Discussion Ada/'Lovelace'? Next gen Nvidia gaming architecture speculation

[Thunder 57]

There has literally never been a 1:1 ratio for performance to price as you move up tiers. The higher the tier, the lower the performance per dollar.

There may have been a time it has happened, though admittingly nothing comes to mind immediately. Wouldn't the 4090 compared to the 4080 break this rule though?

 

[Stuka87]

There may have been a time it has happened, though admittingly nothing comes to mind immediately. Wouldn't the 4090 compared to the 4080 break this rule though?

Yeah, maybe it has happened way back. But going back to at least the Radeon 5K GTX 400 series, performance per dollar gets worse the higher the tier is.

If you are talking the 12GB 4080, that really should be a 4070, and priced way lower than it is, then it looks like it might go against this. But we really don't know the real performance of any of the cards yet. Just what nVidia has shown (which is moot), and then a few videos like DF's video, but that video only showed percentages.

 

[yottabit]

Yeah, maybe it has happened way back. But going back to at least the Radeon 5K GTX 400 series, performance per dollar gets worse the higher the tier is.

If you are talking the 12GB 4080, that really should be a 4070, and priced way lower than it is, then it looks like it might go against this. But we really don't know the real performance of any of the cards yet. Just what nVidia has shown (which is moot), and then a few videos like DF's video, but that video only showed percentages.

If I compare against the X850 XT PE I bought from ZipZoomFly, the 4090 is a great bargain.

That card only had 16 fixed function shaders for $420. The 4090 has 16k shaders for only 4x the price. That’s 100,000% more shaders!

Although I’ll have to dock it a few imaginary perf/$ points because it doesn’t have the All-In-Wonder VIVO capture. Going to make it hard to capture S-Video with the 4090

I just don’t get the sudden focus on this perf/$ nonsense for the halo products. I mean, if it turns out they really are crap at raster I would agree. But with the increase in clocks alone I just don’t see that happening, and pre-emptive claims of poor performance for price seem unfounded

Btw, some sleuthing revealed Ada has a much larger L2 cache? That could help explain the… odd choices in memory bus width. The 256 bit bus is the most concerning thing about even the 4080 16 GB for me as I’m primarily interested in 4K & VR. That bus width is, in fact shared by the old X850 XT PE

I suspect a larger cache would cover up any issues keeping the cores fed at lower resolutions. But NV probably knows what they are doing better than me

 

[igor_kavinski]

If I compare against the X850 XT PE I bought from ZipZoomFly, the 4090 is a great bargain.

You didn't buy any newer gen card after that?

 

[CakeMonster]

Although I’ll have to dock it a few imaginary perf/$ points because it doesn’t have the All-In-Wonder VIVO capture. Going to make it hard to capture S-Video with the 4090

Damn, that brings me back, I had a TNT2U with some amazing capture functions (although learning the quirks of that chip and software was hell), I even sold it for a very good price when I was done with it.

 

[jpiniero]

If you are talking the 12GB 4080, that really should be a 4070, and priced way lower than it is, then it looks like it might go against this. But we really don't know the real performance of any of the cards yet. Just what nVidia has shown (which is moot), and then a few videos like DF's video, but that video only showed percentages.

We have a pretty good idea. The raw specs are nearly identical to the 3090 Ti (except of course it has half the memory bandwidth). nVidia's own marketing claimed it was around the 3090 Ti's performance without using DLSS3.

 

[sniffin]

Now quality, I haven't examined

How many posts about Game Motion Plus and then you say this? It was hot garbage when I tried it. Terrible ghosting and there was judder when the frame rate wasn't perfectly locked. From the DF footage DLSS 3 looks far superior. It is hard to tell how smooth it is through YT, but Alex describes a far better experience than what I had with GMP.

but as defenders of the NVidia fake frames have been pointing out,

I love the accusations of shilling when you so clearly have your own agenda.

