Maple ridge TB4 host controller has 32Gbps bandwidth?
 
Notifications
Clear all

Maple ridge TB4 host controller has 32Gbps bandwidth?  

  RSS

odin
 odin
(@odin)
Estimable Member
Joined: 3 years ago
 

Hidden in this AnandTech article: https://www.anandtech.com/show/16333/intel-maple-ridge-thunderbolt-4-controller-now-shipping

Is the interesting tidbit related to TB4 bandwidth:

Thunderbolt 3's bandwidth sharing mechanism between video and data also put in some dampeners – even in the absence of tunneling DisplayPort streams, 18 Gbps of bandwidth was always reserved for video traffic, and only 22 Gbps available for actual data transfer. Thunderbolt 4 apparently fixes that with up to 32 Gbps of data traffic (full PCIe 3.0 x4 bandwidth) available, allowing devices such as Thunderbolt 4 SSDs to provide 3GBps+ speeds.

I didn't know this aspect, sounds interesting for egpu applications for TB4 enclosures that may launch soon. Especially gaming performance.

LG Gram 17 | Sonnet Breakaway Box 550 | Asus Strix RTX 2070 OC Edition | Win 10 Pro 20H2 + Fedora 32 Dual Boot
Build Link

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


ReplyQuote
NessLookAlike
(@nesslookalike)
Active Member
Joined: 2 months ago
 

I've read a handful of threads around the eGPU.io forums about Thunderbolt 4 vs. 3, read a handful of tech websites and reddit posts about it, and the popular conclusion I'm seeing are 1) that there's no bandwidth change between the two, so there's no performance difference, and 2) TB4 is more about guaranteeing minimum requirements than about performance improvement. But I've also seen a handful of comments (on the eGPU.io forums) that while TB3 limits data throughput to ~22Gb/s, TB4 removes that limit and increases data rate throughput up to ~32Gb/s (?!).

Looking for a sanity check here -- Isn't that 10Gb/s increase a substantial performance improvement? Especially for eGPU applications where effective throughput has such a significant impact on gaming performance (especially with high framerate gaming)? Or am I misunderstanding something here, and TB4 isn't a big deal after all, and I shouldn't grab a TB4 enclosure as soon as a decent solution becomes available, and I should hang on to my Core X Chroma until TB5 etc.? Thanks a bunch. 

2020 15" HP Spectre X360 [11th,4C,G] + RTX 3090 @ 32Gbps-TB4 (Razer Core X Chroma) + Win10 2004 [build link TBD]


ReplyQuote
nando4
(@nando4)
Noble Member Admin
Joined: 4 years ago
 

@nesslookalike sanity heck will be confirmed when somebody pairs a TB4 enclosure with a TB4 system and measures bandwidth using CUDA-Z or AIDA64. Maybe you'd like to do it and report your findings?

eGPU Setup 1.35    •    eGPU Port Bandwidth Reference Table

 
2015 15" Dell Precision 7510 (Q M1000M) [6th,4C,H] + GTX 1080 Ti @32Gbps-M2 (ADT-Link R43SG) + Win10 1803 // compares M.2 vs TB3 performance inc unoptimized H-CPU BIOS [build link]  


ReplyQuote
joevt
(@joevt)
Noble Member
Joined: 3 years ago
 

Titan Ridge host (maybe even Alpine Ridge?) can already do read/write 3000 MB/s using AmorphousDiskMark to a Sabrent Rocket 2TB 4.0 NVMe. That's 24 Gbps. You need another 4 Gbps to reach 3500 MB/s (28 Gbps) which is typical for an NVMe in a PCIe 3.0 x4 slot (= 31.5 Gbps which is impossible because you need to factor PCIe overhead). Tiger Lake can't do 3500 MB/s over Thunderbolt. I don't think Maple Ridge is going to do it.

Mac mini (2018), Mac Pro (Early 2008), MacBook Pro (Retina, 15-inch, Mid 2015), GA-Z170X-Gaming 7, Sapphire Pulse Radeon RX 580 8GB GDDR5, Sonnet Echo Express III-D, Trebleet Thunderbolt 3 to NVMe M.2 case


ReplyQuote
odin
 odin
(@odin)
Estimable Member
Joined: 3 years ago
 
Posted by: @joevt

Titan Ridge host (maybe even Alpine Ridge?) can already do read/write 3000 MB/s using AmorphousDiskMark to a Sabrent Rocket 2TB 4.0 NVMe. That's 24 Gbps. You need another 4 Gbps to reach 3500 MB/s (28 Gbps) which is typical for an NVMe in a PCIe 3.0 x4 slot (= 31.5 Gbps which is impossible because you need to factor PCIe overhead). Tiger Lake can't do 3500 MB/s over Thunderbolt. I don't think Maple Ridge is going to do it.

