Table of Contents
Today marks the release of Nvidia’s new RTX 4000 series graphics cards based on the Ada Lovelace GPU architecture. According to Nvidia, the latest video cards provide the most considerable generational performance upgrade in their history, which is made by a few new key innovations, including the new architecture and core for faster ray tracing, improved harder execution reordering, and the new DLSS 3.0 which is said to improve frame rates up to 2x versus DLSS 2.0. Leaked benchmarks show the new series to be quite the performer, so without further ado, we’ll share all the details of the new architecture and specifications, then get into the benchmarks!
Ada Lovelace Architecture and Technologies
Nvidia’s new Ada Lovelace architecture is fabricated on TSMC’s 4N manufacturing process. The smaller process allowed Nvidia to dramatically increase the transistor count, which turns into a lot more cores (70% more CUDA cores than GA102). The new GPU consists of a whopping 76.3 billion transistors, which they state makes it one of the most complex chips ever made.
With the new core, the Ada architecture also operates at higher clock frequencies using “… high-speed transistors in critical paths that could otherwise restrict the rest of the chip”. The RTX 4090, with clocks of 2.52 GHz, sports the highest out-of-box clock frequency of any Nvidia GPU. With the increased performance, you typically see an increase in power; however, at the same power as the RTX 3090 Ti (~450W), Nvidia claims the RTX 4090 GPU delivers up to 2x more performance. 450W is still a lot of energy to use and dissipate, but there’s a lot more performance available at the same power level so the efficiency (FPS per Watt) has improved. These are lofty claims and are confirmed in some titles when using the DLSS feature. Otherwise, you’ll still see a significant increase in raw raster performance, but it’s more to the tune of a 40-50% increase in our testing.
The full AD102 GPU is equipped with 12 Graphics Process Clusters (GPCs), 72 Texture Processing Clusters (TPCs), 144 Streaming Multiprocessors (SMs), along with a 384-bit memory interface (12x 32-bit memory controllers). There are also 288 FP64 cores (2 per SM) that are not shown in the diagrams. Nvidia states the TFLOP rate for the FP64 cores is 1/64th the TFLOP rate of FP32 operations, and the small FP64 core count is included to ensure compatibility with FP64 code which includes Tensor Core code.
The whole-milk version of AD102 GPU includes 18,432 CUDA Cores (3rd Gen), 144 Ray Tracing (RT) cores (3rd Gen), 576 Tensor Cores, and 576 Texture units. The third generation RT core in these Ada GPUs adds a dedicated unit called the Opacity Micromap Engine and the Displaced Micro-Mesh Engine (DMME). The former evaluates Opacity micromaps that are used to accelerate alpha traversal. The DMME allows the Ada RT core to ray trace complex objects, and the environment with significantly less BVH build time and storage costs which, when simplified, makes for a much faster and more efficient pathing. Together, Nvidia claims the Ada 3rd-gen RT core is the most powerful they ever made.
Another significant improvement over the previous generation is the upgrade from DLSS 2.0 to DLSS 3.0. The new sauce inside is a new frame generation technique that combines optical flow estimation with DLSS to improve the gaming experience. In other words, inserting accurately synthesized frames between existing frames improves the frame rate and provides a smoother gaming experience. The most significant performance uptick, you’ll see, is with using DLSS.
The Tensor cores have also been updated. Compared to Ampere, Ada provides more than double the FP16, BF16, TF32, INT8, and INT4 Tensor TFLOPS and runs the Hopper FP8 Transformer Engine, delivering over 1.3 PetaFLOPS of tensor processing on the 4090.
For those who like to stream content, Nvidia’s NVENC encoder and existing optimizations for OBS (Open Broadcaster Software) eliminated the need for a dedicated PC for video capture so you can play and stream at the PC without sacrificing quality or performance. Ada takes that a step further, incorporating support for AV1 video encoding the 9th gen NVENC hardware encoder. Previous generations only decoded but didn’t encode. The AV1 encoder is said to be 40% more efficient than the H.264 encoder found in the 3000 series GPUs. With this update, users can increase from 1080p to 1440p while running at the same bitrate and quality. The 4080 and 4090 use two NVENC encoders for the work.
