Why don’t they make a low profile gpu with a powerful mobile GPU?

[thepoorgamer]

I wish someone would make a low power gpu based on the rtx 3070 that would draw all of it’s power from the pcie slot. I know there probably is not much of a market for low power 3070s, but it would be really cool (and let you make really powerful tiny pc’s). It would allow you to upgrade any system and turn it into a really fast pc. I have a dell inspiron desktop and it would be really cool (though completely unbalanced) if I could put a 3070 in it! You could also use it to upgrade systems with weak psu’s. After all, the GPU is almost always the bottleneck in a gaming system, even if it is an office pc with an older processor. It would smash a low power 1650 and preform similarly to a desktop 3060. If anyone makes one, I will buy it. With that being said, it would make no sense to design a card for the 5 people that need that powerful of a low power GPU.

[Middcore]

You know there's a limit to how much power can be supplied by just the PCI-e slot, right? 

[thepoorgamer]

Just now, Middcore said:

You know there's a limit to how much power can be supplied by just the PCI-e slot, right? 

Yhea. I was thinking they could set the power limit to 75 watts.

[Middcore]

Just now, thepoorgamer said:

Yhea. I was thinking they could set the power limit to 75 watts.

 

What kind of performance do you think you would get from a 3070 at a power limit of roughly 1/3rd it's normal draw? 

[bellabichon]

Just now, thepoorgamer said:

Yhea. I was thinking they could set the power limit to 75 watts.

A desktop 3070 uses 200 watts. Getting it to run on 75 would be... challenging. 

[Radium_Angel]

2 minutes ago, Middcore said:

 

What kind of performance do you think you would get from a 3070 at a power limit of roughly 1/3rd it's normal draw? 

Really crap performance!(TM)

[lafrente]

That would be like creating a hypercar and limiting it at 100 HP. Waste of resources.

[-rascal-]

Using the GTX 3070 as the example...

• Making the GPU run with 75W means downclocking the GPU and Memory frequency SIGNFICANTLY
• Probably chunks of CUDA / RT cores needs to be disabled
• The full-size RTX 3070 GPU runs hot, so a tiny cooler won't be enough to keep a full RTX 3070 cool

Apply those changes....and you now have a cut-down GA104 Core....which is like a "RTX 3040"...or something.

So much for a RTX 3070.

 

Similarly:

So...instead of a GTX 1070....you now have a Low-Profile GTX 1050...which exists.

[SupaKomputa]

7 minutes ago, -rascal- said:

• Making the GPU run with 75W means downclocking the GPU and Memory frequency SIGNFICANTLY
• Probably chunks of CUDA / RT cores needs to be disabled
• The full-size RTX 3070 GPU runs hot, so a tiny cooler won't be enough to keep it cool

OP asking about using a laptop chip for a low powered gpu.

All your point is what 3080 max-q are made of, lower clock and lower cuda count.

I think it's doable, but noone will do it. 

[silk186]

The issue isn't mobile or desktop chip, it's power draw, thermals and noise. There are many complaints about the 3xxx series GPUs being limited buy insufficient power in laptops. Manufacturers are trying to address this with dynamic power sharing with the CPU.

[tkitch]

20 minutes ago, -rascal- said:

Using the GTX 3070 as the example...

• Making the GPU run with 75W means downclocking the GPU and Memory frequency SIGNFICANTLY
• Probably chunks of CUDA / RT cores needs to be disabled
• The full-size RTX 3070 GPU runs hot, so a tiny cooler won't be enough to keep a full RTX 3070 cool

Apply those changes....and you now have a cut-down GA104 Core....which is like a "RTX 3040"...or something.

So much for a RTX 3070.

 

Similarly:

So...instead of a GTX 1070....you now have a Low-Profile GTX 1050...which exists.

The 1650 also had low profile, non-PCIE powered options, so they could have stepped up a bit. 

[-rascal-]

35 minutes ago, SupaKomputa said:

OP asking about using a laptop chip for a low powered gpu.

All your point is what 3080 max-q are made of, lower clock and lower cuda count.

I think it's doable, but noone will do it. 

 

Given the starting post...he mentions he's got a DELL desktop, and then mentioning about using lower power output PSUs, office / compact PCs.

