External discreet graphics cards

[Chuckmiller1964]

I think it's time for a shift in card design. Go external. This would allow the overall design to become any size and shape. Nearly unlimited VRAM, new cooling designs, and a list of other benefits.  

 

Your thoughts? 

[tkitch]

been done, more than once, and it's terrible.

 

So no, no it won't happen.  

 

Go lookup EGPUs on youtube and see them in action.

[Chuckmiller1964]

There is no reason today it has to be terrible. It could be engineered to be great. For one, it could have it's own power supply and eliminate the 12VHPWR cable fiasco. 

[Eigenvektor]

17 minutes ago, Chuckmiller1964 said:

Nearly unlimited VRAM, new cooling designs, and a list of other benefits. 

Not going to happen. Size of the card is generally not the limiting factor when it comes to VRAM.

 

Also, eGPUs exist. Primary limitation is their lack of bandwidth/increased latency.

[Chuckmiller1964]

Just now, Eigenvektor said:

... Primary limitation is their lack of bandwidth/increased latency.

That could be engineered out. 

[tkitch]

Just now, Chuckmiller1964 said:

That could be engineered out. 

at what cost?

An internal gen 5 GPU Slot has 512gbit of bandwidth.

What external interface are you going to engineer at a reasonable price to even come close to competing with that?

The highest end USB/Thunderbolt available is a small fraction of that speed.

[Salted Spinach]

The existing industry is not alowing this to happen. They want to continue selling full size computers and components.

 

CPUs have progressed a lot, including mobile CPUs. Look at the Ryzen 9000 HX3D series.

 

The moment we get external GPU connection bandwidth on par with PCIE 3.0 x16 (which is what most GPU need to fully stretch their legs) people will start opting for laptops and mini PC with a strong CPU and compatibility with said eGPU connection. The desktop PC hardware market will then head into unstoppable decline as people realize they no longer need a desktop tower to get full gaming capability.

[tkitch]

1 minute ago, Salted Spinach said:

The existing industry is not alowing this to happen. They want to continue selling full size computers and components.

proof?

[OhYou_]

50 minutes ago, Chuckmiller1964 said:

I think it's time for a shift in card design. Go external. This would allow the overall design to become any size and shape. Nearly unlimited VRAM, new cooling designs, and a list of other benefits.  

 

Your thoughts? 

? so cap gpu performance to that of a 5070 because there is no way you're pushing pci-e gen5 speeds over anything resembling a cable or being more than 10cm from the cpu

[Alex Atkin UK]

48 minutes ago, Chuckmiller1964 said:

I think it's time for a shift in card design. Go external. This would allow the overall design to become any size and shape. Nearly unlimited VRAM, new cooling designs, and a list of other benefits.  

 

Your thoughts? 

What list of other benefits?   Most people want less cables and a single box you have to power.

Last thing I want is another PSU and cables I can accidentally unplug.  The 5090 FE showed us that the limiting factor is profit, not that you can't fit a high power GPU in a smaller form factor.

External GPUs just make things even more expensive where an AIB could put that extra cost into making 2 slot GPU if they wanted.  The fact they didn't just shows that they don't feel most end users are willing to pay the premium.

[Salted Spinach]

15 minutes ago, tkitch said:

proof?

eGPU over Thunderbolt 4 is limited to 40Gbps over 4 lanes of PCIE gen 4, which is even slower than old OCuLink

 

The 4 lane limitation is going to remain for Thunderbolt 5, meaning the bandwidth for eGPU for Thunderbolt 5 is only going to match OCuLink.

 

All this is deliberate and we are not going to get the bandwidth needed (parity with PCIE gen 3 x16) for a very long time

[tkitch]

9 minutes ago, Salted Spinach said:

eGPU over Thunderbolt 4 is limited to 40Gbps over 4 lanes of PCIE gen 4, which is even slower than old OCuLink

 

The 4 lane limitation is going to remain for Thunderbolt 5, meaning the bandwidth for eGPU for Thunderbolt 5 is only going to match OCuLink.

 

All this is deliberate and we are not going to get the bandwidth needed (parity with PCIE gen 3 x16) for a very long time

okay, technical facts check out, but...  that's about the end of it.

 

That's quite a conspiracy theory you've got yourself into.

[Tridefender]

1 hour ago, tkitch said:

at what cost?

An internal gen 5 GPU Slot has 512gbit of bandwidth.

What external interface are you going to engineer at a reasonable price to even come close to competing with that?

The highest end USB/Thunderbolt available is a small fraction of that speed.

2*SFF8654=pcie 4.0x16, since it's extending onto same gpu, so it does not need motherboard bifurcation support 

 

ofc if your motherboard supports bifurcation you can run 2 gpu, yay!

[Salted Spinach]

5 minutes ago, tkitch said:

That's quite a conspiracy theory you've got yourself into.

A multi-billion dollar industry doing what it takes to not be made obsolete is not a conspiracy theory

 

Every company involved in manufacturing the following items is going to go into the red if we no longer need a PCIE slot to get all the performance out of GPUs:

• Tower case
• Desktop motherboard
• Desktop memory modules
• Tower coolers and liquid coolers
• PC case fans
• ATX/SFX power supplies
• Aftermarket cable, DIY watercooling

[Eigenvektor]

1 hour ago, Chuckmiller1964 said:

That could be engineered out. 

You can't engineer out latency other than by decreasing the distance

[NavyCobra1417]

35 minutes ago, Salted Spinach said:

A multi-billion dollar industry doing what it takes to not be made obsolete is not a conspiracy theory

 

Every company involved in manufacturing the following items is going to go into the red if we no longer need a PCIE slot to get all the performance out of GPUs:

• Tower case
• Desktop motherboard
• Desktop memory modules
• Tower coolers and liquid coolers
• PC case fans
• ATX/SFX power supplies
• Aftermarket cable, DIY watercooling

The DIY PC market is relatively small compared to the overall computer market.

