NVIDIA GeForce RTX 4070: up to 30% faster than RTX 3090 in gaming

The PCIe version of the H100 is listed at only 350W. That's with 80GB of HBM3 memory. Nvidia's limited benchmark specs show that it runs rings around the A100 80GB which is listed at 300W.

Sorry to disappoint all of the AMD fanbois, but it looks like MLID's 600+W prediction is going to be hilariously wrong - as usual.

It still remains to be seen just how much of this performance jump will translate into fps from the Geforce line, but this is a huge leap in number crunching performance and it comes with only a 16% increase in power consumption (and 200W less than AMD's MI250x):
Benchmark​
H100 PCIe​
A100 80GB PCIe​
FP64249.7+147%
FP64 Tensor Core4819.5+146%
FP324819.5+146%
TF32400156+156%
FP16 Tensor Core800312+156%
INT8 Tensor Core1600624+156%
 
Sorry to disappoint all of the AMD fanbois, but it looks like MLID's 600+W prediction is going to be hilariously wrong - as usual.
You could look at the OAM variants of either teams accelerator cards and claim 600+W is coming. I don’t see it outside of the top tier money is no object <insert sponsored-name here> Spooge edition cards from the big AIB’s.
I mean those kinds of voltages can be expected on the extreme 3090TI variants but that is a major binning process on some custom PCB’s with exotic cooling. Not something remotely suitable for a mainstream product launch.

I’m thinking a more practical 450w on the upper end of the consumer products. Beyond that and you are going to run into some walls because of how PC cases work and how much heat you can realistically dissipate in that format.

I mean they could decide they AiO tripple radiators for GPU’s are the new norm but cases are going to need a redesign to match.
 
You could look at the OAM variants of either teams accelerator cards and claim 600+W is coming. I don’t see it outside of the top tier money is no object <insert sponsored-name here> Spooge edition cards from the big AIB’s.
I mean those kinds of voltages can be expected on the extreme 3090TI variants but that is a major binning process on some custom PCB’s with exotic cooling. Not something remotely suitable for a mainstream product launch.

I’m thinking a more practical 450w on the upper end of the consumer products. Beyond that and you are going to run into some walls because of how PC cases work and how much heat you can realistically dissipate in that format.

I mean they could decide they AiO tripple radiators for GPU’s are the new norm but cases are going to need a redesign to match.
The trend is smaller cases, not bigger, so I agree with you, there are practical limits to board power, especially with power prices going steadily up.
 
The PCIe version of the H100 is listed at only 350W. That's with 80GB of HBM3 memory. Nvidia's limited benchmark specs show that it runs rings around the A100 80GB which is listed at 300W.

Sorry to disappoint all of the AMD fanbois, but it looks like MLID's 600+W prediction is going to be hilariously wrong - as usual.





H100 is Hopper (probably RTX 5XXX), not Lovelace (RTX 4XXX). "Interestingly, the SXM5 variant features a very large TDP of 700 Watts, while the PCIe card is limited to 350 Watts."

As for Lovelace, looks like it will go to 600W. AMD will probably be in the same neighborhood for RX 7XXX.
 
H100 is Hopper (probably RTX 5XXX), not Lovelace (RTX 4XXX). "Interestingly, the SXM5 variant features a very large TDP of 700 Watts, while the PCIe card is limited to 350 Watts."

As for Lovelace, looks like it will go to 600W. AMD will probably be in the same neighborhood for RX 7XXX.
The SXM5 format like the OAM format allows for up to 700w while operating at 48v.
350w is the practical limit on 12v for datacenter use and even with all their massive cooling systems are forced more than not to use liquids on any components exceeding 450w.

The recent NVidia leak made a number of mentions about Lovelace being too silicon intensive and the current supply chains not being capable of providing enough materials to make it feasible. There were reports in there about how their margins on it would be unsustainable when competing with AMD and their RDNA3 cards so they were advising benching Lovelace for their MCM based Hopper.
 
The SXM5 format like the OAM format allows for up to 700w while operating at 48v.
350w is the practical limit on 12v for datacenter use and even with all their massive cooling systems are forced more than not to use liquids on any components exceeding 450w.

