Discussion To hell with nVidia's definition of SFF ready. What should be the actual SFF standard?

[ignsvn]

And just a PITA when fiddling around behind a PC case IMO

In my personal view, if you want to call something SFF-Ready AT ALL it should be based on the original PCI/PCIe spec

View attachment 3706

So I would suggest absolutely max of 111mm tall, 40mm thick (i.e. DUAL SLOT not 2.1, 2.2, etc) with considerations for power connectors and cable routing etc. for full sized cards.

Badge / Category	MAX Length (mm)	MAX Height (mm)	MAX Thickness (mm)	Example Card
SFF Long	312	111	40	IDK if there are any true dual-slot cards that could fit here???
SFF Standard / Reference	267	111	40	AMD Reference RX 6700 XT AMD Radeon RX 6700 XT Specs AMD Navi 22, 2581 MHz, 2560 Cores, 160 TMUs, 64 ROPs, 12288 MB GDDR6, 2000 MHz, 192 bit www.techpowerup.com nVidia RTX 3070 FE NVIDIA GeForce RTX 3070 Specs NVIDIA GA104, 1725 MHz, 5888 Cores, 184 TMUs, 96 ROPs, 8192 MB GDDR6, 1750 MHz, 256 bit www.techpowerup.com
SFF Short	225	111	40	PowerColor Fighter RX 6600XT PowerColor Fighter RX 6600 XT Specs AMD Navi 23, 2589 MHz, 2048 Cores, 128 TMUs, 64 ROPs, 8192 MB GDDR6, 2000 MHz, 128 bit www.techpowerup.com
SFF ITX	150-180	111	40	HP RTX 3060 Single Fan HP RTX 3060 OEM Specs NVIDIA GA106, 1777 MHz, 3584 Cores, 112 TMUs, 48 ROPs, 12288 MB GDDR6, 1875 MHz, 192 bit www.techpowerup.com PowerColor Radeon RX 5600 XT ITX PowerColor RX 5600 XT ITX Specs AMD Navi 10, 1560 MHz, 2304 Cores, 144 TMUs, 64 ROPs, 6144 MB GDDR6, 1750 MHz, 192 bit www.techpowerup.com
SFF LP	190	69	40	Gigabyte RTX 4060 LP GIGABYTE RTX 4060 Low Profile OC Specs NVIDIA AD107, 2475 MHz, 3072 Cores, 96 TMUs, 48 ROPs, 8192 MB GDDR6, 2125 MHz, 128 bit www.techpowerup.com
SFF LP ITX	170	69	40	nVidia RTX 4000 SFF Ada NVIDIA RTX 4000 SFF Ada Generation Specs NVIDIA AD104, 1560 MHz, 6144 Cores, 192 TMUs, 80 ROPs, 20480 MB GDDR6, 1750 MHz, 160 bit www.techpowerup.com

Anyways, just my 2c

Actually this makes sense, yeah.

 

[ignsvn]

Just so we are on the same page - am I thinking about the same thing as rest of you guys? Keeping the possible fitting from pci reference design like this?

This is corresponding with how we assumed 112 mm reference size + 18 mm of room required for PEG cable bend in Sentry.

Note how stupid the recess for 12VHPWR is because of that requirement of 35 mm to the bend - it might make more sense to make it parallel on the top of the card instead of cutting into the board... Maybe the standard should show how stupid this is and require the 12VHWPR to be parallel like in the bottom drawing?

I also thought about potentially drawing the PEG 8-pin as flushed in within the 111,2 mm line, but it could be too greedy to ask for more than what we have had with reference cards. Still the question stands if perhaps when drawing up the standard, maybe we should try to make a step forward?

This is so nostalgic. I often see these kinds of diagrams early on when people build new cases.

 

[yeshyyyk]

My opinion; I think sizes should kinda "correlate" to TDP (not very SFF to put a 100W card in the largest options). But hard to say without some kind of noise-temperature-size-TDP normalized testing...

Badge / Category	MAX Length (mm)	MAX Height (mm)	MAX Thickness (mm)	Target TDP (W), example probably should increase by 15-20% imo (or this can be minimum?)
SFF Long	312	111	40	270
SFF Standard / Reference	267	111	40	230, 285W+ done by Gigabyte 4070Tis AI TOP, 3080 TURBO
SFF Short	225	111	40	200, EVGA example, 3060 Ti XC BLACK GAMING, otherwise easily exceeded (since smaller) by Powercolor Vega 56 Nano
SFF ITX	180	111	40	160, Powercolor 5700 ITX, otherwise easily exceeded (since smaller, much higher TDP) by Powercolor Vega 56 Nano
SFF LP	190	69	40	100
SFF LP ITX	170	69	40	100

 

[Necere]

I'll throw in my two cents.