I got desktop GA104 from that 

 

27 minutes ago, tkitch said:

The 1650 also had low profile, non-PCIE powered options, so they could have stepped up a bit. 

 

Hmm... that is true.

I used the GTX 1080 / GTX 1050, and RTX 3070 / "RTX 3040" as examples.

We didn't get a whole lot of GTX 1600-series out of the TU117 / TU116 / TU106 chips.

 

The GTX 1660 is on TU116 with a 120W TDP..."reference" had a PCI-E 8-pin connector -- while GTX 1650 is on TU117 / TU106.

GTX 1660 vs GTX 1650 has quite a bit of differences, though.

I *could* see a 75W version of the GTX 1650 Super...not sure bout a GTX 1660 / 1660 Super.

[porina]

Seeing some misunderstanding of hardware scaling in this thread.

 

Generally speaking, both CPUs and GPUs run much more power efficiently at lower clocks than higher clocks. A given % reduction in power will result in a smaller % reduction in performance. For best efficiency, you actually want the biggest core you can get, and run it as slow as possible. Due to the workloads, GPUs are a case where going wider still generally scales well, unlike with CPUs.

 

Looking it up a 3070 mobile is apparently: (EDIT: these mobile numbers are different than the nvidia page shot in post below, and may be because they are one actual implementation but doesn't represent the whole range. It is within the range and good enough for a hypothetical exercise like this)

5120 cores, 1110 base, 1560 boost, 1750 mem, 115W TBP.

Desktop version:

5888 cores, 1500 base, 1725 boost, 1750 mem, 220W TBP.

 

Overall, the mobile version is about 79% the peak performance of the desktop version, at just over half the board power. Generally speaking, higher performance desktop products run well into the less efficient zone. Backing off a bit improves efficiency a lot. However if we start at the lower power and try to reduce it further to say 75W, you'll probably see closer to linear scaling there since it will be on a much flatter part of the curve. Not everything will scale, especially more so if we do more extreme power shifts, but I'd estimate a "75W 3070" would be about half the performance of the full power one.

 

The mobile 3070 is pretty close to the desktop 3060, the latter having 170W TBP. That's in part due to tradeoff between cores and clocks mentioned already, although binning may also be a contributing factor. Anyway, for a typical desktop component buyer, would you want to pay (desktop) 3070 price to get (desktop) 3060 performance? Or just buy the cheaper 3060 at higher power? For those who really do want lower power cards, setting a lower power limit on a more expensive model is about the most realistic way of going about it. This in itself wont be enough to eliminate the requirement of the PCIe power connector.

 

Cooling of a low power "75W" 3070 would be very easy. That amount of power is not great, and the relatively big die give a nice area for contact. It would probably be easier to cool than an AMD 75W GPU for example.

[tkitch]

3 minutes ago, -rascal- said:

 

Given the starting post...he mentions he's got a DELL desktop, and then mentioning about using lower power output PSUs, office / compact PCs.

I got desktop GA104 from that 

 

 

Hmm... that is true.

I used the GTX 1080 / GTX 1050, and RTX 3070 / "RTX 3040" as examples.

We didn't get a whole lot of GTX 1600-series out of the TU117 / TU116 / TU106 chips.

 

The GTX 1660 is on TU116 with a 120W TDP..."reference" had a PCI-E 8-pin connector -- while GTX 1650 is on TU117 / TU106.

GTX 1660 vs GTX 1650 has quite a bit of differences, though.

I *could* see a 75W version of the GTX 1650 Super...not sure bout a GTX 1660 / 1660 Super.

Fun screen cap from NVidia's site:

 

Even the 3060 mobile is too beefy to run off just PCI-E.   So, a Mobile 3050 might be doable?  But nothing higher.

[porina]

3 minutes ago, tkitch said:

Fun screen cap from NVidia's site:

Doh! The numbers I looked up elsewhere are different from these, will have to edit my post :D

[tkitch]

2 minutes ago, porina said:

Doh! The numbers I looked up elsewhere are different from these, will have to edit my post :D

Also, with 3800 Cuda in the 3060 Mobile?  