 

[Heartbreaker]

How many posts about Game Motion Plus and then you say this? It was hot garbage when I tried it. Terrible ghosting and there was judder when the frame rate wasn't perfectly locked. From the DF footage DLSS 3 looks far superior. It is hard to tell how smooth it is through YT, but Alex describes a far better experience than what I had with GMP.

I have also seen many YT videos of people raving about Game Motion Plus on Samsung TV's like this one:

It's the same technique, having the same lag, and according to TV owners on YT, showing no visible artifacts. I seriously doubt many people could tell the difference between Game Motion Plus frame interpolation and DLSS 3 frame interpolation.

I love the accusations of shilling when you so clearly have your own agenda.

Where did I say anyone here was shilling?

What agenda would that be?

Being against misleading hype? You got me. I confess. I'm totally against misleading hype.

I just said NVidia frame interpolation should called motion interpolation or frame smoothing. Because that's what it is. Not be rolled into the the overall frame rate, because it's NOT the real frame rate, and doesn't behave like a real frame rate.

To claims it's the real frame is misleading hype. I don't like misleading hype.

 

[jpiniero]

https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvidia-ada-gpu-architecture.pdf

Ada Whitepaper containing specs. Note they compare the 4080 16 GB to the 3080 Ti and the 4080 12 GB to the 3080 12 GB. Which... given the prices actually makes sense.

Seeing the specs like this really makes it look like memory bandwidth is an issue.

 

[Timmah!]

Do you guys think its possible to mount 2x watercooled 4090s into full tower case? Like the MSI one for example? I can see the upper card would have the radiator mounted up top, but where would you mount the rad for the bottom one? Anyway, would it be stupid to have 3 aios inside single case (420 for cpu and 2x 240 for gpus?)

i dont have experience with custom loop, how much it costs, is it expensive? Not taking waterblocks for cards into account, just pump, radiator, reservoir, tubing and all the jazz needed to make it work?

 

[dlerious]

Do you guys think its possible to mount 2x watercooled 4090s into full tower case? Like the MSI one for example? I can see the upper card would have the radiator mounted up top, but where would you mount the rad for the bottom one? Anyway, would it be stupid to have 3 aios inside single case (420 for cpu and 2x 240 for gpus?)

You want the pump at the lowest point in the loop, so top mount and front mount. GPU's are drawing 2x or more power than CPU, so you want more surface area for GPUs. As for expense, I haven't priced out a new loop in a while, but it cost more than AIO while being more flexible (IMO).

 

[jpiniero]

Do you guys think its possible to mount 2x watercooled 4090s into full tower case? Like the MSI one for example? I can see the upper card would have the radiator mounted up top, but where would you mount the rad for the bottom one? Anyway, would it be stupid to have 3 aios inside single case (420 for cpu and 2x 240 for gpus?)

i dont have experience with custom loop, how much it costs, is it expensive? Not taking waterblocks for cards into account, just pump, radiator, reservoir, tubing and all the jazz needed to make it work?

At 2x600 W, that's not going to leave a lot of room left. Assuming you are using a 1500 W PSU. You might have to opt for the 450 W models. Which might be a better deal anyway.

 

[maddie]

At 2x600 W, that's not going to leave a lot of room left. Assuming you are using a 1500 W PSU. You might have to opt for the 450 W models. Which might be a better deal anyway.

People still doing multi-PSUs? A 1500W is inadequate for this use case, because of rest of the system.

 

[Timmah!]

People still doing multi-PSUs? A 1500W is inadequate for this use case, because of rest of the system.

I intend to go with 1600w seasonic prime platinum and i hope/believe its gonna be enough. I count with the 450w power-draw for the cards, i believe thats the default state anyway, 600w is for some OC bios and whatnot, but i am not going to do that. I dont need it to be as fast as possible, what i care about is stability and reasonable temps and noise levels.

 

[maddie]

I intend to go with 1600w seasonic prime platinum and i hope/believe its gonna be enough. I count with the 450w power-draw for the cards, i believe thats the default state anyway, 600w is for some OC bios and whatnot, but i am not going to do that. I dont need it to be as fast as possible, what i care about is stability and reasonable temps and noise levels.