• I'd be happy with the old reference dimensions (267x111x40mm) as a new SFF standard. I'd be cautious of breaking it up into multiple standards though so as to not split an already niche market and cause consumer confusion.
• I'd like to see recessed PCIe power connectors as part of the standard, ideally forward-facing at the top edge of the card. This would maximize compatibility with different styles of SFF case while still being in an easily accessible location (something which IMO would be an issue with @SaperPL's idea of putting it on the lower edge of the card).

Quick mock-up:

I will add, taking into account cooling needs I could see an argument for extending the maximum length up to 290mm or so to allow for triple ~92mm fans. This also maximizes the use of space in M1, sandwich, and console layout cases where you'll need a minimum of 300mm or so for a mini-ITX board and SFX PSU anyway.

 

[Arboreal]

The guru has spoken...This is the way!

 

[SaperPL]

I'll throw in my two cents.

• I'd be happy with the old reference dimensions (267x111x40mm) as a new SFF standard. I'd be cautious of breaking it up into multiple standards though so as to not split an already niche market and cause consumer confusion.
• I'd like to see recessed PCIe power connectors as part of the standard, ideally forward-facing at the top edge of the card. This would maximize compatibility with different styles of SFF case while still being in an easily accessible location (something which IMO would be an issue with @SaperPL's idea of putting it on the lower edge of the card).

Quick mock-up:

I will add, taking into account cooling needs I could see an argument for extending the maximum length up to 290mm or so to allow for triple ~92mm fans. This also maximizes the use of space in M1, sandwich, and console layout cases where you'll need a minimum of 300mm or so for a mini-ITX board and SFX PSU anyway.

I agree with this approach, but there's one piece of the puzzle that you're missing here - I'd even prefer the connectors on the top edge like you've shown here, except maybe at 267mm in cases where you want to go around the card and need to do it in the same plane, you'd need the card to be shorter to squeeze the cables or the case to be longer. But that's not a problem if it's a standard. I'd just want the 12VHPWR to be recessed enough for that offset required for bending to fit withing the outline of the card.

The problem is that nvidia will not allow us to have cards like that and most of the cards will not be able to come up with connectors placed like in quadro and tesla cards because then they would fit in server racks that are dedicated to compute and maybe also in slim towers from HP/Dell etc. That is why my approach is to place the cable in a way it will be inconvenient for E-ATX server boards, but still viable for up to ATX boards size.

Also I'm trying to ride the tide of stealth cable design as this was the only trend in changing the cable placement that got card vendors interested enough to get to a point of manufacturing and selling the cards - I got 4070 from Inno3D and now Sapphire has a 9070 like this. It makes much more sense to propose something that will not only be a selling point for us in SFF niche, but also for general audience with full tower cases.

As for going over 267mm - We can have easily 3 x 85mm fans in it and there are cards like that already in 260mm length, so I'm not sure if it's worth assuming we need 92mm fans. I also agree splitting between multiple tiers doesn't make sense and thus I'd stick to this size of 267mm / 10.5"

 

[robbee]

The problem is that nvidia will not allow us to have cards like that and most of the cards will not be able to come up with connectors placed like in quadro and tesla cards because then they would fit in server racks that are dedicated to compute and maybe also in slim towers from HP/Dell etc. That is why my approach is to place the cable in a way it will be inconvenient for E-ATX server boards, but still viable for up to ATX boards size.

Is that an official thing? I know of a recent example where Lenovo place the connector facing the rear on the RTX 4060 used in their prebuilts. Although these cards don't retail outside of prebuilts, which may explain the different rules set by Nvidia.

 

[Necere]

I agree with this approach, but there's one piece of the puzzle that you're missing here - I'd even prefer the connectors on the top edge like you've shown here, except maybe at 267mm in cases where you want to go around the card and need to do it in the same plane, you'd need the card to be shorter to squeeze the cables or the case to be longer. But that's not a problem if it's a standard. I'd just want the 12VHPWR to be recessed enough for that offset required for bending to fit withing the outline of the card.