So you're not expecting eGPU enclosures that use the Goshen Ridge client controller connected to a TB4 host such as one that has embedded direct to CPU support like the 1165G7 to see any benefit over TB3 eGPU connected to a TB3/4 controller?

 

EDIT: I do expect that the chance of an eGPU box using one soon to be low unless it's backwards compatible with TB3 hosts.

LG Gram 17 | Sonnet Breakaway Box 550 | Asus Strix RTX 2070 OC Edition | Win 10 Pro 20H2 + Fedora 32 Dual Boot
Build Link

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


ReplyQuote
joevt
(@joevt)
Noble Member
Joined: 3 years ago
 

@odin, Correct, I don't expect a speed up from Goshen Ridge to Maple Ridge or integrated Thunderbolt.

I don't see any reason for an eGPU to use Goshen Ridge while Alpine Ridge and Titan Ridge still exist. I think Intel still doesn't support downstream Thunderbolt ports, so Goshen Ridge would be a waste in that case.

Maybe the extra USB ports of Goshen Ridge would be useful? I'm not sure how Goshen Ridge does USB - The OWC Thunderbolt Hub uses a USB hub to do USB for it's four ports - the USB 2.0 part of the hub is Fresco Logic and the USB 3.x part is Intel but I don't know if these are separate chips are built into the Goshen Ridge. If they are separate chips then you don't gain anything in this respect.

Maybe the USB4 functionality of Goshen Ridge might be useful for connecting to USB4 hosts that don't support Thunderbolt (I don't know of any USB4-only hosts that exist at the moment).

 

 

Mac mini (2018), Mac Pro (Early 2008), MacBook Pro (Retina, 15-inch, Mid 2015), GA-Z170X-Gaming 7, Sapphire Pulse Radeon RX 580 8GB GDDR5, Sonnet Echo Express III-D, Trebleet Thunderbolt 3 to NVMe M.2 case


ReplyQuote
odin
 odin
(@odin)
Estimable Member
Joined: 3 years ago
 
Posted by: @joevt

@odin, Correct, I don't expect a speed up from Goshen Ridge to Maple Ridge or integrated Thunderbolt.

I don't see any reason for an eGPU to use Goshen Ridge while Alpine Ridge and Titan Ridge still exist. I think Intel still doesn't support downstream Thunderbolt ports, so Goshen Ridge would be a waste in that case.

Maybe the extra USB ports of Goshen Ridge would be useful? I'm not sure how Goshen Ridge does USB - The OWC Thunderbolt Hub uses a USB hub to do USB for it's four ports - the USB 2.0 part of the hub is Fresco Logic and the USB 3.x part is Intel but I don't know if these are separate chips are built into the Goshen Ridge. If they are separate chips then you don't gain anything in this respect.

Maybe the USB4 functionality of Goshen Ridge might be useful for connecting to USB4 hosts that don't support Thunderbolt (I don't know of any USB4-only hosts that exist at the moment).

 

 

 

I was thinking that eGPUs being able to utilize true 32GB/s instead of always having 10GB/s being allocated to video would be beneficial, as well as hopefully most laptops using controllers embedded in the CPU package instead of a separate controller. Maybe I'm not quite understanding how this all works currently. I assume my current TB3 setup whether I am using my current laptop with an Alpine Ridge controller on the mainboard or a newer one like the 1165G7 with an embedded controller I'd still be under the 22Gb/s maximum throughput. I'd expect pairing TB4 hosts and devices to be able to utilize full 32Gb/s much like m.2 eGPUs currently do, but with the added ability to use the internal panel if desired, which I think doesn't work with m.2 eGPUs. Not to mention that modifying a laptop to actually use an m.2 eGPU without looking bad or having pain in the rear cables restricting portability use of the laptop when not connected as well as needing a laptop with two m.2 slots. I don;'t think many people want to use external drives just to get better GPU bandwidth. That's even more of an edge case than eGPU already is in the space. It seems like TB4 if implemented properly could render the reason to want an m.2 eGPU obsolete once it's deployed at scale in devices and hosts.

Again, maybe I misunderstand what I'm reading in all the papers about it.

LG Gram 17 | Sonnet Breakaway Box 550 | Asus Strix RTX 2070 OC Edition | Win 10 Pro 20H2 + Fedora 32 Dual Boot
Build Link

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


ReplyQuote