I assume that you will want a high power CPU to push those. So, another 250W + rest of system = 1200W at max? Are power spikes no longer an issue with Nvidia?

 

[Timmah!]

I assume that you will want a high power CPU to push those. So, another 250W + rest of system = 1200W at max? Are power spikes no longer an issue with Nvidia?

7950x/x3D. And i wont hesitate to run it in eco mode, if needed.
I dont think i will ever be running all the components under heavy load at the same time, it will either be:

gaming - CPU and one GPU at the time, additionally gaming is not usually as heavy as the likes of Cinebench or Prime95
rendering - there is a part, where the scene is "built" before rendering, this is multi-threaded and taxes all the cores, however GPUs are idle at the time
rendering - actual rendering process, where GPUs are loaded, but CPU only lightly

i truly dont see any other use-case, where all of these would be fully loaded, except maybe some artificial benchmark, like TimeSpy, and i dont need to run those.

Regarding power-spikes, apparently this has been fixed, or at least significantly reduced, at least per Nvidia´s own marketing. Remains to be seen, if its true, when the actual independent reviews pop-up, but i am inclined to believe it. However, it concerns strictly FE card, which i wont be getting, so hopefully this will be true for AIB versions as well.

https://i.ibb.co/1f1jQgR/RTX4090-PCB.jpg

 

[igor_kavinski]

i truly dont see any other use-case, where all of these would be fully loaded, except maybe some artificial benchmark, like TimeSpy, and i dont need to run those.

There's just the issue of some malware blowing up your PSU by mining crypto during idle hours (at night, if you leave your PC on). Make sure to harden your network and Windows security.

 

[coercitiv]

rendering - actual rendering process, where GPUs are loaded, but CPU only lightly

I think you should be fine, not only because Nvidia claims improved transient handling, but rather because of the headroom you will have built in.

According to Seasonic, the bigger brother of the PX 1600W that is the TX 1600W can meet the ATX v3 specs for up to 1000W (this means handling ~2000W transients for 100us). The review in the link which quotes Seasonic also tests OPP for the TX 1600, which goes slightly above 2000W. I expect the Platinum variant to be close in terms of max power delivery.

You'll still need to keep an eye on how 4000 series cards behave in relation to the new power delivery standard (presence or absence of the additional 4 sensing pins on the 12HPWR connectors), the information available out there is a bit of a mess and your planned build may test the more extreme cases of the spec sheet.

 

[Timmah!]

I think you should be fine, not only because Nvidia claims improved transient handling, but rather because of the headroom you will have built in.

According to Seasonic, the bigger brother of the PX 1600W that is the TX 1600W can meet the ATX v3 specs for up to 1000W (this means handling ~2000W transients for 100us). The review in the link which quotes Seasonic also tests OPP for the TX 1600, which goes slightly above 2000W. I expect the Platinum variant to be close in terms of max power delivery.

You'll still need to keep an eye on how 4000 series cards behave in relation to the new power delivery standard (presence or absence of the additional 4 sensing pins on the 12HPWR connectors), the information available out there is a bit of a mess and your planned build may test the more extreme cases of the spec sheet.

Yeah, thats the only worry, the lack of the new connector on those older PSUs, like the one i consider to get.
Someone on Reddit actually gave me a heads-up, there is 1600W PSU with 2 new connectors, Thermaltake one. But i dont know, not really familiar with that brand, are their PSUs any good?

This one:

Toughpower GF3 1650W Gold - TT Premium Edition

Designed with a native PCIe 12+4pin connector, the Toughpower GF3 series is compatible with Intel ‘s ATX 3.0 specification and is built for next-gen PCIe 5.0 graphics cards. The wattage marked on 12VHPWR’s connector is the recommended wattage. (Ref. Intel design guideline) The native 16 PIN...

www.thermaltake.com

 

[igor_kavinski]

But i dont know, not really familiar with that brand, are their PSUs any good?

They should be fine but they are not highly regarded in PSUs. One review about a kilowatt plus PSU model complained about hotspots and noisy power output.