I did think of this, hence 290mm and not 300mm. That way you'd have at least 10mm off the end of the card to run cables in a typical ITX+front SFX case. IMO running the cable at the front makes the most sense for the common layouts.

The problem is that nvidia will not allow us to have cards like that and most of the cards will not be able to come up with connectors placed like in quadro and tesla cards because then they would fit in server racks that are dedicated to compute and maybe also in slim towers from HP/Dell etc. That is why my approach is to place the cable in a way it will be inconvenient for E-ATX server boards, but still viable for up to ATX boards size.

This I had not thought of, and rather unfortunate if true.

I wonder if this is really an issue though. GPUs intended for server racks are either passive or blower-cooled with longitudinal fins, which aren't something you find very often in the AIB space, and axial fans on lateral fin coolers seem like they would perform poorly in a space-constrained server racks designed around front-to-back flow-through airflow.

Also I'm trying to ride the tide of stealth cable design as this was the only trend in changing the cable placement that got card vendors interested enough to get to a point of manufacturing and selling the cards - I got 4070 from Inno3D and now Sapphire has a 9070 like this. It makes much more sense to propose something that will not only be a selling point for us in SFF niche, but also for general audience with full tower cases.

I suppose if the standard just calls for the connector+cable bend to be fully contained within the card dimensions then it's less important where exactly the connector is located. I do still think specifying the location would allow you to more fully optimize the case design, though.

As for going over 267mm - We can have easily 3 x 85mm fans in it and there are cards like that already in 260mm length, so I'm not sure if it's worth assuming we need 92mm fans. I also agree splitting between multiple tiers doesn't make sense and thus I'd stick to this size of 267mm / 10.5"

The thing about 267mm is that it's an akward size when considered relative to the common case layouts: ITX+front SFX has at least 20-30mm of available space that will be there anyway, while shallower layouts would be hindered by the length. IMO the next step down from 290mm that makes sense is 240-250mm for ITX+front flex PSU (though this may be too niche), and 170mm below that to match ITX board depth.

I also like the idea of ~290mm because it leaves open the door to using standard 92mm fans, either as a mod or as an OEM model.

 

[SaperPL]

This I had not thought of, and rather unfortunate if true.

I wonder if this is really an issue though. GPUs intended for server racks are either passive or blower-cooled with longitudinal fins, which aren't something you find very often in the AIB space, and axial fans on lateral fin coolers seem like they would perform poorly in a space-constrained server racks designed around front-to-back flow-through airflow.

I think I did pester Linus on his forum about bringing up this topic in video and based on his response I believe he gave up on talking to nvidia about this because it's how they want it to be, or he's not going to waste time on things that he sees will not have any effect.

About the fins being oriented for racks - while 3-fan solutions often have fins vertically, for smaller cards some vendors still do make horizontal fins like zotac did here for 3060 TI:

But you're right - I think both of these features matter, the fin orientation and the power connector placement.

I suppose if the standard just calls for the connector+cable bend to be fully contained within the card dimensions then it's less important where exactly the connector is located. I do still think specifying the location would allow you to more fully optimize the case design, though.

Getting back to the issue of nvidia not wanting this - this small cutout in my concept to pass the cable out here would probably invalidate the design for them:

Because this essentially works as if the card was a blower with connector placed at the end. So allowing this to exit anywhere may not be a good strategy beside optimising this for us, case designers.

The thing about 267mm is that it's an akward size when considered relative to the common case layouts: ITX+front SFX has at least 20-30mm of available space that will be there anyway, while shallower layouts would be hindered by the length. IMO the next step down from 290mm that makes sense is 240-250mm for ITX+front flex PSU (though this may be too niche), and 170mm below that to match ITX board depth.

It's not an awkward size when talking about ITX + TFX in front or ITX + FlexATX in front. 240-250 I think may be too short unless you're considering that additional space at the end required for the cable and its bend. I would stick to the 267mm requirement here because it shows it can fit 3-fan solutions neatly.

I agree that for ITX + SFX it's great to just have it as 12"/305mm limit. It's just that triple fan cards with 200-250W TDP can fit in this size. RTX 5070 FE is within this size and Arc B580 is almost within this size.

I'm for picking one size because splitting hairs into multiple brackets makes things complicated less likely for anyone to follow. Also it may not make sense for more power hungry cards to try and fit in such reference-sized SFF requirements - we may wish for that, but for anything above 300W even 12"/305mm at reference height of the card may not be reasonable.

