ada lovelace NVIDIA's RTX 4070 will allegedly boost to 2.8GHz offering the same compute power as an RTX 3090 Ti

[BiG StroOnZ]

Summary

The serial Twitter leaker @kopite7kimi provides an update on core clocks and other major spec updates for the RTX 4070 SKU, as the RTX 40 series goes through the last stages of development. The CUDA core spec increased to 7,680 and memory to 12GB capacity. Memory speed is not 18Gbps, but now 21Gbps. Additionally, now there's a maximum boost clock of 2800MHz via GPU Boost 4.0. Lastly, TGP also lowered marginally. 

 

 

Quotes

Quote

Core frequencies are now estimated to be well above anything NVIDIA has ever done before, as the RTX 4070 reportedly has a base clock of 2310MHz and a boost of 2610MHz. Since NVIDIA’s official boost clock is always lower than real-world clock, Kopite also confirms that the maximum boost should exceed 2.8GHz. These core clocks are significantly faster than anything we've seen from NVIDIA before, and is very reminiscent of the frequency jump we saw with the Pascal-based GTX 1000 series in comparison to the 900 series Maxwell architecture. Back then, NVIDIA delivered about a 500MHz core clock improvement across the board, but the past three generations haven't moved the needle much. That looks ready to change with Ada and TSMC 5nm.

 

The memory is no longer GDDR6 memory, but now GDDR6X spec. Higher capacity and higher memory means a significant increase in bandwidth compared to old specs, which is now at 504GB/s.

 

The RTX 3070 had 700-900 MHz lower clocks. The RTX 4070's rumored clock speeds put it well ahead of the previous generation RTX 3070, with 50% higher boost frequencies. Combine this with the new L2 cache and SM improvements rumored to be on the Ada architecture, and we could be looking at an 80% to 100% performance improvement over the RTX 3070 in some workloads (Higher clocks in combination with more cores will enable 40 TFLOPS of compute power [based on official boost clock figure]. That’s 2x the compute power of RTX 3070, or alternatively the same compute power as RTX 3090 Ti [40 TFLOPS]).

 

The RTX 4070 appears to be designed to offer half the performance and half the memory of the RTX 4090 SKU. The only thing not scaling the same way is the power consumption. The RTX 4070 has a rumored total board of 285W, an increase by 45W over RTX 3070. Do note that’s not TDP, but a board limit.

 

While this isn't the official word from NVIDIA, Kopite has a good track record and the specificity of the numbers suggests inside information, perhaps from one of the graphics card vendors.

 

My thoughts

These updated specs look better because it seems they were able to lower the power slightly (5%) compared to previous rumors that had it at 300W. Previous rumors also had the CUDA Core count at 7,168 now it's at 7,680 so a 7% increase there. We went from 10GB of memory to 12GB. 160-bit Memory Bus to 192-bit. 18Gbps Memory Speed to 21Gbps. 360GB/s Bandwidth to 504GB/s, a 40% increase here. Finally, it seems the 3DMark Time Spy Extreme scores jumped up a 1,000 points since late July (from ~10,000 to ~11,000). The 4000 series is expected to launch around October, with the 4070 after the 4080 and 4090. Therefore, we should start getting actual screenshots and possible database entries of some of these claims soon, like we see with many of the Raptor Lake leaks. 

 

Sources

https://videocardz.com/newz/nvidia-geforce-rtx-4070-gpu-rumored-to-boost-over-2-8-ghz-offering-the-same-compute-power-as-rtx-3090-ti

https://www.tomshardware.com/news/nvidia-rtx-4070-allegedly-boosts-to-2800-mhz

[porina]

Random thinking: the 3070 is near enough same gaming perf as 2080 Ti, so there is a kind of precedent for this. 2070 is weaker than a 1080Ti thought. 1070 is about same as 980 Ti. I don't know about older than that. 2 out of 3, not bad, and if this claim is true than that makes it 3 out of 4.

 

I'm still disappointed I didn't get a 3080 within a year of launch. The claimed 4070 would be good enough for me, not needing 4080 or higher. Let's see what pricing and availability is like.

 

I have to wonder if the alleged increase in core count is in part due to the silicon surplus nvidia apparently have. Using a bit more to go wider would allow better efficiency per perf.

[Levent]

I wish Nvidia would stop releasing cards with bare minimum size of VRAM.

[DANK_AS_gay]

4 minutes ago, Levent said:

I wish Nvidia would stop releasing cards with bare minimum size of VRAM.

With that much power, it could do 8K gaming with 24GB of VRAM. Or more. I just don't see anyone getting an 8K display anytime soon. Although 12GB is not enough for the grunt that thing has. if they put 16GB or 20GB I think it would be fine, but like you said, 12GB is no enough. It is pretty imbalanced.