The recent NVidia leak made a number of mentions about Lovelace being too silicon intensive and the current supply chains not being capable of providing enough materials to make it feasible. There were reports in there about how their margins on it would be unsustainable when competing with AMD and their RDNA3 cards so they were advising benching Lovelace for their MCM based Hopper.
Interesting. Do you have a link to the Lovelace benching issue?
 
Interesting. Do you have a link to the Lovelace benching issue?
Not a good source in any way but this is what I can find of it. Leaks and leakers are pretty bad sources in general so I don’t exactly believe it, but their Hopper based workstation cards have a release date and their Lovelace ones don’t.

Granted this could only be true within the context of the datacenter cards and not the gaming/workstation cards.

https://www.tweaktown.com/news/8075...ead-of-ada-lovelace-to-beat-rdna-3/index.html
 
Not a good source in any way but this is what I can find of it. Leaks and leakers are pretty bad sources in general so I don’t exactly believe it, but their Hopper based workstation cards have a release date and their Lovelace ones don’t.

Granted this could only be true within the context of the datacenter cards and not the gaming/workstation cards.

https://www.tweaktown.com/news/8075...ead-of-ada-lovelace-to-beat-rdna-3/index.html
Thanks.

I wouldn't mind seeing Hopper come out instead of Lovelace. I always figure Hopper was going to be the true powerhouse NV wanted to unleash on the world. This fall/winter is going to be fun.
 
Chiplets, so hot right now
 

Attachments

  • 869ed4decfec61cb-hwTcxYgD-zoom.png
    869ed4decfec61cb-hwTcxYgD-zoom.png
    374.3 KB · Views: 0
You could look at the OAM variants of either teams accelerator cards and claim 600+W is coming.
MLID has been very explicit that he's talking about the Geforce lineup, not the OEM-only SXM form-factor enterprise cards.
 
Last edited:
MLID has been very explicit that he's talking about the Geforce lineup, not the OEM-only OAM form-factor enterprise cards.
Good to know, but I don’t know what they are basing their analysis off of, I can’t say their wrong god know the leaked stuff for the 3090TI’s show them to be thirsty bitches, but I’m not sure we can call that the new norm or a top end outlier. I mean if it’s the new norm then I’m going to have to move because my office isn’t equipped to handle those kinds of thermals.
 
Good to know, but I don’t know what they are basing their analysis off of, I can’t say their wrong god know the leaked stuff for the 3090TI’s show them to be thirsty bitches, but I’m not sure we can call that the new norm or a top end outlier. I mean if it’s the new norm then I’m going to have to move because my office isn’t equipped to handle those kinds of thermals.
The 600W is a simple straight line scaling of watts per core using the rumored higher core count. It ignores the node shrink and the fact that the VRAM is a significant TDP contributor.
 
The 600W is a simple straight line scaling of watts per core using the rumored higher core count. It ignores the node shrink and the fact that the VRAM is a significant TDP contributor.
Right…. So since moving a off the table, I’m not house hunting in this market I better look into a heat pump before my next build. Because if you take 600w account for a node shrink, add more VRAM, then increase clock speeds I still come out with a respectable 500w +… so I guess my only hope is the 4060 / 6700XT series chips aren’t so hungry.

It’s not like I get any chance to game anymore anyways and when I do the game is like 3 years old. Just getting around to Witcher 3…. Wooooh!
 
Right…. So since moving a off the table, I’m not house hunting in this market I better look into a heat pump before my next build. Because if you take 600w account for a node shrink, add more VRAM, then increase clock speeds I still come out with a respectable 500w +… so I guess my only hope is the 4060 / 6700XT series chips aren’t so hungry.

It’s not like I get any chance to game anymore anyways and when I do the game is like 3 years old. Just getting around to Witcher 3…. Wooooh!
The VRAM is expected to be the same as the 3090, where it pulls an estimated 60-100W. I'd expect the FE to be 400W while something like a Strix 4090 to be 450W.

You may want to bookmark this post so you can come back to it and mock me in a few months. :)
 
The VRAM is expected to be the same as the 3090, where it pulls an estimated 60-100W. I'd expect the FE to be 400W while something like a Strix 4090 to be 450W.

