Mac Mini [8th,6C,B] + RX 6600 XT @ 32Gbps-TB3 (Sonnet Breakaway 650) + macOS 12.1 & Win10 21H1 [build link]  

itsage liked
ReplyQuote
(@rydell_hall)
New Member
Joined: 3 years ago
 
Posted by: @mini-i5

@rydell_hall

Until you personally test a specific game on a specific laptop+eGPU - and are able to adjust the relevant power and performance settings - go ahead and assume a 30% performance loss. 👨‍💻

Sometimes the loss will be greater, depending on the game, and the laptop’s thermal management (aka thermal throttling).

This thread is a good resource. It explains the various factors involved. There are many.

Appreciate the info. I read through the thread once again and it is seeming dicey. I'm about to pick up this laptop:

https://www.amazon.com/dp/B07HPQPNV1/ref=cm_sw_r_cp_api_i_pr4kEbYXJP4J6

With this card:

https://www.amazon.com/dp/B07HY6QWXN/ref=cm_sw_r_cp_api_i_It4kEb15JFN81

The goal is RDR2 on ultra (4k reso) with my existing 4k monitor w/ <10% performance loss

Seems what I'll have to do is start asking around if anyone has tried it with this specific setup. 

This post was modified 3 years ago

To do: Create my signature with system and expected eGPU configuration information to give context to my posts. I have no builds.

.

ReplyQuote
ron5553792
(@ron5553792)
Active Member
Joined: 2 years ago
 

@enjoy, these are great comparisons, however I am left wondering impact of the CPU on these setup.

The experiment I would have liked to run is different combinations of:

1) CPU (4-8, 6-6, 6-12, 8-8, 8-16 variants of cores-threads, probably ignore overclocking for now)

2) GPU (980ti, 1660s, 1070, 1080, 1080ti, 2080 - not sure I would bother going higher or lower)

3) connection type (PCIe, TB3, TB2 - not sure I would bother with TB1 anymore)

That would be 5 x 6 x 3 = 90 combinations, of which many are covered here but not all (missing the CPU variable). I am also ignoring the internal monitor case - I would only ever use an eGPU with an external monitor.

My hypothesis is that:

A) The bottleneck can be CPU, GPU, or connection type depending on the specific combo, but CPU likely bottlenecking more often than connection type or GPU. 

B) If you already have a laptop, then your CPU + connection will be fixed, but there will be a "maximum" GPU for you to buy, ie. going beyond that GPU will result in higher and higher degradation compared to using that same GPU in a desktop, so it's not really worth it unless you can buy it for less than this "maximum" GPU. For example, if you have a 2-lane 8th gen Intel chip, then maybe getting a 1660 Super is the best you can do, because getting a 3090 will result in only marginally better performance at much higher cost.

C) if you already have a GPU, then there is a "minimum" CPU-connection combo when buying a laptop to pair with it, ie. maybe TB2 with 6 cores + 6 threads, or TB3 with 8 cores and 16 threads. For example, if you already have a 1660 Super, maybe you should hunt for a 2-lane 8th gen Intel chip (or better), while for a 1080, you need 4-lane 10th gen (or better).

One proxy for this experiment (which would take a lot of parts and a lot of time) is to see how your GPU does in "https://gpu.userbenchmark.com/".  My Titan XP for example scores over 100% on my desktop and only 15% on my EGPU setup. What is the highest % anyone has ever gotten on their GPU when running an EGPU setup?  This could also answer the question.

 

 

HP Elitebook 830 G5 (i5-8250U) + 2x16GB 2400 CL16 DDR4 Crucial Ballistix
HP Elitebook 830 G6 (i7-8565U) + 2x16GB 2400 CL14 DDR4 HyperX Impact
Lenovo T480s (i7-8550U) + 2x8GB DDR4 Samsung (stock)
NVidia Titan Xp (Star Wars Collectors Edition)
Asus Rog Station 2


ReplyQuote
Manu89ft
(@manu89ft)
Eminent Member
Joined: 2 years ago
 

I'm trying with the signature configuration to play gears5, in split screen I don't go beyond 30fps. The video card is at 80% and the CPU at 40%, I play in 4k. Is it the bottleneck of the tb3? In my opinion with the latest generations of video cards you lose a lot with the egpu ....

2019 16" MacBook Pro (RP5500M) [9th,8C,H] + RX 6900 XT @ 32Gbps-TB3 (Razer Core X) + Win10 [build link]  

ReplyQuote
odin
 odin
(@odin)
Estimable Member
Joined: 4 years ago
 

In my experience so far the more GPU limited you are the better off you will fare vs desktop performance. This is much the reason 4K to an external monitor (and even internal panel in cases) will show less degradation percentage vs a desktop system than 1080p. Partly due to desktop CPUs being generally much faster. There is always going to be some difference. It also seems some specific games are much more susceptible to PCIe latency or bandwidth limitations. Probably based on how much they rely on expecting at least something like PCIe 3.0 x8 direct connection. eGPU PCI latencies are still an extreme edge case in the grand scope of systems developers expect their games to be running on. Desktops and real gaming laptops far outweigh eGPUs with extremely restrictive PCIe bandwidth/latency. Some dev houses might just be better at optimizing how their engines work and transfer data to the GPU than others as well.