[TVwazhere]

6 minutes ago, porina said:

Random thinking: the 3070 is near enough same gaming perf as 2080 Ti, so there is a kind of precedent for this. 2070 is weaker than a 1080Ti thought. 1070 is about same as 980 Ti. I don't know about older than that. 2 out of 3, not bad, and if this claim is true than that makes it 3 out of 4.

I do believe the 780ti and 970 were close but the 780TI was still stronger (doing some quick googling seems to concur). There was extra emphasis of the 10 series at launch on "how great the 1070 is compared to the 980TI" (i'm trying to find the round-room of youtubers discussing it [notably Jayz2cents was the one that started it] but I cant seem to)

 

So to say the 4070 = 3090ti sounds ludicrous, but generationally is becoming par for the course. So this is less likely a "rumor/leak" and more of a "prediction"

[TetraSky]

Now I'm expecting the card to cost $1k CAD at the very least.

[Levent]

5 minutes ago, DANK_AS_gay said:

With that much power, it could do 8K gaming with 24GB of VRAM. Or more. I just don't see anyone getting an 8K display anytime soon. Although 12GB is not enough for the grunt that thing has. if they put 16GB or 20GB I think it would be fine, but like you said, 12GB is no enough. It is pretty imbalanced.

Its not just about resolution, games tend to have higher quality assets even at 1080p you are looking at upwards of 8G vram usage if you have just 4x MSAA enabled. I would say 16G should have been the minimum that card.

[porina]

10 minutes ago, TVwazhere said:

So to say the 4070 = 3090ti sounds ludicrous, but generationally is becoming par for the course. So this is less likely a "rumor/leak" and more of a "prediction"

Oops, I overlooked the 90 vs 80 part, but once you go to 3080 tier and up the differences between them aren't that great. And I'm looking at ball park performance anyway. Close enough you wouldn't care in practice. We're looking at ball park 10% difference between 3080 Ti and 3090 Ti?

 

That leaker is well known, and I'd give them more benefit of the doubt than the notorious rumour sites.

[Zodiark1593]

17 minutes ago, Levent said:

Its not just about resolution, games tend to have higher quality assets even at 1080p you are looking at upwards of 8G vram usage if you have just 4x MSAA enabled. I would say 16G should have been the minimum that card.

Yeah, framebuffer size alone doesn’t contribute to a lot of usage. A 4K framebuffer (non-hdr) is only 24 MB per frame. 3x that for triple buffering, and you’re still under 100 MB.  Even 8K, you’re under 500 MB for your framebuffer including triple buffering. 

Resolution matters a lot for bandwidth heavy effects, and makes overdraw a hell of a lot more expensive, but a high resolution output does not require a lot of memory itself. 

 

The assets are what matters. 

[Sir Beregond]

Rumor specs keep changing seemingly daily. I'd just wait and see what they actually end up being before making conclusions.

[Brooksie359]

1 hour ago, TVwazhere said:

I do believe the 780ti and 970 were close but the 780TI was still stronger (doing some quick googling seems to concur). There was extra emphasis of the 10 series at launch on "how great the 1070 is compared to the 980TI" (i'm trying to find the round-room of youtubers discussing it [notably Jayz2cents was the one that started it] but I cant seem to)

 

So to say the 4070 = 3090ti sounds ludicrous, but generationally is becoming par for the course. So this is less likely a "rumor/leak" and more of a "prediction"

To be fair I think the 970 was faster than the 780ti sometime after launch. 

[WereCat]

1 hour ago, Levent said:

Its not just about resolution, games tend to have higher quality assets even at 1080p you are looking at upwards of 8G vram usage if you have just 4x MSAA enabled. I would say 16G should have been the minimum that card.

You see... you want 4080 not 4070  

[Sir Beregond]

1 hour ago, Brooksie359 said:

To be fair I think the 970 was faster than the 780ti sometime after launch. 

Would make sense over time. Kepler really died an unceremonious death. Really did not last long term.

[tim0901]

24 minutes ago, Sir Beregond said:

Would make sense over time. Kepler really died an unceremonious death. Really did not last long term.

Yeah... that 3GB of VRAM at the high end just wasn't enough.

[Kisai]

3 hours ago, BiG StroOnZ said:

Summary

 

 

My thoughts

These updated specs look better because it seems they were able to lower the power slightly (5%) compared to previous rumors that had it at 300W. Previous rumors also had the CUDA Core count at 7,168 now it's at 7,680 so a 7% increase there. We went from 10GB of memory to 12GB. 160-bit Memory Bus to 192-bit. 18Gbps Memory Speed to 21Gbps. 360GB/s Bandwidth to 504GB/s, a 40% increase here.

