Question 'Ampere'/Next-gen gaming uarch speculation thread

[Ottonomous]

How much is the Samsung 7nm EUV process expected to provide in terms of gains?
How will the RTX components be scaled/developed?
Any major architectural enhancements expected?
Will VRAM be bumped to 16/12/12 for the top three?
Will there be further fragmentation in the lineup? (Keeping turing at cheaper prices, while offering 'beefed up RTX' options at the top?)
Will the top card be capable of >4K60, at least 90?
Would Nvidia ever consider an HBM implementation in the gaming lineup?
Will Nvidia introduce new proprietary technologies again?

Sorry if imprudent/uncalled for, just interested in the forum member's thoughts.

 

[Stuka87]

Agreed. My PSU can handle 4x 8-pins... why the move to a non-standard plug? I don't see the benefit to doing so.

At one point in time people said the same thing about PCIe power connectors. "Why have a special connector that requires a new PSU when we are already using molex connectors?"

If this connector turns out to be real, I would be willing to bet its not an nVidia only "standard", but is actually part of the ATX 12VO which OEM's are switching too (And will eventually come to DIY PSU's).

 

[uzzi38]

One thing really doesn't make sense to me, why wouldn't this be on the A100?

826mm, 40GB HBM2, five figure price, deployed in machines purpose built to power the GPU, and yet they went with a lower tune to keep the power envelope at 250 watts.

Something really doesn't add up here.

Server clients and adopting new standards quickly are two things that don't go well together.

 

[uzzi38]

https://twitter.com/x/status/1283663908096466945

Interesting thread on the state at TSMC. As stated, they beat on 2Q and have given strong guidance on H2. As for how capacity is being filled up from declining smartphone market and Huawei ban? Short term there will be a hit, long term demand for 5G and HPC will fill up the rest of the capacity.

No mention of consumer parts.

 

[ozzy702]

At one point in time people said the same thing about PCIe power connectors. "Why have a special connector that requires a new PSU when we are already using molex connectors?"

If this connector turns out to be real, I would be willing to bet its not an nVidia only "standard", but is actually part of the ATX 12VO which OEM's are switching too (And will eventually come to DIY PSU's).

I understand, but why release this when there's not a single available power supply to drive it without adapters? Just seems like a strange move. Do two 8 pins and a six pin if you need overhead, or three 8 pins. Once ATX 12V0 PSUs are more standard then by all means move to it. Just seems like jumping the gun.

As long as GPUs come with the adapters I suppose it's not a big deal.

 

[Stuka87]

I understand, but why release this when there's not a single available power supply to drive it without adapters? Just seems like a strange move. Do two 8 pins and a six pin if you need overhead, or three 8 pins. Once ATX 12V0 PSUs are more standard then by all means move to it. Just seems like jumping the gun.

As long as GPUs come with the adapters I suppose it's not a big deal.

The founders edition cards while sold stand alone to consumers, are the primary cards that OEM's use. OEM's are switching to 12VO before DIY PSU's will be. So from that stand point, it makes sense. OEM's are the main driver behind 12VO, because it makes wiring simpler. For them, having some cards that use 6pin, some use 8, some use both, it means they need PSU's that have all of those. It is way cheaper and simpler to just have a single power connector that works for everything.

But on the DIY side, it will be a while before it becomes standardized as it also required a new motherboard that supports the standard. But in the end, I think it will be better and easier for all. But the middle area will kind of suck.

 

[Kenmitch]

The founders edition cards while sold stand alone to consumers, are the primary cards that OEM's use. OEM's are switching to 12VO before DIY PSU's will be. So from that stand point, it makes sense. OEM's are the main driver behind 12VO, because it makes wiring simpler. For them, having some cards that use 6pin, some use 8, some use both, it means they need PSU's that have all of those. It is way cheaper and simpler to just have a single power connector that works for everything.

I was thinking similar thoughts. I'd imagine they'll include a cheesy looking adapter in the box. I'm leaning towards Nvidia also selling a nice spiffy looking one at the same time....Of course they'll dabble on a nice helping of Nvidia tax if so. I guess in the end it is what it is.

 

[ozzy702]

The founders edition cards while sold stand alone to consumers, are the primary cards that OEM's use. OEM's are switching to 12VO before DIY PSU's will be. So from that stand point, it makes sense. OEM's are the main driver behind 12VO, because it makes wiring simpler. For them, having some cards that use 6pin, some use 8, some use both, it means they need PSU's that have all of those. It is way cheaper and simpler to just have a single power connector that works for everything.

