NVIDIA didn’t send me an RTX 5090 for testing (that is fine it is not obliged to send me anything), so I invested 2.5K euros to buy one from a local shop (Msystems.gr) to bring you today’s review. I believe you will find some info that is unique and you will also find out why you need a good and strong power supply for the RTX 5090. In general the results I got are not the expected ones, especially when it comes to performance per watt.
During CES, I had the opportunity to watch NVIDIA’s CEO speech, which included the new GPUs and graphics cards. In the end, after the smoke settled down, things proved not to be as he described them, especially regarding the pricing strategy. Unfortunately, this is typically the case with every new GPU release.
Since I am not in talking terms with NVIDIA and no other board partners could send me an RTX 5090, I was forced to buy one to bring you this review, with detailed power and noise analysis information, besides performance on countless games and settings. I find it difficult to believe that some believe that the RTX 5090 is efficient, from the moment it maxes out the power capacity of the 12V-2×6 cable, leading to countless problems so far. In today’s review, I will provide data showing that the RTX 5090 is a step backward in terms of efficiency (aka performance per watt), and I admit that I didn’t expect such results. From the moment NVIDIA is left entirely alone in the high-end GPU field, with AMD staying in the mid and low market segment and Intel restricted to the budget GPU categories, there is no one to push NVIDIA to improve its products further. However, I must admit that AMD delivered way more than expected with the Radeon RX 9000 models, which I will evaluate next week.
Besides not having optimal performance and efficiency, NVIDIA can also decrease production and focus more on AI, which is more profitable than gaming, leading to sky-high prices. To make matters even worse, besides the RT 50 paper launch we also have to deal with several significant issues with the few graphics cards sold, from missing ROPs to melted connectors, because NVIDIA’s engineers weren’t capable or didn’t have the required time or weren’t instructed to implement a proper protection features circuit on this design. Given the 600W of sustained power that passes through the RTX 5090’s 12V-2×6 header, there should be over current and over temperature protection on all required pins to save the day when cable/header temperatures go wild. The needed protection features are not there though, on these super expensive graphics cards, leaving users in distress and feeling helpless, trying to find solutions to protect their costly parts, given that they were lucky enough to score an RTX 5090 or 5080. This is unacceptable, but we can do nothing since NVIDIA currently has the monopoly in the high-end (and not only) GPU market.
Because I bought this card out of my own money and with no sponsorships, I won’t perform a part analysis since I need to keep it under warranty if anything happens. Given the problems reported with 12V-2×6 headers on RTX 5090 cards, I think this is the best way.
The tables below show the specs of the RTX 50/40 series and AMD models.
NVIDIA Graphics Cards Major Specs
RTX 50 Series
| RTX 5090 | RTX 5080 | RTX 5070 Ti | RTX 5070 | |
| Architecture | GB202 | GB203 | GB203 | GB205 |
| Process Technology | TSMC 4N | TSMC 4N | TSMC 4N | TSMC 4N |
| Base Clock (MHz) | 2010 | 2300 | 2300 | 2160 |
| Boost Clock (MHz) | 2410 | 2620 | 2450 | 2510 |
| VRAM Size (GB) | 32 | 16 | 16 | 12 |
| VRAM Type | GDDR7 | GDDR7 | GDDR7 | GDDR7 |
| VRAM Speed (Gbps) | 28 | 30 | 28 | 28 |
| VRAM Bus Width (bit) | 512 | 256 | 256 | 192 |
| Transistors (Billions) | 92.2 | 45.6 | 45.6 | 31.0 |
| Shading Units | 21760 | 10752 | 8960 | 6144 |
| TMUs/Tensor Cores | 680 | 336 | 280 | 192 |
| ROPs | 176 | 128 | 96 | 64 |
| SM/RT Cores | 170 | 84 | 40 | 48 |
| TDP (W) | 575 | 360 | 300 | 250 |
| Launch Month/Year | 1/2025 | 1/2025 | 1/2025 | 1/2025 |
| MSRP ($) | 2000 | 1000 | 750 | 550 |
RTX 40 Series
| RTX 4090 | RTX 4080 Super | RTX 4070 Ti Super | RTX 4070 Super | RTX 4070 | RTX 4060 Ti | |
| Architecture | AD102 | AD103 | AD103 | AD104 | AD104 | AD106 |
| Process Technology | TSMC 5N | TSMC 5N | TSMC 5N | TSMC 5N | TSMC 5N | TSMC 5N |
| Base Clock (MHz) | 2235 | 2295 | 2340 | 1980 | 1920 | 2310 |
| Boost Clock (MHz) | 2520 | 2550 | 2610 | 2475 | 2475 | 2535 |
| VRAM Size (GB) | 24 | 16 | 16 | 12 | 12 | 8 |
| VRAM Type | GDDR6X | GDDR6X | GDDR6X | GDDR6X | GDDR6X | GDDR6 |
| VRAM Speed (Gbps) | 21 | 23 | 21 | 21 | 21 | 18 |
| VRAM Bus Width (bit) | 384 | 256 | 256 | 192 | 192 | 128 |
| Transistors (Billions) | 76.3 | 45.9 | 45.9 | 35.8 | 35.8 | 22.9 |
| Shading Units | 16384 | 10240 | 8448 | 7168 | 5888 | 4352 |
| TMUs/Tensor Cores | 512 | 320 | 264 | 224 | 184 | 136 |
| ROPs | 176 | 112 | 96 | 80 | 64 | 48 |
| SM/RT Cores | 128 | 80 | 66 | 56 | 46 | 34 |
| TDP (W) | 450 | 320 | 285 | 220 | 200 | 160 |
| Launch Month/Year | 9/22 | 1/2024 | 1/24 | 1/24 | 4/23 | 4/23 |
| Street Price ($) | 2500 | 1600 | 915 | 640 | 570 | 399 |
The table below contains the specs of AMD’s current-generation graphics cards for reference purposes.
