MSI GeForce RTX 3090 Ti Suprim X

[ratirt]

As for this, we'll have to disagree on that. While this "problem" will no doubt become more noticeable in the future, at the same time its absolute compute performance (whether rasterization or RT) will simultaneously decrease relative to the demands put on it by games, meaning that by the point where this is a dominating issue (rather than an extreme niche case, like today), those GPUs likely wouldn't produce playable framerates even if they had infinite VRAM. Remember, Doom Eternal is just about the easiest-to-run AAA shooter out there in terms of its compute requirements (and it can likely run more than fine at 2160p RT on a 3070 if you lower the texture quality or some other memory-heavy setting to the second highest setting). And it's not like these two games are even remotely representative of RT loads today - heck, nothing is, given that performance for the 3090 Ti at 2160p varies from ~137fps to ~24fps. The span is too wide. So, using these two edge cases as a predictor for the future is nit-picking and statistically insignificant. So again, calling the cards "handicapped" here is ... well, you're picking out an extreme edge case and using it in a way that I think is overblown. You can't expect universal 2160p60 RT from any GPU today, so why would you do so with an upper mid-range/lower high end GPU? That just doesn't make sense. Every GPU has its limitations, and these ones clearly have their limitations most specifically in memory-intensive RT at 2160p - the most extreme use case possible. That is a really small limitation. Calling that a "handicap" is making a mountain out of a molehill.

I’m not talking about to which degree this is a problem but rather it is here already and obviously in the future it will be more noticeable. These cards are RT capable and if you pay cash for something, you don’t want to be constraint. Saying this card is 1440p doesn’t mean you must play at that resolution, and you can’t go up and crank down the details to play comfortably. Some people might see that as a huge disadvantage you know. Infinite Vram? I'm talking about bare minimum to play a game which apparently this GPUs (3070 and 3070 Ti) would have been capable of with the games I have mentioned. I really don’t understand what are you trying to prove? That he RT implementation sucks because in different games RT is more demanding? So is rasterization in those games obviously and since RT hits the performance when enabled that’s the natural cause what happens. Listen, I asked about the RAM because I was curious. Nonetheless these cards could play RT 4k with those games but can’t already. Not tomorrow but now. So, lets just leave it at that. If you are ok with it that perfectly fine.

That is a way too simplistic solution to this conundrum. As a 6900XT owner using it on a 1440p60 display, I know just how low that GPU will clock and how efficiently it will run if it doesn't need the power (that 75W figure I gave for Elden Ring isn't too exceptional). I've also run an undervolted, underclocked profile at ~2100MHz which never exceeded 190W no matter what I threw at it. The point being: RDNA2 has no problem clocking down and reducing power if needed. And, to remind you, in the game used for power testing here, the 6900XT matches the performance of the 3080Ti and 3090 at 1440p while consuming less power. Despite its higher clocks, even at peak. And, of course, all of these GPUs will reduce their clocks roughly equally, given an equal reduction in the workload. Yet what we're seemingly seeing here is a dramatic difference in said reductions, to the tune of a massive reversal of power efficiency.

So, while you're right that power consumption and performance scaling are not linear, and that a wide-and-slow GPU will generally be more efficient than a fast-and-narrow one, your application of these principles here ignores a massive variable: architectural and node differences. We know that RDNA2 on TSMC 7nm is more efficient than Ampere on Samsung 8nm, even at ~500MHz higher clocks. This is true pretty much true across the AMD-Nvidia product stacks, though with some fluctuations. And it's not like the 3090Ti is meaningfully wider than a 3090 (the increase in compute resources is tiny), and by extension not a 6900XT either. You could argue that the 3080Ti and 3090 are wider than the 6900 XT, and they certainly clock lower - but that runs counter to your argument, as they then ought to be more efficient at peak performance, not less. This tells us that AMD simply has the architecture and node advantage to clock higher yet still win out in terms of efficiency. Thus, there doesn't seem to be any reason why these GPUs wouldn't also clock down and reduce their power to similar degrees, despite their differing starting points. Now, performance scaling per frequency for any single GPU or architecutre isn't entirely linear either, but it is close to linear within the reasonable operating frequency ranges of most GPUs. Meaning that if two GPUs produce ~X performance, one at 2GHz and one at 2.5GHz, the drop in clock speeds needed to reach X/2 performance should be similar, not in MHz but in relative % to their starting frequencies. Not the same, but sufficiently similar for the difference not to matter much. And as power and clock speeds follow each other, even if non-linear, the power drop across the two GPUs should also be similar. Yet here we're seeing one GPU drop drastically more than the other - if we're comparing 3090 to 6900 XT, we're talking a 66% drop vs. a 46% drop. That's a rather dramatic difference considering that they started out at the same level of absolute performance.

