Even setting aside Frame Generation, this is a seriously fast, power-hungry GPU.
See full article...
See full article...
Review: Nvidia’s GeForce RTX 5090 is the first GPU that can beat the RTX 4090
- Thread starter JournalBot
- Start date
You are correct. No one is going to pay $2000 for it. You can be assured the scalpers will pay them all up on day one and charge $2500-$3000 for them and people will buy them.
Upvote
14
(14
/
0)
Ohhhh that is absolutely the headline they should have gone with. The power/price/performance ratios are essentially unchanged (based on MSRP).
I'm just assuming that is what the actual retail prices will settle at, like it did with the 4090.
Upvote
2
(2
/
0)
Heh, my workstation 4090 + w3495x can heat the room by several degrees when I ask it to run a erosion simulation. It's an underappreciated problem. A machine that's pulling 800-1100 watts for hours at a time is basically a space heater. They are at the point where for a high end machine you may need to talk with your HVAC contractor it keeping the temperature comfortable is a requirement.
The other problem is the UPS. The current 1500va ups occasionally beeps because the attached computer goes over it's rated capacity. With a 5090, that's going to be a regular occurrence. Consumer grade UPSs are not ready for this.
Upvote
19
(19
/
0)
Back in the early oughts, we had LAN parties late at night since they produced so much heat back then and we lived in teh desert. Would be hard on home circuits with multiples of these.
Upvote
7
(7
/
0)
That’s good to hear. I hope they work their way up the performance ladder.
Upvote
3
(3
/
0)
This video about the cooling system of the 5090 shows they had a 4-slot prototype:
View: https://www.youtube.com/watch?v=RDr1pr_c6ts
Upvote
5
(5
/
0)
that’s my understanding from what I have read. They seem hell-bent on sticking with a monolithic chip architecture. they’re gonna ride that horse all the way into the ground. I just can’t help but see that for a 30% performance increase use 35% more power? yeah that’s definitely a scalability issue.
Upvote
7
(7
/
0)
We're in interesting times - picked up a very well used 3090 MSI RTX for roughly 650$ last year, I am doing some local RAG stuff so my primary need is the VRAM, the card is actually more expensive to find used today...at least here in Europe
Upvote
1
(1
/
0)
Heh. Wondering now - DLSS and frame generation add artifacts, would those be worse if you used this thing for VR? I remember VR had trouble pushing the needed amount of frames at the right res even with the biggest, baddest cards running tandem.
Disclaimer: I don't even own games that are VR-compatible, everything secondhand with me.
Disclaimer: I don't even own games that are VR-compatible, everything secondhand with me.
Upvote
1
(1
/
0)
Chiplet architectures are much harder for GPUs than CPUs. AMD got an efficiency improvement from RDNA3 versus RDNA2, but not an enormous one.
CPUs rely on low latency (delivered via on-chip caches) while putting relatively little pressure on bandwidth, which means there's less of a power / performance penalty from connecting things together via chiplets and a central I/O die. GPUs are fairly latency tolerant, but require a lot of memory bandwidth across the core. If you put two chips side-by-side on the same PCB and connect them via an interconnect, the resulting solution will consume more power transferring data than if you built the entire chip in monolithic silicon.
AMD had to design an all-new approach for RDNA3 -- it couldn't just port its Ryzen approach over. While RDNA3 did improve on RDNA2's power efficiency and overall performance, it's a real challenge to build a disaggregated GPU.
Upvote
10
(10
/
0)
As I'm actually in the market for a new build right now (quite possibly my "last one" as I will hopefully retire in a couple of years), this was a fun comment thread (no offense, Ars, but for the details I'll go to GN).
I may actually "have to" go for a 5090 over a 5080, because it's the AI that I now dive into nolens-volens, grumbling, kicking and screaming. Gotta do what pays the bills - and also what makes the whole thing tax deductible where I live. My 3090 is starting to creak (no OC)
Before our home reno, I was always concerned that turning on the a/c and the stovetop - which tended to blow a breaker - would take my system down in mid-work, so had a UPS just for that (rather than what most people do, in case of a mains power loss
). Fortunately I live in a 240V country - but if I were in the US, I really would be looking at it like the "Concern" meme. The tables have turned, induction cooktop!