 

And the thing is, an Asus would love to sell you a $2000 laptop that you have to replace every 2 years because everything is soldered.

 

They aren't gating eGPUs so that they can sell you a $200 motherboard every 5 years

[highduke]

2 hours ago, tkitch said:

at what cost?

An internal gen 5 GPU Slot has 512gbit of bandwidth.

What external interface are you going to engineer at a reasonable price to even come close to competing with that?

The highest end USB/Thunderbolt available is a small fraction of that speed.

Is gen 5 required for GPUs though?

[tkitch]

1 minute ago, highduke said:

Is gen 5 required for GPUs though?

almost all modern CPUs, and all current Gen GPUs are now gen 5.  

[highduke]

2 hours ago, Alex Atkin UK said:

What list of other benefits?   Most people want less cables and a single box you have to power.

Last thing I want is another PSU and cables I can accidentally unplug.  The 5090 FE showed us that the limiting factor is profit, not that you can't fit a high power GPU in a smaller form factor.

External GPUs just make things even more expensive where an AIB could put that extra cost into making 2 slot GPU if they wanted.  The fact they didn't just shows that they don't feel most end users are willing to pay the premium.

They are wasting 16 gen 5 lanes on a GPU that doesn't need it and they use gen 4 for everything else that would benefit from gen 5. Faster SSDs and chipset link etc.

 

1 minute ago, tkitch said:

almost all modern CPUs, and all current Gen GPUs are now gen 5.  

My question was if it is required. They are doing it for marketing purposes, but is there any benefit to it? Specifically for current GPUs?

[Alex Atkin UK]

1 hour ago, Salted Spinach said:

A multi-billion dollar industry doing what it takes to not be made obsolete is not a conspiracy theory

 

Every company involved in manufacturing the following items is going to go into the red if we no longer need a PCIE slot to get all the performance out of GPUs:

• Tower case
• Desktop motherboard
• Desktop memory modules
• Tower coolers and liquid coolers
• PC case fans
• ATX/SFX power supplies
• Aftermarket cable, DIY watercooling

People still build DIY PCs even if they don't use an dedicated GPU, because its more flexible than a NUC-style PC.

The market for SFF PCs vs big towers is very different.

 

It just makes no sense as a SFF case that can fit a high-end CPU and GPU on an ITX motherboard is a vastly superior solution, as its not much bigger than an eGPU.  So this is really a fight between ATX and ITX.

[tkitch]

5 minutes ago, highduke said:

They are wasting 16 gen 5 lanes on a GPU that doesn't need it and they use gen 4 for everything else that would benefit from gen 5. Faster SSDs and chipset link etc.

 

My question was if it is required. They are doing it for marketing purposes, but is there any benefit to it? Specifically for current GPUs?

wait, you're saying that the GPUs don't need gen 5 and that they'd be better used on SSDs?

dude....  what?  0.o

[highduke]

4 minutes ago, tkitch said:

wait, you're saying that the GPUs don't need gen 5 and that they'd be better used on SSDs?

dude....  what?  0.o

So if you downgrade the GPU to PCIe 4.0 are you gimping your GPU's performance?

What are the gains in GPU performance having it on PCIe 5.0 instead of 4.0, compared to the gains you'd have in the rest of your system from having extra 16 PCIe 5.0 lanes all around the motherboard?

Let alone the fact that you cannot even use a second GPU on most motherboards, other x16 slots are x4 lanes tops (and 4.0). I see no justification for wasting so many of them on...what gains REALLY?

[highduke]

4.0 on x4 chipset link with everything behind it is a joke, while wasting 16 5.0 ones on irrelevant GPU gains. There's no solid justification for it.

[Eigenvektor]

14 hours ago, highduke said:

Is gen 5 required for GPUs though?

No, provided you have 16 lanes. There are multiple PCIe scaling benchmarks on TPU that show this.

 

But as I said above, you also need to take latency into account and that is not something you can solve with more bandwidth. That can only be solved by moving components closer together.

 

In any case, I don't really see what problem this is trying to solve.

 

A major issue with GPUs is availability and price, not their size. Moving them into an extra case that requires additional cooling, power and engineering will make that situation worse, not better.

 

We don't need a paradigm shift in their form factor, we need affordable GPUs.

[highduke]

1 hour ago, Eigenvektor said:

We don't need a paradigm shift in their form factor, we need affordable GPUs.

I agree with that.

Although GPU PCBs have become smaller and there's not much on them anyway, that isn't already on a motherboard. VRMs/connectors, same thing. Have motherboard GPU socket next to CPU socket, and have VRAM on GPU die (these days hardly different memory sizes for one GPU chip), that will certainly help. Same with how you can use either 65W either 170W CPU, you could also upgrade GPU chip, and bypass board partners and get GPU chips from AMD/Nvidia/Intel just like we now get CPUs from AMD/Intel.

GPU PCBs with everything extra are wasteful, and removes board partners from pulling all kinds of tricks on customers.

And a good bit of current price madness is partners abusing customers with the increased demand. There's $200-$300 extra price for same GPU chip but "better" fans and pretty lights and graphics on GPU cards. As in price difference between two GPU boards with same GPU chip from same board partner. It's crazy. Considering they sell full fledged ATX motherboards for $100 and have way more parts, including APU VRMs for video with displayport connectors, VRM heatsinks, and everything else that can be found on your $100 motherboard.