Ok, simplistic, you ask a question, and you answer it yourself. These are two different architectures if you want to compare those just by the result you can. I didn’t miss anything. I don’t see how you can compare two different architectures saying one uses more or less power than the other giving the same performance or similar. Different nodes, different architectures. Obviously, that’s the case. So comparison of these two to understand the difference lies there. Different node and different architectures. There’s no point on dwelling on it. Obviously, they are different, and the difference will be there in the results. I focused on the results themselves and performance/consumption. Here is your answer. Node difference, architecture difference since these are completely different products just have the same goal.

One possible explanation: That the Ampere cards are actually really CPU limited at 1080p in CP2077, and would dramatically outperform the 6900XT there if not held back. This would require the same to not be true at 1440p, as the Ampere GPUs run at peak power there, indicating no significant bottleneck elsewhere. This would then require power measurements of the Ampere cards at 1080p without Vsync to check. Another possible explanation is that Nvidia is drastically pushing these cards beyond their efficiency sweet spot in a way AMD isn't - but given the massive clock speeds of RDNA2, that is also unlikely - both architectures seem to be pushed roughly equally (outside of the 3090 Ti, which is ridiculous in this regard). It could also just be some weird architectural quirk, where Ampere is suddenly drastically more efficient below a certain, quite low clock threshold (significantly lower than any of its GPUs clock in regular use). This would require power testing at ever-decreasing clocks to test.

Either way, these measurements are sufficiently weird to have me curious.

I remember HWUB talking about the NV driver overlay or something. A CPU is being utilized more due to lower resolution than AMD’s counterpart and thus the lower results since the resources are taken. I don’t know if that issue has been fixed by Nvidia or not. I know there was an instance brought by HWUB. Bump resolution up and you have no or way less driver overlay. Also, if they are limited at low res (obviously are) just as AMD counterparts are, maybe it is due to game itself. There was also a mention of the architecture of the Ampere GPUs. The FP32 processing mainly that some has pointed out to have an impact on the lower res high FPS performance. If I remember correctly.

I was curious about the RT and ram insufficiency and thus my question to Wiz. Any conclusions I leave for you but I did share mine which you don't need to agree with.

 

[nguyen]

I'm genuinely curious as to why this is. Looking at CP2077 performance results, sadly these cards seem CPU limited at 1080p, so there isn't much to gather there - though at 1440p the 3090Ti clearly pulls ahead of the 3080Ti, 6900XT and 3090 which are tied. Yet in this power consumption graph the Nvidia cards are closely grouped while the 6900XT is ~40% higher. That strikes me as odd, given that the 6900XT uses less power than a 3090, 3080ti, and even the 3080. I understand that the test scenarios for these measurements aren't the same, but the difference seems very strange to me. The Gaming power consumption numbers are also CP2077, though at 1440p, not 1080p - but in this scenario, the 6900XT delivers the essentially identical performance at ~40W less. So how come the situation is so dramatiaclly reversed at 1080p60? @W1zzard, got any thoughts on this?