But on the DIY side, it will be a while before it becomes standardized as it also required a new motherboard that supports the standard. But in the end, I think it will be better and easier for all. But the middle area will kind of suck.

Ah, that does make sense for the founders cards. I wonder if the AIBs will just go with 8 pins then.

 

[FaaR]

It is way cheaper and simpler to just have a single power connector that works for everything.

There's different connectors for different power levels. One size fits all connector only works if every power supply is upsized to deliver 600W (or whatever will be peak power of this connector.) Otherwise users will either short out the power supply or trip the OCP when pulling more juice than the PSU can actually deliver if they install an oversized GPU board in their low-end or mid-range OEM PC...

So there's going to have to be different power connectors for different performance tiers regardless (because OEMs like to skimp on components like PSUs when they can), and then we're back to square one. Modular power supplies mated with traditional PCIe power connectors can easily adapt to a broad range of power consumption conditions. One huge outsized power connector is no inherent advantage there I'd say.

 

[Stuka87]

There's different connectors for different power levels. One size fits all connector only works if every power supply is upsized to deliver 600W (or whatever will be peak power of this connector.) Otherwise users will either short out the power supply or trip the OCP when pulling more juice than the PSU can actually deliver if they install an oversized GPU board in their low-end or mid-range OEM PC...

So there's going to have to be different power connectors for different performance tiers regardless (because OEMs like to skimp on components like PSUs when they can), and then we're back to square one. Modular power supplies mated with traditional PCIe power connectors can easily adapt to a broad range of power consumption conditions. One huge outsized power connector is no inherent advantage there I'd say.

GPU's already suggest certain power supply wattage requirements. Just because a PSU has dual 8 pin connectors, does not mean that rail has the power to fully supply both 8 pin connectors.

There is no reason to have different connectors. If an RTX 3070 says it requires at least a 600W PSU, thats what the user goes off of. Not the connectors that happen to be on it.

 

[maddie]

There's different connectors for different power levels. One size fits all connector only works if every power supply is upsized to deliver 600W (or whatever will be peak power of this connector.) Otherwise users will either short out the power supply or trip the OCP when pulling more juice than the PSU can actually deliver if they install an oversized GPU board in their low-end or mid-range OEM PC...

So there's going to have to be different power connectors for different performance tiers regardless (because OEMs like to skimp on components like PSUs when they can), and then we're back to square one. Modular power supplies mated with traditional PCIe power connectors can easily adapt to a broad range of power consumption conditions. One huge outsized power connector is no inherent advantage there I'd say.

Shouldn't there be a minimum level for making products idiot proof? Want to fiddle, then learn something first.

 

[FaaR]

Just because a PSU has dual 8 pin connectors, does not mean that rail has the power to fully supply both 8 pin connectors.

What quality power supply has more power connectors than juice to supply them? That's nuts. Stuff like that is what you find on noname chinese junk, if there's two 8-pins on there it absolutely suggests they can both be used. Obviously. If the power supply then shuts down (or worse, burns out) it is obviously faulty as designed. The spec says 8-pin can deliver up to 150W per connector, so the PSU better be able to fulfil that. Otherwise it's out of spec.

 

[Stuka87]

What quality power supply has more power connectors than juice to supply them? That's nuts. Stuff like that is what you find on noname chinese junk, if there's two 8-pins on there it absolutely suggests they can both be used. Obviously. If the power supply then shuts down (or worse, burns out) it is obviously faulty as designed. The spec says 8-pin can deliver up to 150W per connector, so the PSU better be able to fulfil that. Otherwise it's out of spec.

There are plenty of PSU's that could easily supply 300W for both 8 pin connectors. However, being able to supply peak current to every single connector at once, not so much. Once you add in the 24pin mobo connection, the 8pin CPU 12V feed, all the molex/SATA power connectors, and the PCIe connectors, quite a few PSU's that fail in this regard.

So, just because a future PSU has the new 12pin connector, does not mean it can supply full current. This depends on the size of the PSU, just as it does now.

 

[FaaR]

So, just because a future PSU has the new 12pin connector, does not mean it can supply full current. This depends on the size of the PSU, just as it does now.

You're using cheapo, out of spec power supplies as an example of how a 600W power connector could be used to "work" with any size power supply? Wow, that's a baaaad idea.