AMD Graphics Cards Major Specs
| RX 9070 XT | RX 9070 | RX 7900 XTX | RX 7900 XT | RX 7900 GRE | RX 7800 XT | RX 7700 XT | |
| Architecture | Navi 48 | Navi 48 | Navi 31 | Navi 31 | Navi 31 | Navi 32 | Navi 32 |
| Process Technology | TSMC N4 | TSMC N4 | TSMC N5/N6 | TSMC N5/N6 | TSMC N5/N6 | TSMC N5/N6 | TSMC N5/N6 |
| Base Clock (MHz) | 1660 | 1330 | 1929 | 1387 | 1880 | 1295 | 1435 |
| Boost Clock (MHz) | 2970 | 2520 | 2498 | 2394 | 2245 | 2430 | 2544 |
| VRAM Size (GB) | 16 | 16 | 24 | 20 | 16 | 16 | 12 |
| VRAM Type | GDDR6 | GDDR6 | GDDR6 | GDDR6 | GDDR6 | GDDR6 | GDDR6 |
| VRAM Speed (Gbps) | 20 | 20 | 20 | 20 | 18 | 19.5 | 18 |
| VRAM Bus Width (bit) | 256 | 256 | 384 | 320 | 256 | 256 | 192 |
| Transistors (Billions) | 53.9 | 53.9 | 57.7 | 57.7 | 57.7 | 28.1 | 28.1 |
| Shading Units | 4096 | 3584 | 6144 | 5376 | 5120 | 3840 | 3456 |
| TMUs | 256 | 224 | 384 | 336 | 320 | 240 | 216 |
| ROPs | 96 | 80 | 192 | 192 | 160 | 96 | 96 |
| Compute Units | 64 | 56 | 96 | 84 | 80 | 60 | 54 |
| Ray Tracing Cores | 64 | 80 | 96 | 84 | 80 | 60 | 54 |
| TDP (W) | 304 | 220 | 355 | 300 | 250 | 263 | 245 |
| Launch Month/Year | 3/25 | 3/25 | 11/22 | 11/22 | 7/23 | 8/23 | 8/23 |
| Street Price ($) | 599 | 549 | 930 | 750 | 550 | 500 | 420 |
Intel Graphics Cards Major Specs
| Arc B580 | Arc B570 | |
| GPU (Architecture) | BMG-G21 (Xe2) | BMG-G21 (Xe2) |
| Process Technology | TSMC 5N | TSMC 5N |
| Base Clock (MHz) | 2670 | 2500 |
| Boost Clock (MHz) | 2670 | 2500 |
| VRAM Size (GB) | 12 | 10 |
| VRAM Type | GDDR6 | GDDR6 |
| VRAM Speed (Gbps) | 19 | 19 |
| VRAM Bus Width (bit) | 192 | 160 |
| Transistors (Billions) | 19.6 | 19.6 |
| Shading Units | 2560 | 2304 |
| TMUs/Tensor Cores | 160 | 144 |
| ROPs | 80 | 80 |
| EU/RT Cores | 20 | 18 |
| TDP (W) | 190 | 150 |
| Launch Month/Year | 1/2025 | 1/2025 |
| MSRP ($) | 249 | 219 |
- GPU: GB202-300-A1
- Architecture: Blackwell 2.0
- CUDA Cores: 21760
- Boost Clock: 2407MHz
- OC mode: 2437 MHz
- Memory Speed: 28 Gbps
- Standard Memory Config: 32 GB
- Memory Interface Width 512-bit GDDR7
- Memory Bandwidth: 1792 GB/sec
- PCI-E 5.0
- Fans: 3x 92mm
- Fan Stop: Yes
- Ports: 3x DisplayPort 2.1b, 2x HDMI 2.1b
- Dimensions(with Bracket): 348 * 146 * 72mm
- Required slots: 3.6x
- Weight: 2315gr
- Warranty: three years
- MSRP (excluding VAT): unknown
- Prologue & Technical specifications
- NVIDIA’s Key Technologies
- Box & Contents
- Part Analysis
- Specifications Comparison
- Test System
- Game Benchmark Details
- Raster Performance
- RT Performance
- RT Performance + DLSS/FSR Balanced
- Raytracing Performance + DLSS/FSR Balanced + FG
- DLSS/FSR Balanced (No RT)
- DLSS/FSR Balanced + FG (No RT)
- Relative Perf & Perf Per Watt (Raster)
- Relative Perf & Perf Per Watt (Raster + DLSS/FSR)
- Relative Perf & Perf Per Watt (RT)
- Relative Perf & Perf Per Watt (RT + DLSS/FSR)
- Relative Perf & Perf Per Watt (RT + DLSS/FSR + FG)
- Rendering Performance
- Operating Temperatures
- Operating Noise & Frequency Analysis
- Power Consumption
- Clock Speeds & Overclocking
- Cooling Performance
- Epilogue
Can you do coil whine analysis? I’ve bought and returned so many cards because coil whine is unbearable with a silent pc.
