NVIDIA Might Consider Major Design Shift for Future 300 GPU Series

[Nomad76]

NVIDIA is reportedly considering a significant design change for its GPU products, shifting from the current on-board solution to an independent GPU socket design following the GB200 shipment in Q4, according to reports from MoneyDJ and the Economic Daily News quoted by TrendForce. This move is not new in the industry, AMD has already introduced socket design in 2023 with their MI300A series via Supermicro dedicated servers. The B300 series, expected to become NVIDIA's mainstream product in the second half of 2025, is rumored to be the main beneficiary of this design change that could improve yield rates, though it may come with some performance trade-offs.

According to the Economic Daily News, the socket design will simplify after-sales service and server board maintenance, allowing users to replace or upgrade the GPUs quickly. The report also pointed out that based on the slot design, boards will contain up to four NVIDIA GPUs and a CPU, with each GPU having its dedicated slot. This will bring benefits for Taiwanese manufacturers like Foxconn and LOTES, who will supply different components and connectors. The move seems logical since with the current on-board design, once a GPU becomes faulty, the entire motherboard needs to be replaced, leading to significant downtime and high operational and maintenance costs.



View at TechPowerUp Main Site | Source

 

[SL2]

Waiting for a comment asking why this isn't a thing on consumer cards.

 

[lexluthermiester]

Waiting for a comment asking why this isn't a thing on consumer cards.

I'll fall on that grenade...

Seriously? Why the hell hasn't GPU sockets been a thing? VDIMM modules? It's not like it would all that difficult..

 

[Onasi]

@lexluthermiester
VRM difference. Traces difference on the board for different memory buses and configs (and PCI-e lanes nowadays). Different sized chips requiring different cold plates on coolers since there is no unified IHS. Lots of reasons, really. They are POTENTIALLY solvable, but you would have to have a board and cooler combo that would be interchangeable between a hypothetical 4060 and 4090. This would be hilariously costly. Do we really want GPUs coating even more?

 

[Philaphlous]

I'll fall on that grenade...

Seriously? Why the hell hasn't GPU sockets been a thing? VDIMM modules? It's not like it would all that difficult..

Money? Nvidia selling directly with a GPU will grant far less money than a full board and chip from a 3rd party... Plus then NVIDIA gets to control the wholesale GPU chip price and 3rd party gets to set the margins..win win for the middleman...

 

[Daven]

@lexluthermiester
VRM difference. Traces difference on the board for different memory buses and configs (and PCI-e lanes nowadays). Different sized chips requiring different cold plates on coolers since there is no unified IHS. Lots of reasons, really. They are POTENTIALLY solvable, but you would have to have a board and cooler combo that would be interchangeable between a hypothetical 4060 and 4090. This would be hilariously costly. Do we really want GPUs coating even more?

Traditional PC architectures haven't been significantly revamped since the beginning of consumer client desktops. While you are correct, that just adding another socket to existing motherboards is probably not the answer, we can create a two daughter board standard connection through x16 PCIe Gen 5 interlink. Each daughter board can have up its own socket, memory slots, I/O connections and even separate power supplies/delivery.

Cost is hard to gauge as you could have two 600W PSU's (one for CPU daughter board and one for GPU daughter board) which can be cheaper than one 1.2 kW power supply. The ability to upgrade just the VRAM and GPU itself could generate cost savings over time.

Either way, we are stuck in the socketed CPU, DIMM slot, multiple PCIe slots off of one motherboard era as it has been since the 80386 days (circa late 80s). Only the integration of I/O to the motherboard has really changed from those days requiring less expansion slots. Oh and I'm glad the era of jumpers is over. I hated those little plastic pieces of you know what.

 

[Caring1]

@lexluthermiester
VRM difference. Traces difference on the board for different memory buses and configs (and PCI-e lanes nowadays). Different sized chips requiring different cold plates on coolers since there is no unified IHS. Lots of reasons, really. They are POTENTIALLY solvable, but you would have to have a board and cooler combo that would be interchangeable between a hypothetical 4060 and 4090. This would be hilariously costly. Do we really want GPUs coating even more?