Also, most of us are running 4c 15W ultrabooks or lightweight notebooks, since that's the main use case of an eGPU, having a portable laptop you can bring home and game with instead of having two systems. It comes with massive compromises and wer're CPU limited there. The 28W variant of those fares better but it's less common.

I wish more TB sporting notebooks had 6c/12t CPUs. It makes me sad that the i7-10710U wasn't used more often since it was 6c. I think it's probably because more companies thought it was better to add Xe/Iris graphics than 6c CPU.

2018 17" LG Gram 17 [8th,4C,U] + RTX 2070 @ 32Gbps-TB3 (Sonnet Breakaway 550) + Win10 [build link]  

ReplyQuote
Manu89ft
(@manu89ft)
Eminent Member
Joined: 2 years ago
 

@odin, in your experience, what affects this latency the most? Can it be improved in any way? Maybe removing the HDR, or doing some configuration ...

 

Certainly very demanding games, such as flight simulators, have similar performance to the desktop, so your reasoning is correct.

 

2019 16" MacBook Pro (RP5500M) [9th,8C,H] + RX 6900 XT @ 32Gbps-TB3 (Razer Core X) + Win10 [build link]  

odin liked
ReplyQuote
odin
 odin
(@odin)
Estimable Member
Joined: 4 years ago
 
Posted by: @manu89ft

@odin, in your experience, what affects this latency the most? Can it be improved in any way? Maybe removing the HDR, or doing some configuration ...

 

Certainly very demanding games, such as flight simulators, have similar performance to the desktop, so your reasoning is correct.

 

I think going on the baseline assumption that someone else already said that you should expect 20% performance loss on average is an acceptable assumption. It just varies so wildly and is susceptible to performance degradation, I think it might get worse as the PCI express bus gets faster and wider over time and TB doesn't seem to be keeping up. Some just bullet point thoughts I have (I might be wrong about any of them I'm no expert, still learning)

  • 4K will still be closer to true desktop/discrete GPU laptops due to GPU limited, especially on mid-range GPUs
  • Using internal panel is always going to impose a limitation due to sending back framebuffer data output, still seems GPU limited scenarios benefit internal panel as well though vs desktop
  • We often use 15W CPUs, desktop GPU reviews often use very high wattage CPUs.. we're usually CPU limited depending on the game
  • Seems some games just rely on PCIe bandwidth or latency much heavier than others, and that usually doesn't seem to affect desktops or laptop dGPUs as negatively. I think signs of this is significant performance deltas in that game between eGPU/desktop when CPU/GPU utilization is not near 100% on the eGPU scenario.
  • I fear that PCIe bus width keeps growing and thunderbolt is staying relatively static, although it still appears for now that 8x PCIe 3.0 might still be around the point where even high-end GPUs are not constrained by the PCI bus for gaming in most cases. This might start changing or might be dependent on GPU limiting again.
  • I fear that things like raytracing and resizable BAR placing higher demands on the PCI bus between CPU<->GPU might start affecting eGPU performance given TB's seemingly glacial pace of advancement
  • Things like what Asus is doing with their 8x PCIe 3.0 connector in the Flow X13 seem pretty cool. It would be neat if laptop manufacturers could rally around something standard for that but I don't see that happening at all. Probably more like TB5 or USB5. If Asus implements it in more laptops that would be nifty as just one laptop is going to limit its user base to people that can get by with a 13" panel. If they released a 16-17" one I would seriously consider that.
  • I wonder if 4x PCIe 4.0 m.2 eGPUs have the potential to perform almost as well as an 8x PCIe 3.0 GPU. Perhaps one already exists and I missed posts about it. But that still has all the baggage that goes along with using an m.2 eGPU in a laptop. Seems cool for edge case but a pain to seamlessly implement without frankenstien'ing your laptop and restricting it's storage potential. I would imagine unless you have two NVMe slots in the notebook it's almost a non-starter except for just kicks.

 

I'm seriously considering just ditching eGPU and going back to a discrete GPU laptop, but I waver back and forth on it. I might just sit tight until summer or so and analyze the changing landscape and my individual needs.

 

2018 17" LG Gram 17 [8th,4C,U] + RTX 2070 @ 32Gbps-TB3 (Sonnet Breakaway 550) + Win10 [build link]  

ReplyQuote
 of  11