I love the idea some commenters threw around of somehow integrating these modern monsters with a heat exchange system for the house. Imagine taking a nice comfy bath courtesy the waste heat after a long hard sloggy day training LLM models! Instead of turning the planet into a second sun (which would make Mars the equivalent of Tatooine, I suppose?) Yes, I know the heat is still being produced, but at least we use PV, trying to minimise fossil power use, and like most, probably failing miserably.
That all said, who knows if and when I will be able to get one here, despite the fact that those US$2K will turn into over twice that in our local currency (while salaries are nominally close or even lower).
I may actually "have to" go for a 5090 over a 5080, because it's the AI that I now dive into nolens-volens, grumbling, kicking and screaming. Gotta do what pays the bills - and also what makes the whole thing tax deductible where I live. My 3090 is starting to creak (no OC)
Before our home reno, I was always concerned that turning on the a/c and the stovetop - which tended to blow a breaker - would take my system down in mid-work, so had a UPS just for that (rather than what most people do, in case of a mains power loss
). Fortunately I live in a 240V country - but if I were in the US, I really would be looking at it like the "Concern" meme. The tables have turned, induction cooktop!I love the idea some commenters threw around of somehow integrating these modern monsters with a heat exchange system for the house. Imagine taking a nice comfy bath courtesy the waste heat after a long hard sloggy day training LLM models! Instead of turning the planet into a second sun (which would make Mars the equivalent of Tatooine, I suppose?) Yes, I know the heat is still being produced, but at least we use PV, trying to minimise fossil power use, and like most, probably failing miserably.
That all said, who knows if and when I will be able to get one here, despite the fact that those US$2K will turn into over twice that in our local currency (while salaries are nominally close or even lower).
Upvote
2
(5
/
-3)
Honestly, I was about to get one from a 4090 (I'm one of those, I know, it's a sickness. At least I use the Vram). But after seeing the frame gens and dlss4 is going to be on 40s too, I'm not so keen. My monitor is a 2k, and I already hit 110fps on cyberpunk at max which is (no kidding) what i need to not have a headache. Beyond work related things, the 5090 doesn't seem like it will offer much more than what I already have outside of work.
Along with going up from a 5950x to a 9950x3d makes more sense due to bottlenecking i already have a tiny bit (what? I think thats going to be maybe a few percent at worst, I do OC). Then cruze along happily until I get a 6090 or whatever down the road, on sale to barely bottleneck the 9950x3d. Save money in the process.
Along with going up from a 5950x to a 9950x3d makes more sense due to bottlenecking i already have a tiny bit (what? I think thats going to be maybe a few percent at worst, I do OC). Then cruze along happily until I get a 6090 or whatever down the road, on sale to barely bottleneck the 9950x3d. Save money in the process.
Upvote
0
(0
/
0)
TGP is apparently not 575W but 725-750W in "some" reviews online? What are these reviewers doing wrong or not doing correctly? (this is with the founders edition, not AIB, so one would assume it should be locked)
Why would Nvidia publish TBP/TGP at 575W when its exceeding the limitations of their connector?
Why would Nvidia publish TBP/TGP at 575W when its exceeding the limitations of their connector?
Upvote
5
(5
/
0)
Anecdote: I have a VR display (Samsung Galaxy), but I hate using it, so I'm out of the VR loop by a couple years.
However, it is generally considered that nvidia cards have the best frame reprojection for VR displays, which syncs camera movements to your head movements regardless of whether the update happens during a frame refresh or not.
I expect the 5000 series to be at least as good as previous gens, and I also expect all the DLSS4 tech (including FG) to play nicely with their existing infrastructure.
Upvote
1
(1
/
0)
Sounds more like either total system power or the reviewer messed up bad.
Upvote
6
(6
/
0)
And the performance increases of 2000 - 2006 were proportionally much larger than those from 2012 - 2018. Remember when Nvidia had an ironclad rule about a new architecture every year, with a six-month kicker refresh in-between?
Things changed. Dennard Scaling stopped. Lithography improvements slowed -- first to walk, and now to a crawl. GPU architectures matured. There are fewer radical departures from how things were done in the past than there once were.