I guess Nvidia driver allows for memory clocks to drop when high VRAM clock is unnecessary
Here I try to downclock the core, which also drop VRAM clock to 10Gbps compare to 19.5Gbps stock

I think the worst offender to Ampere's efficiency is the GDDR6X and not the Samsung 8N fab. These 21Gbps GDDR6X on 3090Ti use 100-110W in-game while the GDDR6 on my 2080Ti only use around 30-40W (probably the same for 6900XT)
GPU-Z screenshot from Tweaktown review

 

[Fluffmeister]

60% slower? You're looking at a 3050 as an upgrade to a 980 Ti?

The 3070 based on the 4K graph. It's 63% slower.

 

[Valantar]

@Valantar
I don't see that power consumption so complicated. Each GPU simply have some predefined performance states at some frequencies and it depends on current GPU core percentage usage when the GPU decide to change that state to other one. Both amd and nvidia have different percentage/frequency/... ranges for different performance states count, ...so it is very hard to tell at default conditions if the card could or could not be more power efficient.

I would advise take some static camera angle in game at 1080p60fps, set manually some lower performance state, start manually locking GPU frequencies to lower values and stop this process when gpu reaches almost 100% core utilization. Then do undervolting and after that check what is the power consumption. Without this procedure we are all totaly just guessing.

Could you post what are the frequencies (core, vram) of each performance state on your 6900XT?

Current GPUs don't have anything practically understandable as "predetermined performance states", they have voltage/frequency curves with heaps and heaps of points, and they scale their clock speeds dynamically based on load, thermals, voltages, and a bunch of other factors - and they will clock quite low if the load is low. This is precisely why I'm curious as to these results: from the widely documented characteristics of Ampere and RDNA2 as well as my own experiences, I don't quite understand why Ampere cards would drop their power consumption so much more drastically than equally performing RDNA2 cards in the same scenario. This indicates an unevenness of scaling that seems outside the bounds of normal variations to me. Hence my curiosity.

I guess Nvidia driver allows for memory clocks to drop when high VRAM clock is unnecessary
Here I try to downclock the core, which also drop VRAM clock to 10Gbps compare to 19.5Gbps stock

View attachment 241805

I think the worst offender to Ampere's efficiency is the GDDR6X and not the Samsung 8N fab. These 21Gbps GDDR6X on 3090Ti use 100-110W in-game while the GDDR6 on my 2080Ti only use around 30-40W (probably the same for 6900XT)
GPU-Z screenshot from Tweaktown review
View attachment 241806

That's really interesting! An intermediate memory speed state could indeed go some way towards explaining that. Also, holy crap, 107W on the memory rail? That's crazy. Though it is 24GB of the stuff, it's still a lot of power. Memory speed scaling is also something where AMD has a long history of struggling - there have been widely reported issues across many different GPUs where memory speed for various reasons gets stuck at peak clocks even on the desktop, leading to very high idle power. It would stand to reason that this could cause behaviour like this. Would love to have someone look into this in that specific test scenario.

I’m not talking about to which degree this is a problem but rather it is here already and obviously in the future it will be more noticeable. These cards are RT capable and if you pay cash for something, you don’t want to be constraint. Saying this card is 1440p doesn’t mean you must play at that resolution, and you can’t go up and crank down the details to play comfortably. Some people might see that as a huge disadvantage you know.

No, but I never said that either. But using a "1440p card" to play at 2160p with lowered details also handily avoids these VRAM issues, rendering that point moot. What we're looking at here is at Ultra, after all.

Infinite Vram? I'm talking about bare minimum to play a game which apparently this GPUs (3070 and 3070 Ti) would have been capable of with the games I have mentioned. I really don’t understand what are you trying to prove?

That what you're describing is a very niche issue, and one likely quite easily overcome by - as you yourself brought up above - lowering settings a bit.

That he RT implementation sucks because in different games RT is more demanding?

Nope, I just said that RT performance varies wildly across titles and that extrapolating from a couple of worst-case scenarios is thus a poor predictor of future performance.

Ok, simplistic, you ask a question, and you answer it yourself.