You may want to bookmark this post so you can come back to it and mock me in a few months. :)
450w would be “reasonable” for a high end card….. but not 600. A 120mm aluminum radiator is good for about 150w of thermal dissipation before it saturates, so 450w could be handled by an AiO running a 360mm radiator, 600w would require 4 and I don’t know how many cases that would fit unless they split it up. But anything past 450w is incredibly hard to cool on air if you value your eardrums and don’t have the room climate controlled.
 
Every year the 4030 will be 800% faster than the 3090 Titan founders edition and it never happens according to early clickbait leaks.
Highly doubt they'll push Board power that high. Most average Pcs can barely properly run a vega or 3k/6k series transient peak.
 
Competition is great! One reason Pascal topped out at 250W (1080 Ti) is that Nvidia didn't have to push the envelope. AMD had nothing to challenge the high end.
No! Nvidia was scared of Vega so went near full out, needlesly. 250W was considered eschelon in the day noobs. The market got a bonus because of it. Expect similar this next round.
 
I’m always shopping in the at 300w or less range power draw for GPUs. Personally I like 250w or less since getting into SFF building and finding that mid to upper mid range performance is fine by me. If that is unplayable in tomorrow’s games then I’ll just do something else.
 
The VRAM is expected to be the same as the 3090, where it pulls an estimated 60-100W. I'd expect the FE to be 400W while something like a Strix 4090 to be 450W.

You may want to bookmark this post so you can come back to it and mock me in a few months. :)
My 3090 Pulls ~49W on the ram, 59W max I've recorded when overclocked.
1648792479115.png
 
Just my opinion, I'm thinking the rumors might be true though, that's a lot of performance to be kicking out at that core count and what not. Also my 3090 Strix draws an insane amount of power.

Image below. This is at full load though, mining atm, while also running Topaz Labs Video Enhance AI. I normally run a lower power profile, but set it to stock for this image with a +900 bump on memory. I had a Founders Edition before my strix, definitely drew a lot less power compared to the strix. In a gaming environment though, no joke, my strix pulls just above 450w. Not sure why it's lower when I'm running a mining session + video enhance AI atm.

So considering the bump in core count, giant monolithic die, I'm expecting easily to creep closer to that 600w mark. Time will tell though. Gonna swap out to a 4090 when it hits, although I'm more curious about how AMD's upcoming RDNA 3 will perform. Wouldn't get one, as I need cuda for my workload, but looking forward to the reviews on that more than anything.

2022-03-31 23_06_44-NVIDIA GeForce RTX 4070_ up to 30% faster than RTX 3090 in gaming _ Page 5...png
 
My 3090 Pulls ~49W on the ram, 59W max I've recorded when overclocked.

I don't have a screencap from a 3090 handy, but this is from an A6000. 48GB instead of 24GB, but also GDDR6 instead of the more (iirc) power-hungry GDDR6X of the 3090. If we cut this in half, it's still over 100W:
1648819850895.png


Don't forget that in addition to the 4000 series upping the core count in a big way, there is also a major node shrink that goes along with it which will make each core more power efficient.

MLID's 600W lie assumes there is no node shrink and neither the 3090 nor the 4090 have any VRAM on them.

It's fun to make guesses at the future, but keep in mind the rumor source here. The very first time MLID did a video mentioning GPUs, it's a 20 minute long rant about how it literally is Nvidia's fault that AMD's new-at-the-time GPU sucked so hard. MLID is so far up AMD's butthole that he could brush Lisa's teeth while her mouth is closed.
 
My 3090 Pulls ~49W on the ram, 59W max I've recorded when overclocked.

don't have a screencap from a 3090 handy, but this is from an A6000. 48GB instead of 24GB, but also GDDR6 instead of the more (iirc) power-hungry GDDR6X of the 3090. If we cut this in half, it's still over 100W:
The difference between an A series and 3090 is the A series is going to be using all that memory, on a 3090 most of it is sitting in an idle state. I’m sure the 3090 could consume more than that with the memory usage at full load.
Note: GDDR6x uses about 15% less power per transferred bit, but as it moves more bits over a given time it consumes more overall under load.