I'm not saying Nvidia couldn't be doing more than it has -- bringing more VRAM to lower price points is an obvious place to start -- but the entire reason the industry has pivoted towards AI is because continuing to increase rasterization / RT performance by building bigger, ever-more power-hungry chips isn't feasible. Chiplet GPU designs are difficult because the need for high bandwidth puts a lot more pressure on interconnect performance and power consumption than a CPU chiplet design does. They aren't a "Get Out of Scaling Problems Free" card -- if they were, AMD would've swept Nvidia's legs out from under it with the 7000 family.
To put it differently: If Intel or AMD thought that Nvidia was lying about the importance of AI to the future of graphics, they'd be saying so. They'd be release power point decks and PDFs decrying "fake frame" bullshit and emphasizing how they deliver REAL FRAMES(TM).
Instead, both companies have taken a page from Nvidia's book and worked on their own answers to technologies like DLSS and FG.
The reason everyone is moving in this direction is because the graphics industry wants to re-enable the very scaling you remember fondly, and believes AI is the only way to simultaneously improve performance and deliver new capabilities like RT. I'm not making any predictions over how successful that's going to be, but that's definitely the goal. Underneath Nvidia's marketing push for AI are some very hard truths about physics that we don't have answers for.
Upvote
14
(14
/
0)
I wish I can do that.... the CF reinforced plastic prints aren't as good as the stamped steel cases...
Upvote
0
(0
/
0)
I've had that a few times myself with a 4090 and 13900k. I'm seriously considering a 2nd ups for just the computer and put the monitor and assorted peripherals on the first only.
Upvote
2
(2
/
0)
Upvote
4
(4
/
0)
That definitely helps. Did that a while back and it mostly solved it for me. Should completely fix your issue with the lower TDP on the 13900k.
Upvote
1
(1
/
0)
My guess is a mix of all but perhaps mostly 1. If you use an AIO, using a pull setup rather than push may make more sense with a 5090FE in the case. That way the CPU will get fresh cool air rather than air that has been mixed with the 5090FE's exhaust.
Upvote
3
(3
/
0)
Not really but I would look for number and size of mounts on front back and top for fans. That is really what is going to be the thing.
Upvote
0
(0
/
0)
Nvidia is on its way to 1000 Watts, while Apple is headed in the other direction 100 Watts for everything…… with convergence of GPU performance by the M5-M6 Ultras.
Upvote
-5
(2
/
-7)
It is a great piece of engineering, however I have two issues that I cant shake from my mind: the first one is the ratio power performance, in order to get 30% extra performance it uses 30% extra power. Does this mean we already hit the limits and we are just trying to build around it (frame generation technology seems to indicate this).
My other concern, and it is related to the first one, is if the amount of power drawn in the prev gen was already causing serious problems (read melted connectors) is there a way to avoid it now? The tilted plug helps with cable management, but does it help with the temp itself?
My other concern, and it is related to the first one, is if the amount of power drawn in the prev gen was already causing serious problems (read melted connectors) is there a way to avoid it now? The tilted plug helps with cable management, but does it help with the temp itself?
Upvote
8
(8
/
0)
Don't try to do it locally. Just use your 3090 for development.
Use a Cloud Computer from Google or AWS or the like to actually run your larger test and training on. A 3090 is WELL above what is needed to actually run a fully-trained AI, and even a 5090 isn't anywhere near good enough to do real training on. Just rent a cloud system with Teslas on it, and it will be done in 10% of the time a 5090 will. for under $5/hour.
Upvote
1
(4
/
-3)
Techspot measured 698W at the PCIE slot and 12VHPWR connector in Cyberpunk 2077 at 4K.
Upvote
3
(4
/
-1)
Do you mean you have the Samsung Odyssey VR Headset?
I had this headset many years ago. Not a very good headset IMO and severely outdated now. I sold it and replaced it with a Vive Pro 1 (also pretty old by now). I don't think you can even use the Odyssey with modern versions of windows after Microsoft stopped supporting windows Mixed Reality.