Have I answered anything? Not as far as I've seen, at least. I've just explained why your simplistic "scaling isn't linear" statement is insufficient to explain the dramatic differences in scaling between these GPUs.

These are two different architectures if you want to compare those just by the result you can. I didn’t miss anything. I don’t see how you can compare two different architectures saying one uses more or less power than the other giving the same performance or similar. Different nodes, different architectures. Obviously, that’s the case. So comparison of these two to understand the difference lies there. Different node and different architectures. There’s no point on dwelling on it. Obviously, they are different, and the difference will be there in the results. I focused on the results themselves and performance/consumption. Here is your answer. Node difference, architecture difference since these are completely different products just have the same goal.

But that's the thing: even different nodes and architectures broadly follow predictable voltage/frequency curves. They don't match or align, but the curves tend to look similar, have similar shapes. Yet the scaling here seems to indicate quite dramatic differences between overall similarly performing products in relatively similar workloads. This indicates that something more than just simple DVFS variance is happening here. Something like what @nguyen indicated above, for example.

I remember HWUB talking about the NV driver overlay or something. A CPU is being utilized more due to lower resolution than AMD’s counterpart and thus the lower results since the resources are taken. I don’t know if that issue has been fixed by Nvidia or not. I know there was an instance brought by HWUB. Bump resolution up and you have no or way less driver overlay. Also, if they are limited at low res (obviously are) just as AMD counterparts are, maybe it is due to game itself. There was also a mention of the architecture of the Ampere GPUs. The FP32 processing mainly that some has pointed out to have an impact on the lower res high FPS performance. If I remember correctly.

This is exactly why I was curious - there are tons of possible explanations here. Isn't that an interesting thing to find out? Saying "duh, power and clock speed don't scale linearly" is ... well, not that.

The 3070 based on the 4K graph. It's 63% slower.

Ah, you mean the 3090 Ti is 63% faster? Relative percentages don't work in both directions, remember. The 3070 is 39% slower than the 3090 Ti (61% of its 100%); the 3090 Ti is 63% faster than the 3070 (163% of its 100%) - it depends what number you start from, and the wording used needs to match the reference point. And comparing upwards from a smaller number always seems more dramatic than comparing downwards from a larger one (if I have $25 and you have $100, you have 400% more money than me, while I have 75% less money than you, etc.), which is why mixing up these two directions of comparison gets troublesome - it effectively overstates (or understates!) the difference. It's still absolutely a significant difference, don't get me wrong. But it's not quite that bad. Saying it's 63% slower implies that if the faster card gets 100fps, the slower will only get 37fps, when the reality is that it will get 61fps. And that's quite a difference. (And yes, I'm well aware I'm being a bit of a pedantic a** right now, sorry!)

 

[Fluffmeister]

Yes, that is indeed what I meant. I could afford to buy even this 3090 Ti with cash in the bank but I like many bulk at the current prices. Sorry I got my last wires crossed with my initial post too.

 

[Valantar]

Yes, that is indeed what I meant. I could afford to buy even this 3090 Ti with cash in the bank but I like many bulk at the current prices. Sorry I got my last wires crossed with my initial post too.

Understandable I keep catching myself doing the same type of thing (or just getting confused AF when trying to compare percentages), hence why I noticed it here I guess. And yeah, it's still a travesty even if prices are coming down. I'm rather pessimistic about future outlooks as well - given that most AIB partners have been operating on razor-thin margins for years, I can't avoid thinking they kind of like this new normal, and want to keep it that way. Which says more about the inherent structural flaws of the computer hardware industry than anything else - when Intel, Nvidia and AMD operate on ~40% gross margins while AIB partners struggle to break even on retail GPUs even above MSRP (even if that's a gross v. net margin comparison and thus not equal), that's a problem. But the solution to that problem can't be "GPUs start at $400 and go up way past $1500 now, tough luck". That will erode the entire customer base for PC gaming as a whole. Something has to give, somewhere.