You'd be looking at probably 2500 cuda in a 3050.  (Based on the 3070 to 3060 step down.)  The RTX 2060 has about 2000 already.  Just sayin'

[-rascal-]

5 minutes ago, tkitch said:

Fun screen cap from NVidia's site:

Even the 3060 mobile is too beefy to run off just PCI-E.   So, a Mobile 3050 might be doable?  But nothing higher.

 

Hmm...

 

[Mark Kaine]

1 hour ago, -rascal- said:

Using the GTX 3070 as the example...

• Making the GPU run with 75W means downclocking the GPU and Memory frequency SIGNFICANTLY
• Probably chunks of CUDA / RT cores needs to be disabled
• The full-size RTX 3070 GPU runs hot, so a tiny cooler won't be enough to keep a full RTX 3070 cool

yeah the 75w limit makes no sense - but reading the title I thought op meant something like this...

 

which would make a lot of sense! (like I guess 1500/1600mhz maximum boost, 150w power limit...) it would basically be a low power 3060ti, so why not... 

 

 

[tkitch]

2 minutes ago, Mark Kaine said:

yeah the 75w limit makes no sense - but reading the title I thought op meant something like this...

 

which would make a lot of sense! (like I guess 1500/1600mhz maximum boost, 150w power limit...) it would basically be a low power 3060ti, so why not... 

 

 

The 75 W Power Limit is because that's what the PCI-E Slot delivers.

 

most low-profile systems don't have a spare 150W to push a full GPU.  Squeezing in a 75W Card is the limit for most of them. 

 

Also, that is not a low-profile card, that's a single slot card.

[Mark Kaine]

1 minute ago, tkitch said:

The 75 W Power Limit is because that's what the PCI-E Slot delivers.

yes but you only need 75w more to 150, which is more realistic imo. 

 

3 minutes ago, tkitch said:

most low-profile systems don't have a spare 150W to push a full GPU.  Squeezing in a 75W Card is the limit for most of them. 

I mean ok, then we have to figure out how powerful such a '3070' would be - it'd be more like a 3030 I suppose... it's not a problem to make it I guess it just won't be very fast. 

 

5 minutes ago, tkitch said:

Also, that is not a low-profile card, that's a single slot card.

B... but - it's smol and looks cute! 

[Middcore]

Yep, the Katana still used an external 8-pin PCI-e power connector. It was a neat card and I wish someone would make something similar now but it's super niche.

[Mark Kaine]

3 minutes ago, Middcore said:

Yep, the Katana still used an external 8-pin PCI-e power connector. It was a neat card and I wish someone would make something similar now but it's super niche.

I mean they could even make it shorter, then it's a 'low profile' card... tho it would make even less sense imo. (since it'll run hot) 

 

but I get it op basically wants a mobile card for desktop 'mini' PC's, not sure how feasible this is... 

 

 

[porina]

If cost were no object, what could you do to make the highest performing 75W GPU? At the end of the day, the mobile GPUs aren't that different from desktop ones, it is more where they are operated than any special sauce. On nvidia side, I'd start with a 3090.

 

We can quickly save a bunch of power by simply running it with 12GB VRAM, which should still be plenty. If more power saving is needed, that could go down further. Ref Igor's Lab link below, 3090 ram uses 60W itself. If you're setting a 75W power budget, obviously that's not going to work well. I'm assuming half that capacity will also halve the power, but it might not. Ideally for performance you'd want the full bus width used, so you'd just switch to lower capacity chips. Power is used in two scenarios: when data is transferred, and when data is internally refreshed (assuming GDDR is not so different from DDR in that area). If the power consumption is from the data transfer, not refresh of the vram, the savings might not be as much as hoped. Running the vram at a lower speed could see efficiency gains. Especially with reduced core performance, the bandwidth required to feed it could be scaled back also.

https://www.igorslab.de/en/350-watts-for-nvidias-new-top-of-the-line-geforce-rtx-3090-ampere-model-explained-chip-area-calculated-and-boards-compared/

 

As discussed before, the core clocks would be run way down from desktop models. Performance is approximated by width (cores) multiplied by clock. Efficiency per core is non-linearly related to clock. That's why fewer and faster is generally less efficient than wider but slower of same overall throughput.