Overall I don't know whether we should focus on 305mm or 267mm, I have a preference for smaller one, but I just believe we should pick one.

Is that an official thing? I know of a recent example where Lenovo place the connector facing the rear on the RTX 4060 used in their prebuilts. Although these cards don't retail outside of prebuilts, which may explain the different rules set by Nvidia.

I think it's not publicly official, but I think its exactly so if you buy pre-builts from Lenovo/HP/Dell, replacing the gpu with something retail gets tricky and you need to go for upgrade/replacement from them, or go for a quadro card. This would make sense to fight those recycling scenarios of buying used 1 or 2 generations old office PCs and putting in an retail GPU easily. That's why they have stupidly placed x16 slots and custom cables on their power supplies and motherboards etc.

I believe in previous generation there was this situation that few partners said early on they would release blower cards with RTX 4000 series early on, probably because nvidia missed this requirement somewhere assuming it's clear how they roll with this topic, and nvidia was like "oh no you won't" and for some of the partners we got these really late or scarce quantities, and some did something like gigabyte and remade them into things like this card:

https://www.gigabyte.com/Graphics-Card/GV-N4090WF3V2-24GD-rev-10-11#kf

In all this you need to consider the big crypto miners situation as well - if they can get hands on some standardised racks/used servers made for quadros and teslas or whatever its called right now that don't have room for that connector on top, they're forced to either use professional cards OR mod the retail consumer cards to fit there. If nvidia allowed that to happen, we could have floods of used cards from last generation more often than currently when new generation launches and crypto miners switch onto next generation. Yes - there are janky open-bench solutions, but when it comes to optimising space, rack gives you best optimisation, and if someone goes for this, just having taller/thicker rack case just because you need to fit in connectors doesn't make sense. So now you have to solder cables to each card to use it, and when selling those cards back, you have to resolder the connector back so you can lie that it wasn't running 24/7 for it's whole lifetime so far.

It's all about making multiple attempts at preventing using consumer cards where nvidia would want to make more money on the chips with professional cards, so while they cannot prevent you from hacking together a custom solution for yourself, the clients that buy things in volume and want warranty will be forced to use what they are allowing to use for each segment.

 

[SFFMunkee]

I think I did pester Linus on his forum about bringing up this topic in video and based on his response I believe he gave up on talking to nvidia about this because it's how they want it to be, or he's not going to waste time on things that he sees will not have any effect.

About the fins being oriented for racks - while 3-fan solutions often have fins vertically, for smaller cards some vendors still do make horizontal fins like zotac did here for 3060 TI:

But you're right - I think both of these features matter, the fin orientation and the power connector placement.


Getting back to the issue of nvidia not wanting this - this small cutout in my concept to pass the cable out here would probably invalidate the design for them:

Because this essentially works as if the card was a blower with connector placed at the end. So allowing this to exit anywhere may not be a good strategy beside optimising this for us, case designers.


It's not an awkward size when talking about ITX + TFX in front or ITX + FlexATX in front. 240-250 I think may be too short unless you're considering that additional space at the end required for the cable and its bend. I would stick to the 267mm requirement here because it shows it can fit 3-fan solutions neatly.

I agree that for ITX + SFX it's great to just have it as 12"/305mm limit. It's just that triple fan cards with 200-250W TDP can fit in this size. RTX 5070 FE is within this size and Arc B580 is almost within this size.

I'm for picking one size because splitting hairs into multiple brackets makes things complicated less likely for anyone to follow. Also it may not make sense for more power hungry cards to try and fit in such reference-sized SFF requirements - we may wish for that, but for anything above 300W even 12"/305mm at reference height of the card may not be reasonable.

Overall I don't know whether we should focus on 305mm or 267mm, I have a preference for smaller one, but I just believe we should pick one.



I think it's not publicly official, but I think its exactly so if you buy pre-builts from Lenovo/HP/Dell, replacing the gpu with something retail gets tricky and you need to go for upgrade/replacement from them, or go for a quadro card. This would make sense to fight those recycling scenarios of buying used 1 or 2 generations old office PCs and putting in an retail GPU easily. That's why they have stupidly placed x16 slots and custom cables on their power supplies and motherboards etc.