Well see what next gen brings when it gets here, not a lot of good press coming from Samsung and their processes. So we should all be thankful Nvidia is going back to TSMC different leaks from Samsung and Qualcomm each show the yield rates from Samsung 5nm to be as low as 35%. Qualcomm is pulling the production of the Snapdragon 8 from Samsung as a result.

Samsung 8 also proved to be far inferior to TSMC 7, so moving to TSMC 5 will bring large improvements on its own.
 
Last edited:
Jarrod's Tech showed that the differences between a 2060->2070->2080 and 3060->3070-3080 are marginal at best in laptops, likely due to power and thermal limitations in a small form factor. With the specs of the 4000 series being so high, assuming the mobile chips will have the same proportional increase in power from the last gen, what's the point in getting a high end gpu in a laptop without a screaming loud cooling solution? Even if they cost the same, I would rather trade 10% performance for more than 10% battery and more than 10% less fan noise.
 
Isn't 384.1W the overall board draw? Or is that a reply to something I missed?
 
Isn't 384.1W the overall board draw? Or is that a reply to something I missed?
They were commenting earlier about memory draw on an RTX 3090 vs an A4000. Somebody commented that the 3090 drew far less and why was an A4000 drawing 4x more while only having double the ram. But their 3090 test was flawed because it was only using a small fraction of its memory while the above mining test is using all of it and showing its memory usage properly.
 
They were commenting earlier about memory draw on an RTX 3090 vs an A4000. Somebody commented that the 3090 drew far less and why was an A4000 drawing 4x more while only having double the ram. But their 3090 test was flawed because it was only using a small fraction of its memory while the above mining test is using all of it and showing its memory usage properly.

Very close. My screenshot of the 3090's VRAM pulling 100W is just while playing PUBG at stock clocks.

So, what we see is ~380W total board draw. Take 100W off of that and we have 280W left for everything else. For simplicity, let's count 100% of that as GPU core usage. If a 4090 has 50% more cores, then its gpu could pull 420W plus 100W for the VRAM, thus making the 4090 a 520W card. That is far lower than MLID's 600W claim.

Of course, this accounts for the improved power efficiency from neither the node shrink nor from the generally more efficient architecture. What improvement is expected there? 25% brings the full board TDP down to 415W. If only 15%, then it's 457W. Plus, this is from a premium factory-OCed AIB card like a Strix.
 
My main thing is, do you guys think this will require PCIE 5.0 to use at all? Or will PCIE4 suffice? I just did an upgrade so that's kind of rough, supposing I do even get the opportunity to try to upgrade...
 
My main thing is, do you guys think this will require PCIE 5.0 to use at all? Or will PCIE4 suffice? I just did an upgrade so that's kind of rough, supposing I do even get the opportunity to try to upgrade...
PCI-E 3.0 is still fine. Why would you think PCI-E 4.0 wouldn't be enough?
 
My main thing is, do you guys think this will require PCIE 5.0 to use at all? Or will PCIE4 suffice? I just did an upgrade so that's kind of rough, supposing I do even get the opportunity to try to upgrade...
PCIE4 will be fine for a while, you have to get into some pretty specific workloads to make use of 5 right now.

DirectStorage will see slight improvements from 3 to 4 but you won’t see any gaming benefits from moving to 5 until 3 becomes a rarity.
 
PCI-E 3.0 is still fine. Why would you think PCI-E 4.0 wouldn't be enough?
I just remember the GamerNexus overview for these new GPU's mentioning PCIE 5.0. I figured it was mainly for high end compute loads or something, because iirc most of the time the latest PCIE specs remained well ahead of GPU bandwidth requirements, but I wasn't sure. There was some mention of having more power from PCIE on the slot, though, for the 5.0 spec... had me worried.
PCIE4 will be fine for a while, you have to get into some pretty specific workloads to make use of 5 right now.

Thanks, I sincerely assume and hope that's the case.
 