 

[dlerious]

HWBusters recommendations

Best PSUs Picks 2023 – Hardware Busters - Hardware Busters

Hardware Busters - Best PSUs Picks 2023 – Hardware Busters - Best Picks

hwbusters.com

 

[Leeea]

There's just the issue of some malware blowing up your PSU by mining crypto during idle hours (at night, if you leave your PC on). Make sure to harden your network and Windows security.

I have repeatedly pushed PSUs past their limits on a mining rig.

Does not really hurt anything with a brand name unit. Resets after a few minutes.

 

[Leeea]

I intend to go with 1600w seasonic prime platinum and i hope/believe its gonna be enough. I count with the 450w power-draw for the cards, i believe thats the default state anyway, 600w is for some OC bios and whatnot, but i am not going to do that. I dont need it to be as fast as possible, what i care about is stability and reasonable temps and noise levels.

7950x/x3D. And i wont hesitate to run it in eco mode, if needed.

Your doing this wrong.

Your spending a few thousand $ on this system, and skimping on the power supply.


Just spend $200 - $400 and have a licensed electrician run you a 240v outlet. Be sure to have them install the right type of outlet. Buy the PSU first and show them the plug. Be sure to have the spec sheet so you can show him it will run at 240v 60hz. They may replace the plug on your PSU.


Get a mining PSU. Models are available at 2000w, 2400w, 2500w, and 3000w.


If you want brand name you recognize, EVGA has a 2000w model:
https://www.newegg.com/p/1HU-00J7-009F3

 

[Timmah!]

Your doing this wrong.

Your spending a few thousand $ on this system, and skimping on the power supply.


Just spend $200 - $400 and have a licensed electrician run you a 240v outlet. Be sure to have them install the right type of outlet. Buy the PSU first and show them the plug. Be sure to have the spec sheet so you can show him it will run at 240v 60hz. They may replace the plug on your PSU.

Spoiler: EU outlets in US discussion

https://diy.stackexchange.com/questions/237632/can-i-install-a-240v-outlet-in-my-new-us-home-for-my-uk-kitchen-appliances

Then get a mining PSU. Models are available at 2000w, 2400w, 2500w, and 3000w.


If you want brand name you recognize, EVGA has a 2000w model:
https://www.newegg.com/p/1HU-00J7-009F3

Thank you for suggestions.

I do however disagree that i am skimping on PSU with 1600W Seasonic one It costs as much as that Evga 2K one.

Regarding electricity things, outlets and whatnot, will look into this, i am acquainted with bunch of construction/electrical engineers due to my profession, so will consult with them. I dont think any outlet changes will be needed, 1600W~2KW is inline with some other regular home appliances, like ovens, space heaters and whatnot, which you can plug into those outlets no problem.

Anyway, back on topic of custom watercooling. I literally dont know what to buy, if i wanted to set one up. Well i know i would need>

- obviously waterblocks for cards, or card with premade block, like for example Inno3D iChill Frostbite
- reservoir
- pump
- hoses
- those "fittings", but what kind, how many, etc...
- radiator
- coolant liquid
- anything else?

Now obviously, there is crapton of option for every one of these things, and i have no clue what to get. Is there anywhere on web some kind of guide?

 

[Leeea]

It costs as much as that Evga 2K one.

There is a good reason for that.

Generating 2000w from 240v is much easier then generating 1600w from 120v.


Not just a little easier, but massively easier. In math, the 120v unit would need to pull down 59.96% more amps down the line. But once we factor in voltage drop across your household wiring, that number is going to probably climb closer to 65% more amps down the line. That is 65% more heat to dump out of the unit, 65% more heat to dump out of your household* wiring, 65% more waste, 65% better components, etc.

*if you do go with the 120v option, you should pay a certified electrician to install an outlet just for your computer with a dedicated 20 amp circuit.


edited multiple times:
I did my percentage math wrong originally, it is as follows:
1600 watts / 120v = 13.33 amps
2000 watts / 240v = 8.33 amps
13.333 / 8.333 = 1.5996, or 59.96% more amps