 

[wheresmycar]

Relative percentages don't work in both directions, remember. The 3070 is 39% slower than the 3090 Ti (61% of its 100%); the 3090 Ti is 63% faster than the 3070 (163% of its 100%) - it depends what number you start from, and the wording used needs to match the reference point. And comparing upwards from a smaller number always seems more dramatic than comparing downwards from a larger one (if I have $25 and you have $100, you have 400% more money than me, while I have 75% less money than you, etc.), which is why mixing up these two directions of comparison gets troublesome - it effectively overstates (or understates!) the difference. It's still absolutely a significant difference, don't get me wrong. But it's not quite that bad. Saying it's 63% slower implies that if the faster card gets 100fps, the slower will only get 37fps, when the reality is that it will get 61fps. And that's quite a difference.

lol i've been a (and on-going) victim of this too. Some time back i saved a snapshot of a relative performance chart from TPU and based a bunch of decisions off it... luckily i didn't pull the trigger on anything at its behest. I tend to eventually look at actual FPS returns in specific games at a given resolution before pulling the trigger. I'm glad you pointed out the above, makes so much sense now.

(And yes, I'm well aware I'm being a bit of a pedantic a** right now, sorry!)

A useful pedantic a** lol with a happy and better informed customer here!!

 

[the54thvoid]

So, after a full day, you can still buy the FE 3090ti from Scan in the UK. £1879.

It states the power draw of the FE card is 350 Watts. Recommended PSU is 750W. Wish there was a review of the FE version. Need to go Google it. 350W isn't as bad as 450W.

 

[Prima.Vera]

2500Euros in Europe and Japan, without VAT?? Seriously, this is getting out of control. We need Intel to break those callous prices with something significant...

 

[ratirt]

I guess Nvidia driver allows for memory clocks to drop when high VRAM clock is unnecessary
Here I try to downclock the core, which also drop VRAM clock to 10Gbps compare to 19.5Gbps stock

So you are saying that, you downclock the Core frequency and somehow the Memory frequency drops as well?

 

[Bzuco]

@Valantar
There is also another reasonable explanation, CP2077 could be designed to more utilize shader units than sampling textures or doing rasterization tasks, they have custom inhouse developed realtime global illumination(non RT), which is maybe half precomputed or fully computed at realtime on cuda cores, do not know. CDPR was certainly informed, that the upcoming generation of nvidia cards will have doubled the shader counts, so why not use the potential for their game(or better said, adapt their game to new core units ratio). With the double amount of shader unit no need to use higher core frequencies at 60Hz.
Scenario for AMD would be: do less shader calculations and more texture sampling thanks to infinity cache.

Predefined performance levels in bios on asus 2060s nvidia:
P0 - 1470MHz + boost, vram max.1750MHz
P2 - max. 825MHz clock, vram max.1700MHz
P3 - max. 825MHz clock, vram max.1250MHz
P5 - max. 645MHz clock, vram max.202MHz
P8 - max. 645MHz clock, vram max.101MHz

I had always only nvidia card, and number of those P-states has always been different, same as locked/unlocked voltages/clocks for certain P-state. Too bad I can't set vram e.g. 600MHz, forcing vram frequency is somehow ignored when using official nvidia-smi command-line utility.
Those P-states can be forced, good for e.g. video playback, video encoding, forcing minimizing power consumption, etc.

Therefor I am curious if AMD cards have something similar and also what P-states are on 3090/ti cards.

 

[ratirt]

No, but I never said that either. But using a "1440p card" to play at 2160p with lowered details also handily avoids these VRAM issues, rendering that point moot. What we're looking at here is at Ultra, after all.

You have suggested that saying the card is designed to play 1440p as a counter to my can't play 4k with RT on.

But that's the thing: even different nodes and architectures broadly follow predictable voltage/frequency curves. They don't match or align, but the curves tend to look similar, have similar shapes. Yet the scaling here seems to indicate quite dramatic differences between overall similarly performing products in relatively similar workloads. This indicates that something more than just simple DVFS variance is happening here. Something like what @nguyen indicated above, for example.