 

Power feeding circuits on board will probably have to be re-optimised. If you're not pushing high powers through them, the bigger VRM designs don't make sense any more. Smaller ones will need to be present and better optimised for that. I think there is a further complication in that the PCIe slot power of 75W is not entirely available on 12V, but split with 3.3V. That'll mean either running more conversion, or setting an even lower budget. I wonder how "75W" cards actually do it? It is approx. 66W on 12V rail and 9W on 3.3v rail.

 

Anyway, as discussed previously I estimate a hypothetical 75W "3070" to have about half the performance of the desktop model. That'll put it well below 3060 levels. If we assume some further efficiency from using a 3090 as basis, it might be a bit higher but I think performance will still be not that exciting. Obviously, the bigger the core, the bigger the price.

 

If anyone has an Ampere card and some time to spend, some more practical data could be obtained to see how this might go. Find a representative workload and monitor clocks, voltage and power. Reduce power limit and repeat, until you get a curve showing how that scales. You can then try extrapolating that to see what might happen if you could set it ever lower. Again, it wont be exact because design changes would be needed to go lower, so this wouldn't go down as low as it could.

 

A similar exercise could be carried out using RDNA2 based chips, but even then I'm not sure it'll get you to a much different conclusion.

 

Anyway, this is just a bit of "what if" fun for me. I didn't cover all variables I could throw at it, as it is enough of a stretch without going into finer detail

[RTX 3090]

It would be cool if they could make some sort of adapter to convert the mobile GPU's into desktop ones

edit : looks like they make them already

[spartaman64]

18 minutes ago, porina said:

If cost were no object, what could you do to make the highest performing 75W GPU? At the end of the day, the mobile GPUs aren't that different from desktop ones, it is more where they are operated than any special sauce. On nvidia side, I'd start with a 3090.

 

We can quickly save a bunch of power by simply running it with 12GB VRAM, which should still be plenty. If more power saving is needed, that could go down further. Ref Igor's Lab link below, 3090 ram uses 60W itself. If you're setting a 75W power budget, obviously that's not going to work well. I'm assuming half that capacity will also halve the power, but it might not. Ideally for performance you'd want the full bus width used, so you'd just switch to lower capacity chips. Power is used in two scenarios: when data is transferred, and when data is internally refreshed (assuming GDDR is not so different from DDR in that area). If the power consumption is from the data transfer, not refresh of the vram, the savings might not be as much as hoped. Running the vram at a lower speed could see efficiency gains. Especially with reduced core performance, the bandwidth required to feed it could be scaled back also.

https://www.igorslab.de/en/350-watts-for-nvidias-new-top-of-the-line-geforce-rtx-3090-ampere-model-explained-chip-area-calculated-and-boards-compared/

 

As discussed before, the core clocks would be run way down from desktop models. Performance is approximated by width (cores) multiplied by clock. Efficiency per core is non-linearly related to clock. That's why fewer and faster is generally less efficient than wider but slower of same overall throughput.

 

Power feeding circuits on board will probably have to be re-optimised. If you're not pushing high powers through them, the bigger VRM designs don't make sense any more. Smaller ones will need to be present and better optimised for that. I think there is a further complication in that the PCIe slot power of 75W is not entirely available on 12V, but split with 3.3V. That'll mean either running more conversion, or setting an even lower budget. I wonder how "75W" cards actually do it? It is approx. 66W on 12V rail and 9W on 3.3v rail.

 

Anyway, as discussed previously I estimate a hypothetical 75W "3070" to have about half the performance of the desktop model. That'll put it well below 3060 levels. If we assume some further efficiency from using a 3090 as basis, it might be a bit higher but I think performance will still be not that exciting. Obviously, the bigger the core, the bigger the price.

 

If anyone has an Ampere card and some time to spend, some more practical data could be obtained to see how this might go. Find a representative workload and monitor clocks, voltage and power. Reduce power limit and repeat, until you get a curve showing how that scales. You can then try extrapolating that to see what might happen if you could set it ever lower. Again, it wont be exact because design changes would be needed to go lower, so this wouldn't go down as low as it could.

 

A similar exercise could be carried out using RDNA2 based chips, but even then I'm not sure it'll get you to a much different conclusion.

 

Anyway, this is just a bit of "what if" fun for me. I didn't cover all variables I could throw at it, as it is enough of a stretch without going into finer detail

we would probably need to use something like carbon nanotubes or optical interconnects to reduce the power usage and heat output