Details aside, the table below lists the specifications for the new ADA GPUs, including our review sample, MSI RTX 4090 Suprim Liquid X.
Specifications and Features
|Nvidia RTX 4000 Series Specifications|
|Model||MSI RTX 4090 SUPRIM Liquid X||RTX 4090||RTX 4080|
(AD102, AD103, AD104)
|Manufacturing||TSMC 4N (5 nm)|
|Tensor Cores |
|L2 Cache||96 MB||96 MB||64 MB||48 MB|
|Base Clock||2,235 MHz||2,230 MHz||2,210 MHz||2,310 MHz|
|Boost Clock||2,626 MHz||2,520 MHz||2,510 MHz||2,610 MHz|
|Memory Speed||1027 GBps||1008 GBps||736 GBps||504 GBps|
|PCIe 5 16-pin|
|PCIe 5 16-pin|
|PCIe 5 12-pin|
|PCIE 5 12-pin 2x PCIe 8-pin|
3x DisplayPort (1.4a)
|Max Resolution||8K (7680 x 4320)|
|TDP||450+ W||450 W||320 W||285 W|
Before we get into the MSI card, here are a few stock images of the RTX 4090 FE.
The MSI RTX 4090 Suprim Liquid X
As usual, all board partners released their custom cards on day one. This includes MSI, who kindly sent their RTX 4090 Suprim Liquid X, the flagship, for review. Like most aftermarket cards, MSI’s Suprim Liquid X has more robust power delivery, an AIO water cooling system, and total graphics power of over 450W out of the box. Like the previous generation 3090 Ti, the new 4090 card sources its power from the latest PCIe 5.0 12(+4) 12VHPWR connector that offers up to 600 W. MSI recommends a 1,000 W power supply for this card with a minimum of 850W.
Both Suprim Xs include a dual BIOS for silence and a more aggressive profile for gaming. Both BIOS offer the same performance but with different temperatures and fan speeds/noise. The 2,626 MHz boost clock is one of the fastest listed on the market, but as many are aware, performance among its peers will be very close regardless. The Suprim Liquid X variant is priced higher than the Founders Edition by $150 ($1599) and the air-cooled Suprim X by $50. Which for the latter seems to be a good deal (high prices notwithstanding) as it’s able to keep the card running cooler than the air-cooled version for a pittance more.
The RTX 4090 Suprim Liquid X hits the market with an improved over the Founder’s Edition (FE) card power delivery and a 2x120mm AIO to keep the entire card cool. If you’re not into AIOs on your graphics card, MSI has you taken care of with the ‘base’ Suprim X that will do the job and is less expensive. If the fancy design and price point is still too expensive, further down the product stack is the RTX 4090 Gamin X Trio which uses a custom cooler and is likely working with the reference card hardware specification. In short, there’s an RTX 4090 option for everyone.
As far as the appearance goes, the Suprim Liquid X sports a metallic shroud with gold trim covering the business parts on the front while a matching grey backplate protects the rear and helps keep the card cool. Also on the front is a large fan (Torq 5.0) that sits inside an octagon-shaped cutout that further helps remove the heat from the memory and VRMs generated by this 450W monster. Like the air-cooled options, this fan also uses MSI’s Zero Frozr technology that keeps the fan off until cooling is needed. At idle/desktop and watching videos, the fans never spun up and were completely silent (the pump isn’t audible over the existing chassis fans). Even when fully loaded, the cooling solution was barely audible over the system it was installed in, thanks to the 240mm aluminum radiator and the micro-fin copper base. If the noise is still too much, flip the dual bios switch to the Silent mode instead of gaming and the fans ramp up more slowly.
If RGB lighting is your thing, the Suprim Liquid X sports an illuminated chevron next to the fan, along with SUPRIM writing and a Suprim design feature on the back. While the features are certainly visible, it’s not bright enough to steal the thunder from anything inside the chassis. Still, it is a nice compliment to an RGB-laden motherboard or other integrated lighting (controlled through the Mystic Light application).