Well, these are two different architectures. Maybe I don't understand exactly. I tried, like you said 'simple solution' but you mentioned architecture and node difference but you were not satisfied with it and brought up the arch and node differences.

This is exactly why I was curious - there are tons of possible explanations here. Isn't that an interesting thing to find out? Saying "duh, power and clock speed don't scale linearly" is ... well, not that.

well they dont scale linearly. There can be other issues but obviously there is a huge correlation between power and clock. If you want to eliminate the for instance driver overlay while testing. Simply, switch to 4K and the problem is gone. Then start tweaking the card and you will find out the power and clock relations for sure. There might be tons and all of them may affect the performance. As usual, these are only valid in certain scenarios. Power to clock, no matter what you do you will have that relation.

 

[nguyen]

So you are saying that, you downclock the Core frequency and somehow the Memory frequency drops as well?

The driver will decide whether the VRAM run at full speed or 10Gbps depending on graphical demand (unless the "Prefer Maximum Performance" option is On). The Max FPS option can save a lot of unnecessary power usage since the driver will downclock as appropriate.

For example running Dota2 1440p 120hz

vs 4K 120hz

VRAM power consumption goes from 69W at full speed down to 47W at 10Gbps.
So yeah when you don't need maximum performance, these Ampere GPU can be quite efficient.

 

[W1zzard]

1440p 120hz

4K 120hz

If you set your desktop to these resolutions, you're getting the same memory clocks as in dota 2?

The card thinks it's idle, and will clock down to its lowest performance state

 

[nguyen]

If you set your desktop to these resolutions, you're getting the same memory clocks as in dota 2?

The card thinks it's idle, and will clock down to its lowest performance state

this is desktop idle 4K 120hz

Dota2 4k120hz vs 1440p120hz

 

[ratirt]

The driver will decide whether the VRAM run at full speed or 10Gbps depending on graphical demand (unless the "Prefer Maximum Performance" option is On). The Max FPS option can save a lot of unnecessary power usage since the driver will downclock as appropriate.

For example running Dota2 1440p 120hz
View attachment 241895

vs 4K 120hz
View attachment 241896

VRAM power consumption goes from 69W at full speed down to 47W at 10Gbps.
So yeah when you don't need maximum performance, these Ampere GPU can be quite efficient.

Well, How I see it. If you limit FPS the driver will lock the resources to save power. The card will use necessary resources to sustain the 120FPS limit you have set. So no wonder the power drops when the frequency of the memory drops as well. It does make sense

 

[wolf]

I think the worst offender to Ampere's efficiency is the GDDR6X and not the Samsung 8N fab

Very much so, even on my 3080 I see 80+w to the memory often. When I'm undervolted pulling roughly 225w board power, the GPU is 75-100w, it's really not that bad, but the memory and circiutry make up the rest.

Honestly if people are worried about RTX 40 series being 600w monsters they have a very short memory. Halo cards from many past generations have pulled crazy power, and yeah, the halo cards next gen probably will too, but it doesn't mean a RTX 4060 will want 450w. Get your expectations in check.

GTX 295 - 2009 - 289w
GTX 590 - 2011 - 365w
R9 295X2 - 2014- 500w

Silly halo products for people with more money than sense have been creeping power consumption for bad gains for many years. The only difference now is dual GPU is dead. If you find yourself put off by the price or power consumption of the 3090Ti, this product was never going to be for you in the first place.

 

[Valantar]

Very much so, even on my 2080 I see 80+w to the memory often. When I'm undervolted pulling roughly 225w board power, the GPU is 75-100w, it's really not that bad, but the memory and circiutry make up the rest.

But the 2080 has GDDR6, same as AMD's RDNA2 cards, not GDDR6X - if anything, your GDDR6 pulling that much power is an argument against the GDDR6X being the main reason for Ampere's relatively poor efficiency. That would after all indicate similar power draws for AMD's VRAM.