I think at this point in time, nobody really needs anything over a 4070 in any capacity. I'm sure there may be a few games that might benefit if the RTX features are all turned on (ef FFXV still absolutely crushes 30-series cards when everything is on), but those features aren't selling points when few games use more than one of them. The parts are likely geared for workstation/server use.

 

That said, I do wish the 70 GPU's would all have 16GB standard, be it for 4K/8K video or VR. The 30-series felt like a bit of a joke when the 50's 60's (other than the TI which had 12) and 70's had 8GB, 80 had 10, and then 90 had 24. Like hello, the 80 should have been a 16GB card.

 

Nvidia might also be trying to head offer another landrush on GPU's in the event another crypto coin becomes popular shortly after launch. Like they might release TI models later with double the video memory and higher-yield on the CUDA cores.

 

At any rate, I feel that unless you're waiting for an AV1 encoder, buy the 30-series cards while they're being dumped.

 

[tim0901]

37 minutes ago, Kisai said:

At any rate, I feel that unless you're waiting for an AV1 encoder, buy the 30-series cards while they're being dumped.

It'll certainly be interesting to see where pricing lands if these performance claims turn out to be accurate. I imagine the extra VRAM will keep the 3090Ti priced higher than the 4070, but it'll be very tempting.

 

Of course, there are likely other improvements that aren't obvious here. Next-gen RT cores for example - how low can they reduce the overhead?

[Brooksie359]

1 hour ago, Sir Beregond said:

Would make sense over time. Kepler really died an unceremonious death. Really did not last long term.

Well I think a big issue was that Kepler was just quite a bit different than maxwell architecturally so it really didn't get great driver support and as a result didn't age well. 

[Supreme Calamitas]

4090 = 12k Gaming 

[Sir Beregond]

17 hours ago, Brooksie359 said:

Well I think a big issue was that Kepler was just quite a bit different than maxwell architecturally so it really didn't get great driver support and as a result didn't age well. 

This too.

[Senzelian]

Cool.

Let's hope the 7700XT keeps up and is 10% cheaper.

[NunoLava1998]

I'm afraid of the 4090 Ti now

[Middcore]

Well that's nice but if raw compute power was all it took to make a good GPU the Vega cards would have ruled their generation.

[Sorenson]

On 8/10/2022 at 10:59 AM, Levent said:

Its not just about resolution, games tend to have higher quality assets even at 1080p you are looking at upwards of 8G vram usage if you have just 4x MSAA enabled. I would say 16G should have been the minimum that card.

You need to make sure you’re looking at actual VRAM usage and not allocated VRAM usage. A lot of games will allocate around 8+ GB of VRAM but only use 3-6GB in reality. However, some games like Horizion Zero Dawn will actually use 12+ GBs of VRAM. I haven’t had an issue playing 4K with my 3070, so getting an extra 4GBs at a similar price would be a godsend. 

[Levent]

17 minutes ago, Sorenson said:

You need to make sure you’re looking at actual VRAM usage and not allocated VRAM usage. A lot of games will allocate around 8+ GB of VRAM but only use 3-6GB in reality. However, some games like Horizion Zero Dawn will actually use 12+ GBs of VRAM. I haven’t had an issue playing 4K with my 3070, so getting an extra 4GBs at a similar price would be a godsend. 

I am monitoring usage. Not to mention allocation vs usage argument doesnt matter. If you dont have enough VRAM, RAM is allocated (DVMT) and assets are swapped to ram when VRAM runs out. I can easily run out of VRAM in FH5 using extreme preset at 1080p (even 12G doesnt cut it for that preset) and have visual artifacts and stutters purely due to running out of vram and having to stream assets in and out of RAM.

 

Also, Afterburner/RTSS monitors VRAM usage, not allocation.

[Sorenson]

32 minutes ago, Levent said:

I am monitoring usage. Not to mention allocation vs usage argument doesnt matter. If you dont have enough VRAM, RAM is allocated (DVMT) and assets are swapped to ram when VRAM runs out. I can easily run out of VRAM in FH5 using extreme preset at 1080p (even 12G doesnt cut it for that preset) and have visual artifacts and stutters purely due to running out of vram and having to stream assets in and out of RAM.

 

Also, Afterburner/RTSS monitors VRAM usage, not allocation.

I'm not sure how your computer is using that much VRAM when every report I've seen has it using ~8.5GB of VRAM at 4k Ultra settings. I know the game has pop-in issues but I don't see any artifacts running it on a G9 with a 3070.