Why not make the GPU socket a standard and all GPUs have to comply so one cooler would suffice for any unit used?

 

[Onasi]

@Caring1
Because it doesn’t work this way even with CPUs? Good luck cooling a 14900KS with a stock cooler. Same here - if you have a hypothetical 600W 5090 and standardize with THAT in mind you’d have to make it so that “any cooler” can reasonably keep THAT cool. Which is costly. I suppose the argument can be made for establishing a ceiling for power usage on GPUs and going from there, but it wouldn’t fly.

Essentially, socketed GPUs, be it for AIBs or directly onto MoBos, would require a complete overhaul of everything about the modern PC form-factor, as @Daven has mentioned. And it’s just not really feasible as it stands right now - ATX and its offspring has a profoundly ingrained effect on how every single PC part is created and functions. And it has been this way for decades now. Shifting the entire ecosystem just for some potential upgradeability benefits wouldn’t fly with any of the current players.

 

[Caring1]

@Caring1
Because it doesn’t work this way even with CPUs? Good luck cooling a 14900KS with a stock cooler. Same here - if you have a hypothetical 600W 5090 and standardize with THAT in mind you’d have to make it so that “any cooler” can reasonably keep THAT cool. Which is costly. I suppose the argument can be made for establishing a ceiling for power usage on GPUs and going from there, but it wouldn’t fly.

Essentially, socketed GPUs, be it for AIBs or directly onto MoBos would require a complete overhaul of everything about the modern PC form-factor, as @Daven has mentioned. And it’s just not really feasible as it stands right now - ATX and its offspring has a profoundly ingrained effect on how every single PC part is created and functions. And it has been this way for decades now. Shifting the entire ecosystem just for some potential upgradeability benefits wouldn’t fly with any of the current players.

You can use the same socket for anything from an i3 to an i9, the cooler can be whatever you want it to be.
The same should apply to GPUs.

 

[Vya Domus]

Yeah not happening, part of Nvidia's strategy is making upgrading the hardware as expensive as possible because customers don't have a choice.

 

[Onasi]

@Caring1
I mean, sure, in that you can theoretically put the 14900K into the shittiest board possible and slap an Intel stock cooler on top of it. You won’t be able to actually access the full potential of the chip, but I suppose by the metric of “it can be physically done” it works.
You also haven’t provided any actual argument other than “it should be like this just cause” which… okay. I will be sure to forward your feedback to whoever it may concern.

 

[lexluthermiester]

I'm not going to respond to anyone separately, instead offering the following thought to all: Anything we can do with CPUs, we can do just as easily with GPUs. It's a design shift and spec that should have happened decades ago. Build a GPU daughter-card that has a socket and VDIMMs, have it plug into the motherboard with power circuitry to support whatever you want to install. I mean after all, we have the MXM standard, and that's fairly close.

 

[Jism]

Waiting for a comment asking why this isn't a thing on consumer cards.

It was.

90's had their own chip(s) with each and every function. Dozens of examples, 3DFX for example. Chips and tech got advanced, now pretty much house everything inside of it, but Nvidia's approach right now is extremely inefficient. High failure rates on wafers is expensive. AMD on the other hand now makes the compute dies seperate from the memory cache dies - yielding much better on wafers and thus get more out of it.

 

[SL2]

I'm not going to respond to anyone separately, instead offering the following thought to all: Anything we can do with CPUs, we can do just as easily with GPUs.

As if anyone questioned the hypotetical possibility to do it.

 

[OSdevr]

Surprised this is possible given the speed of modern VRAM, unless they're thinking everything will use HBM. Not sure how practical GPU sockets would be on consumer GPUs with separate VRAM for signal integrity reasons, that was ostensibly the reason for ditching upgradable VRAM so long ago (it used to be a thing).

 

[Vayra86]

Ah, improved yields! Cheaper GPUs!

Stop laughing pls!