Honestly if people are worried about RTX 40 series being 600w monsters they have a very short memory. Halo cards from many past generations have pulled crazy power, and yeah, the halo cards next gen probably will too, but it doesn't mean a RTX 4060 will want 450w. Get your expectations in check.

GTX 295 - 2009 - 289w
GTX 590 - 2011 - 365w
R9 295X2 - 2014- 500w

Well, that latter GPU is a dual GPU card, so effectively two GPUs in one, and the next GPU down pulled ~300W. Craziness is kind of expected in that regard. And please note that the two others you listed are lower power than current flagships, let alone the rumored 4000-series. Yes, there have always been stupidly power hungry halo cards, even if not every generation. But we've never before been in a situation where it's rumored that even upper midrange cards will exceed 300W. That is completely unprecedented.

Silly halo products for people with more money than sense have been creeping power consumption for bad gains for many years. The only difference now is dual GPU is dead. If you find yourself put off by the price or power consumption of the 3090Ti, this product was never going to be for you in the first place.

That still isn't an argument against noting that these ballooning power figures are worrying. Dual GPU is inherently inefficient, after all, yet now we're seeing single GPUs reach the power levels of dual GPUs previously. Of course the GPU market is much, much larger and has a lot more wealthy whale customers than previously, as well as the ever growing enterprise markets that gobble up as much performance as they can get, but it's still worrying when the design trend moves from "peak performance in X power envelope" (which has typically been ~300W) to "peak performance, power be damned".

IMO, what we're seeing is the effects of a highly competitive market under massive shareholder pressure to maintain an unrealistic level of growth now that silicon manufacturing gains are slowing and architectures are ever more advanced. If shareholders have seen X% growth once, they expect - nay, demand - that to continue, and as companies are legally obligated to appease shareholders (or risk lawsuits), we see designs increasingly move towards the ridiculous. Of course enthusiasts expecting continued massive performance increases isn't helping either.

 

[wolf]

But the 2080 has GDDR6

Typo, I meant 3080 in agreement with nyguen and as my SIG shows.

Craziness is kind of expected in that regard.

That still isn't an argument against noting that these ballooning power figures are worrying

Yeah, crazy is expected for halo products. And if nvidia and AMD alike have learned anything, there are people that will buy these ever increasingly mental products.

Performance per watt will still go up, you will still be able to pick and upgrade that fits a power budget that is reasonable to you. Silly halo products may well push beyond 450w but that doesn't mean there won't be a whole stack of more efficient than previous gen parts too. This is such an outrage nothing burger.

 

[nguyen]

Typo, I meant 3080 in agreement with nyguen and as my SIG shows.


Yeah, crazy is expected for halo products. And if nvidia and AMD alike have learned anything, there are people that will buy these ever increasingly mental products.

Performance per watt will still go up, you will still be able to pick and upgrade that fits a power budget that is reasonable to you. Silly halo products may well push beyond 450w but that doesn't mean there won't be a whole stack of more efficient that previous gen parts too. This is such an outrage nothing burger.

3090 Ti has a major problem is that the people who don't care about prices and only want the best GPU already got the 3090 some 18 months ago and put them in a WC loop , and watercooled + OCed 3090 is gonna perform the same as stock 3090 Ti anyways.

Beside making 3090 Ti as prototype for 4090, I don't see the point of 3090Ti either.

I could see the point for 450W+ next gen GPU if the uarch efficiency is maintained, for example if the 4090 had the same FPS/watt as the 4070 (GPU power consumption only). That way a PC with 4090 actually has better Perf/Watt than one with 4070 when whole system power consumption is taken into account. Yes the PC with 4090 will use more power, but seriously it's just chump change.

I bought a watt-meter for fun and so far my PC (which use ~550W at the wall when gaming, avg 4h/day gaming for me) barely cost me 9usd in electricity a month (0.13usd per KWh here), if I swapped out the 3090 for 3090Ti it would cost me ~11usd a month in electricity, so yeah the power consumption outrage is just nothing burger