A power supply which could not supply 600W to a 600W power connector would still be out of spec. If the connector is on the PSU, it better match spec. Or bad things will start to happen, I wouldn't be surprised if what you suggest would even be illegal. If there aren't any electricity-related safety regs being violated, it sure sounds fraudulent to me.

 

[Stuka87]

You're using cheapo, out of spec power supplies as an example of how a 600W power connector could be used to "work" with any size power supply? Wow, that's a baaaad idea.

A power supply which could not supply 600W to a 600W power connector would still be out of spec. If the connector is on the PSU, it better match spec. Or bad things will start to happen, I wouldn't be surprised if what you suggest would even be illegal. If there aren't any electricity-related safety regs being violated, it sure sounds fraudulent to me.

Ok, let make this a bit easier for you. Below is representative of your average 650W Gold rated PSU. Not some "cheap chinese one".

24pin: 3.3V-12A, 5V-24A, 12V-22A = 355W
EPS 8 pin: 235W
PCIe 8 pin: 150 x2
SATA: 54W each x4

Grand Total: 1106W

So again, just because a connector can handle X amount of power, doesn't mean the PSU has to.

 

[FaaR]

So again, just because a connector can handle X amount of power, doesn't mean the PSU has to.

That's still a different matter compared to what's being discussed, here you're adding together all the different power connectors of a power supply. But this is about one specific (set of) connector(s) here - and particularly when it's a friggin' 600W connector it would be ridiculous to put that one on a 650W PSU because it would be more than obvious to any reasonably knowledgeable person the PSU could not handle it.

Meanwhile there's all kinds of computer greenhorns out there who can't tell their molexes from their EPS12v plugs who would assume that if the shoe fits, then it works, and happily plug their 650W PSU via that monster 600W plug into a high-powered GPU and end up with a repeatedly crashing or even dead system.

I can guarantee you that situation was never the intent of whomever designed that plug, or even anything near the like. It's obvious to me you will never natively find that plug on any PSU which does not fulfil minimum spec for any high-end GPU that will use the plug. So like, 800W+ capacity at least, and quite possibly more considering power draw of the rest of a high-end PC plus reasonable overprovisioning.

Oh, and lay off the condescension, thanks. That's not needed.

 

[peonyu]

Is there a good chance that these new cards can deliver 2x the performance of a GTX 1070 for under $500? I know that no benchmarks are out yet but this is a new architecture and it is 2 generations removed from the 1070.

 

[psolord]

Is there a good chance that these new cards can deliver 2x the performance of a GTX 1070 for under $500? I know that no benchmarks are out yet but this is a new architecture and it is 2 generations removed from the 1070.

As a 1070 owner, I am looking for the same thing. The 2070 is a little less than 50% faster than the 1070 already, while the 2070 super, is solidly above that. So after that 1.5X of processing power, you need another 1.33X to reach 2X of the 1070. I guess 1.33X is doable for ampere.

What bugs me is the video ram. Cause I ain't paying 500 for 8GBs again. No way man. 10GBs bare minimum, 12GBs acceptable, 16GBs would be a bomb! These, paired with 32GBs of RAM, would provide plenty of addressable space for data to move around.

 

[BFG10K]

Wow. I’m kind of excited to see what kind of GPU is crazy enough to bring it’s own 600w connector to the table.

I'd use the word "horrified" instead. This is trending to put multi-GPU power consumption onto a single card, which is a very bad direction. It's absolutely nasty gaming on a furnace. 8800U, 6800U and Fermi are the worst video cards I've ever used for this reason. I hated them.

Hard to believe the 9700 Pro which absolutely dominated its generation only needed a floppy power connector and a tiny cooler.

Ahhahaha. More proprietary NV BS. Undoubtedly, they'll "offer" this connector to PSU manufacturers for an outsized "licensing fee" (AKA JHH's vig.)

We'll they locked motherboards and monitors to their cards in the past, so maybe PSUs are next.

 

[MrTeal]

Based on what? I'm feeling a bit sceptical over this whole thing, it would be a big departure, introducing a new connector that nobody (outside possibly some few industry insiders) has ever seen or heard of, and doing so before there's any ecosystem in the wild.

https://www.molex.com/pdm_docs/ps/2064600000-PS-000.pdf

With tin connectors and 20 AWG wire, it's 4 amps per pin. That gives 288W for the connector. Mandating 18AWG or gold plated connectors would get you to 400W though.