In game menus, if there is no FPS limiter, most have coil whine, at desktop this is not the case.
Thank you for the review. Very good stuff.
When can we expect a review and the cybernetics report for the new FSP Mega TI 1350W PSU ? Looking forward to it i think it could be a very good performer at least i hope so.
Will ask from FSP review samples. Unfortunately lately I am over my head with work, and trying to find a solution on the 12V-2×6 thing, so limited time for PSU reviews.
Aris did Gamemax brings you some NDA for their new LION CORE PSU, i’ve seen that they just claimed tested & certified by Cybenetics ?
We did test a Gamemax PSU, but it is already on Cybenetics and it isn’t this model. I will look into this now that you mentioned it.
https://www.youtube.com/watch?v=D-7TbwosSC0 After watching this video, I want to ensure their brand never overclaims Cybenetics certifications. If they overclaim, it could be a fatal flaw. Alternatively, you can advise their brand to never overclaim that a product wasn’t tested by Yours.
I paired one with the FSP Hydro Ti PRO 1000W as you recommend in your tier lists, I was really looking forward to reviews of it!
Just received it 2 days ago and first impressions are great, I have it running at 45% constant fan speed and it’s increadibly silent, only when it gets closer to 65ºC or so I ramp up to 55% and there in really high loads it increases until 60% or so, keeping it under 70ºC without problems, while still being very quiet.
I tried a bit of undervolting and can do 0.93V with +412 MHz keeping pretty much 98% of the performance in 3DMark tests, and passing the Steel Nomad stress test.
I have measured the connector temperature with a FLIR E4 both at the PSU (~35ºC) and at the GPU (~55 to 65ºC as you say) while the stress test was running, I don’t see a problem at all. It’s clear that it gets hotter on the GPU as it is very close to the heatsink and very hot parts on the PCB.
I would be curious to see if you could do some undervolt analyses 🙂 I think there’s a lot of potential to reduce the crazy power it is uses. I agree that using it at maximum power for long times is bonkers, my entire computer enclosure was really hot after the stress test.
Cheers!
I also noticed in this graphics card, and not only, that the 12+4 header is so close to the GPU’s heatsink. It might not affect as much as I initially thought, but probably isn’t the best spot.
This period I am over my head with various projects and reviews, but I will do more work on the 5090 and examine more usage scenarios. This card required strong airflow for sure!
Yes, I have it in a good old Phanteks Enthoo Evolv (front modded) and it becomes a sauna in there…. But my 13700KF with a LF III modded to P12 MAX fans does a pretty good job to keep it cool in real usage scenarios, while the GPU also remains at decent temperatures.
I’m a big fan of silent PCs and I know that temperatures are not that big of a deal for electronics components (electronic engineer myself working in automotive industry) so I’m fine with that.
But of course, power is power and it has to go somewhere, so the environment around gets pretty warm.
Looking forward to seeing more tests when you have the time 🙂