They were commenting earlier about memory draw on an RTX 3090 vs an A4000. Somebody commented that the 3090 drew far less and why was an A4000 drawing 4x more while only having double the ram. But their 3090 test was flawed because it was only using a small fraction of its memory while the above mining test is using all of it and showing its memory usage properly.
My test was using a portion of the ram, but it is based on real world gaming. I've recently played both Cyberpunk 2077 for say a 5 hour stretch, and Siberian Mayhem for the same stretch, and never saw the max recorded value on my FE hit 60W. About 13Gb of the vram is max ram usage I have seen, and even with data in the vram, it isn't always all being accessed at all times.

So unless MLID was talking about Mining use cases, the power draw is still overstated imho.
A Strix hitting 100watts playing pubg... maybe it's just the games I am playing.

I have flashed the 3090 to the resize bar bios, and its enabled in my mobo bios. Maybe that improves vram power usage, hell if I know. But no matter the game I have played, I have never seen 60W power usage on my vram. And it's OC'd +750
1649350445855.png
 
Mining hammers VRAM, even the most demanding games won't come close to what mining does.
 
My test was using a portion of the ram, but it is based on real world gaming. I've recently played both Cyberpunk 2077 for say a 5 hour stretch, and Siberian Mayhem for the same stretch, and never saw the max recorded value on my FE hit 60W. About 13Gb of the vram is max ram usage I have seen, and even with data in the vram, it isn't always all being accessed at all times.

So unless MLID was talking about Mining use cases, the power draw is still overstated imho.
A Strix hitting 100watts playing pubg... maybe it's just the games I am playing.

I have flashed the 3090 to the resize bar bios, and its enabled in my mobo bios. Maybe that improves vram power usage, hell if I know. But no matter the game I have played, I have never seen 60W power usage on my vram. And it's OC'd +750
View attachment 461612
The 3090 has more ram than a gaming GPU needs while gaming even at 4K your not going to be using much more than 12. Which is why in practical usage 50-60w will the the norm. The A series workloads or mining applications can tear into ram like a bag of chips and will run it constant and in that case you will see your worst case numbers.

I don’t know where MILD gets their numbers and what they speculate on to land where they do.

But I’m pretty sure most PUBG cheats balloon out GPU memory usage so…. I can see a few of those making use of that memory.
 
Last edited:
The 3090 has more ram than a gaming GPU needs while gaming even at 4K your not going to be using much more than 12. Which is why in practical usage 50-60w will the the norm. The A series workloads or mining applications can tear into ram like a bag of chips and will run it constant and in that case you will see your worst case numbers.

I don’t know where MILD gets their numbers and what they speculate on to land where they do.

But I’m pretty sure most PUBG cheats balloon out GPU memory usage so…. I can see a few of those making use of that memory.
I've used up to about 14-16gb in a handful of pc games with everything maxed out.

Only time I've hit 23gb+ was in Yuzu playing xenoblade 1 with serious memory leaks. If you want to know what happens, it's as you would imagine. Framerate starts to tank, eventually either the system ram crashes the game, or the gpu memory crashes the game. Usually only when later in the game, after long playing sessions traveling to many different locations.

Still, you can imagine my surprise when I saw afterburner showing over 20 gigs of gpu memory in the corner.
 
I have a 1000W PSU but my UPS is 900W max. Take out the 3 monitors and network gear hooked up to it and I have about 750W left for the computer.

12900K system without video card maxes out at 300W, so 450W is left for the video card. Hopefully 4080 is under that.
 
You don't have to run it balls to the wall on battery power though, Spicedaddy. That being said I have a different UPS for the network closet. Every computer has it's own 900w UPS, 475w x 2 for the network stuff.
 
I guess only question is whether I wait for one of these cards, or try to find a 3080Ti for a good price instead, probably later. I think either one will be probably way more than I need at 3440x1440, and it looks like GPU prices are (surprisingly) finally dropping back down to earth (kind of)... so... yeah...

TBH considering the power targets of these new things, 4070 would probably be absolute max I would go for.

But no pressing need to upgrade, just kind of an itch since I upgraded to a 5950x recently... but my 2080 is doing fine...
 
Let me know when the '1060' variant of Hopper actually hits retail availability at MSRP.
 
Back
Top