I also am not sure on this. It seems like something that would appear in a full new revision of the ATX standard, not randomly appear mid cycle. I would be interested to see what they propose if they go this route though. Perhaps a 12 pin @ 288W, with a 6 pin @ 144W where each cable could have 6+6.

Edit: I should probably add that the picture in the FCPowerUp report isn't some inside picture from nvidia, it's the exact picture on Molex's product page for that product.

https://www.molex.com/molex/products/part-detail/crimp_housings/0430251200

 

[DeathReborn]

I'd use the word "horrified" instead. This is trending to put multi-GPU power consumption onto a single card, which is a very bad direction. It's absolutely nasty gaming on a furnace. 8800U, 6800U and Fermi are the worst video cards I've ever used for this reason. I hated them.

Hard to believe the 9700 Pro which absolutely dominated its generation only needed a floppy power connector and a tiny cooler.


We'll they locked motherboards and monitors to their cards in the past, so maybe PSUs are next.

Just as well you avoided the 2900 XT as it was slower & used more power than the 8800U. It does seem like the 50W or less Graphics Card has been pretty much devoured by IGP's which is a shame really.

 

[Bubbleawsome]

I'd use the word "horrified" instead. This is trending to put multi-GPU power consumption onto a single card, which is a very bad direction. It's absolutely nasty gaming on a furnace. 8800U, 6800U and Fermi are the worst video cards I've ever used for this reason. I hated them.

It’s more that I doubt Nvidia would decrease perf/watt during a shrink to 7/8nm. Double 2080ti power next gen should mean more than double 2080ti performance, hopefully by a fair bit. If Nvidia thinks launching a GPU that crazy is worth it, then that gives me hope for AMD’s offering this time around as well. So until further notice, still kinda excited.

 

[jpiniero]

It’s more that I doubt Nvidia would decrease perf/watt during a shrink to 7/8nm. Double 2080ti power next gen should mean more than double 2080ti performance, hopefully by a fair bit. If Nvidia thinks launching a GPU that crazy is worth it, then that gives me hope for AMD’s offering this time around as well. So until further notice, still kinda excited.

Seems the perf/watt will be about the same, at least in raster games. Samsung said SS7 used 30% less power than SS10 which explains why it needs to be over 300 W.

All of the GA102 will still be less than what Dual 8pin gives (375 W total), I would have to think... regardless of what the power connector can do.

 

[CastleBravo]

It’s more that I doubt Nvidia would decrease perf/watt during a shrink to 7/8nm. Double 2080ti power next gen should mean more than double 2080ti performance, hopefully by a fair bit. If Nvidia thinks launching a GPU that crazy is worth it, then that gives me hope for AMD’s offering this time around as well. So until further notice, still kinda excited.

I'm sure Nvidia would love to launch a GPU that powerful. I bet they would sell a lot more "Titan" cards if that ~$2k price actually provided a significant increase in performance. I just don't think a card that powerful is feasible without MCM. Maybe they are planning a future 500+ watt MCM monster card that will fill in the market void left by the death of SLI/Crossfire?

 

[FaaR]

I'd use the word "horrified" instead. This is trending to put multi-GPU power consumption onto a single card, which is a very bad direction. It's absolutely nasty gaming on a furnace.

Friggin' aye. I've had a number of hot-running PCs in my life, and it always sucks.

With tin connectors and 20 AWG wire, it's 4 amps per pin. That gives 288W for the connector. Mandating 18AWG or gold plated connectors would get you to 400W though.

Hm, pretty sure I saw one of the images (linked on TPU's news post page about this connector IIRC) stating 9A per pin? If all of the pins in the connector are dedicated to power delivery, that's almost 650W actually.

...AKA, not a graphics card I would ever want to own. lol We badly need a new form factor for GPU boards, the card slots dating back to the original IBM PC just don't cut it with available space restraint and air flow/cooling ability. Power delivery is the easy part: dissipating all the resulting heat is...

If nothing is done, unless all high-end GPUs start coming with their own built-in closed loop water coolers there's going to be a definitive cap on GPU performance pretty soon the way power consumption keeps going up despite process shrinks.

...And transistors can't shrink for much more anyhow, regardless of what we do. Soon they get so small they stop working properly and that's the end of the road for "free" performance gains. After that, fundamentally (assuming well-engineered hardware of course), the only way to make things faster is adding more compute hardware, and spending more power...

 

[BenSkywalker]

Seems the perf/watt will be about the same

The same as what?