I believe in previous generation there was this situation that few partners said early on they would release blower cards with RTX 4000 series early on, probably because nvidia missed this requirement somewhere assuming it's clear how they roll with this topic, and nvidia was like "oh no you won't" and for some of the partners we got these really late or scarce quantities, and some did something like gigabyte and remade them into things like this card:

https://www.gigabyte.com/Graphics-Card/GV-N4090WF3V2-24GD-rev-10-11#kf

In all this you need to consider the big crypto miners situation as well - if they can get hands on some standardised racks/used servers made for quadros and teslas or whatever its called right now that don't have room for that connector on top, they're forced to either use professional cards OR mod the retail consumer cards to fit there. If nvidia allowed that to happen, we could have floods of used cards from last generation more often than currently when new generation launches and crypto miners switch onto next generation. Yes - there are janky open-bench solutions, but when it comes to optimising space, rack gives you best optimisation, and if someone goes for this, just having taller/thicker rack case just because you need to fit in connectors doesn't make sense. So now you have to solder cables to each card to use it, and when selling those cards back, you have to resolder the connector back so you can lie that it wasn't running 24/7 for it's whole lifetime so far.

It's all about making multiple attempts at preventing using consumer cards where nvidia would want to make more money on the chips with professional cards, so while they cannot prevent you from hacking together a custom solution for yourself, the clients that buy things in volume and want warranty will be forced to use what they are allowing to use for each segment.

I also strongly lean towards 267mm and push the reference dimensions, but with considerations for cable routing as to why it's SFF friendly. Those two design elements are what makes it more friendly for case designers and builders alike!

You raise the right question though, how do we make it SFF designer/builder friendly AND not suitable for professional rack/high-density compute/etc use cases so that nVidia will actually support it...

LOL at that Windforce card that was so clearly an afterthought retrofit!

 

[yeshyyyk]

Overall I don't know whether we should focus on 305mm or 267mm, I have a preference for smaller one, but I just believe we should pick one.

As I mentioned in my post

Badge / Category	MAX Length (mm)	MAX Height (mm)	MAX Thickness (mm)	Target TDP (W), example probably should increase by 15-20% imo (or this can be minimum?)
SFF Long	312	111	40	270
SFF Standard / Reference	267	111	40	230, 285W+ done by Gigabyte 4070Tis AI TOP, 3080 TURBO

267mm "can" do ~300W, so only if you wanted to cool well beyond (5080 comfortably, or 5090 with really well designed cooler) that would it be necessary to go to 305 or even 290mm

Non-blower example would be Inno3D X2 / X3 and mayybe Nvidia FE, but all have to be 111mm width

I wouldn't mind having a 5080/5090 in shorter/longer dimension that can only get quiet+ <79C if undervolted+PL, since that'll drop 75W/140W or so

 

[Medeyer]

I'm seeing a lot of support for 267mm as the maximum length, though I suspect that's largely in part due to nostalgia for the 780/1080 reference models. However, I think we should take into consideration to the size of motherboards themselves. The width of ATX and mATX boards is 244mm (9.6") and would be an easy metric to set as a potential standard.

 

[SFFMunkee]

I'm seeing a lot of support for 267mm as the maximum length, though I suspect that's largely in part due to nostalgia for the 780/1080 reference models. However, I think we should take into consideration to the size of motherboards themselves. The width of ATX and mATX boards is 244mm (9.6") and would be an easy metric to set as a potential standard.

The reason for using this number, is that it's based on the existing PCI and thus PCIe standard itself.

 

[SaperPL]

I'm seeing a lot of support for 267mm as the maximum length, though I suspect that's largely in part due to nostalgia for the 780/1080 reference models. However, I think we should take into consideration to the size of motherboards themselves. The width of ATX and mATX boards is 244mm (9.6") and would be an easy metric to set as a potential standard.

Take a look a the stealth/hidden cable problem as well. Card needs to be longer than the motherboard to be able to route the cable to the backplane there, otherwise you would have to sqash cables between the motherboard and card to go around the board.

Supporting or even expecting the stealth cable design works in favor of not needing to figure out how much space on top of the card is needed for the cable, as well as this kind of form factor will be interesting for tower pcs with tempered glass as well so these cards will have a higher chance of gaining traction and surviving on the market if there's more usecases than just sff.

Also if we do actually have problem with nvidia not wanting consumer cards with cable oriented horizontally, then this design may make it hard to use with eatx boards, potentially guving us chance for it to be acceptable for nvidia.