Last edited:
Upvote
1
(1
/
0)
Why stop there? Blower-design cooling can make a comeback, but this time using traditional 120mm fans aimed out the back of a 7-slot graphics card. Especially since the actual PCB area isn't much longer than 120mm, and the height of PCI slots are just above 120mm. Just imagine the beauty of TWO gigantic heatpipe blocks inside your case. Two cubed is four times as cool! That's reason enough to skip water-cooling, right there.
Upvote
2
(2
/
0)
IIRC, it was cable connectivity issues that caused the melting, not just the amount of power drawn.
Regarding your "Have we hit the end of scaling?" question: No. I'm certain we haven't.
Not every GPU architecture is a tremendous efficiency leap above the one that came before. This is also true in CPUs, even though CPUs are architecturally quite different than GPUs. AMD, Intel, and Nvidia have historically improved performance in a range of ways, including IPC improvements, boosted memory bandwidth, larger cores, and higher clockspeeds.
To my eye, it looks like Blackwell is overwhelmingly designed to improve AI and secondarily RT workload performance, including the adoption of GDDR7. It isn't a big step forward for the architecture because AI workloads are far more memory bandwidth (and memory capacity) bound than compute bound.
Nvidia was able to make long-term plans like this because it doesn't have a great deal of market competition right now. Intel did something very similar in the mid-2010s when AMD wasn't competitive in server, but Nvidia is hauling in profits that Intel could only dream of back then.
Blackwell's design reflects Nvidia's dominant market interest right now. But that doesn't mean the company has no interest in continuing to improve gaming performance per watt. The entire pivot towards AI is an effort to continue delivering improved performance per watt and higher raw frame rates.
In a world where MFG worked flawlessly and was universally supported, we'd all be hailing Nvidia as a champion for actually delivering RTX 4090 performance ($1600 in 2022) for $549 in 2025. That kind of improvement would be worthy of the Good Old Days. In the real world, the RTX 5070 isn't going to reasonably match the RTX 4090, but I do think technologies like DLSS / FSR4 will be increasingly important to gaming. If MFG can be made to work well, it could revolutionize both handheld and high frame rate gaming by dramatically improving perf/W and maximum FPS to take advantage of high refresh rate monitors.
MFG doesn't work flawlessly and it isn't universally supported, but you don't have to drown in Nvidia's Kool-Aid to see the larger argument the company is making.
Scaling will continue outside of AI. Lithography gains are nothing like they once were, but there will still be stepwise improvements that add up to meaningful gains over 5-10 year timescales. Rasterization performance and IPC will still improve. But the hope is that AI can boost that limited rate of return into something that better approximates the improvement rate of 10-20 years ago.
Upvote
5
(5
/
0)
Reviews for the 5080 are going to be interesting. The 5090 seems like it might “get away” with 4-frame generation without incurring unreasonable lag because it has also enough grunt to achieve a decent “real” frame rate. Meanwhile the 5080 is a lot less powerful in terms of raster and raytracing; frame gen might feel the same playability-wise, or it might feel a lot worse.
Upvote
5
(5
/
0)
My problem is I have a 4K display for work and all resolutions except HD and 4K look blurry. Is there a 4K display (or higher) that looks sharp at 1440p? Don't have space for 2 displays either.
Upvote
0
(0
/
0)
That's because 4K is a perfect multiple of 1080p. This allows for integer scaling.
There are 5120x2880 displays (5K) and above. They are less common, but they exist. There are even higher resolutions, if you need them.
Upvote
11
(11
/
0)
Replace card with car and you'll understand how remarkable the chip revolution of the last decades has been.
Upvote
4
(4
/
0)
Buy the card instead and let it mine Bitcoin while it is heating the room. It may pay itself back eventually.
Upvote
-8
(1
/
-9)
nVidia is the best popularizer of consoles.
After watching the joy of pricing announcements for new cards my wallet voted for PlayStation.
After watching the joy of pricing announcements for new cards my wallet voted for PlayStation.
Upvote
-5
(1
/
-6)
Upvote
5
(5
/
0)
Can’t say I agree that slapping on top a giant fan and heat sink to cram in as much transistors as possible is great engineering.
Upvote
-4
(0
/
-4)