TORX FAN 5.0
Design improvements to TORX FAN 5.0 result in +23% airflow compared to an axial fan and +10% airflow compared to TORX FAN 4.0.
MICRO-FIN COPPER BASE
The four sides of the copper base feature a thicker surface, allowing for improved heat dissipation on the VRAM and the GPU.
A 240 mm black aluminum radiator dissipates heat out of the liquid stream and allows SUPRIM LIQUID to sustain intense performance.
GOT YOUR BACK
Fine details elevate the backplate such as its brushed metal finish and the beveled edges polished with a light-gold color. Thermal pads beneath the sturdy plate provide additional cooling to the graphics card
MSI MEG SILENT GALE P12
The anti-vibration gaskets sitting on the four corners of the fan are used to reduce rattling caused by the vibration.
The fans completely stop when temperatures are relatively low, reducing the noise significantly when active cooling is not needed. The fans will automatically start spinning again when the heat is on during gaming.
Retail Packaging and Accessories
The retail packaging for the MSI Suprim comes in a predominately white box that includes a picture of the card along with the typical Nvidia Geforce RTX colors and branding. The back of the box shares some features about the card. After opening the box, you’re greeted by a black pad that covers the card inside and the included accessories. Lifting that up exposes the card, tubes, and radiator sitting snug inside the form-fitting foam. Additionally, you’ll find the included mouse pad, instructions, and the all-important 4x 8-pin PCIe to 12+4-pin PCIe 5.0 adapter to power the card.
Meet the MSI RTX 4090 Suprim Liquid X
The MSI RTX 4090 Suprim Liquid X is one of the better-looking cards with the polygon-inspired silver cover, highlight RGB lighting, and the MSI Gaming Dragon sitting on the fan hubs. The top of the card sports the Suprim X branding, which is also adorned with RGB lighting. In all, the Suprim would make a great centerpiece for your system and fits in with most build themes without sticking out. In fact, it would match the Godlike motherboards (among others) well.
A Closer Look at the MSI RTX 4090 Suprim Liquid X
Zooming in on the I/O, we see three DisplayPort ports (v1.4a), and one HDMI (v2.1) port. The former supports a maximum resolution of 8K (7680×4320) while the HDMI port handles up to 4K @ 120Hz. The Suprim Liquid X’s onboard fan pipes some warm air outside of the chassis while the rest goes through the radiator. You’ll want good airflow for this card and especially the air-cooled version to keep things running as they should.
Power is sent through the new 12+4-pin PCIe 5.0 connector, which takes up a lot less space than the traditional 6+2-pin PCIe connectors that we’re a lot more familiar with. Not only does the new connector take up less real estate on the card, but the PCIE-SIG designed 12VHPWR plug is also capable of delivering up to 600 W which is a lot more than a single 6+2-pin is capable of (150 W). MSI includes the proper adapter to connect with the card which, in this case, requires FOUR independent PCIe 6+2-pin connections – do not use piggybacked connectors!
Also pictured in this image is the dual-BIOS switch. While the clock speeds and power limits don’t change, fan behavior does. Just as it’s labeled, the Silent BIOS has a slower ramp-up and Zero Frozr capability, while the Gaming BIOS ramps up quicker and peaks at faster speeds offering more cooling (also uses Zero Frozr).
We have some pictures of the core and memory, but getting the plate off the PCB proved to be a more difficult exercise than I thought. So instead of risking breaking it, we’ve gone without.
Below is a screenshot of GPU-Z showing the clocks we achieved at stock speeds. Out of the box, the MSI RTX 4090 Suprim Liquid X has a listed boost clock of 2,625 MHz, which it maintained much higher (2,790 MHz) throughout our testing. I imagine you won’t see these clocks on the air-cooled version as these typically start to drop boost bins at higher temperatures.
MSI RTX 4090 Suprim Liquid X on the test bench…
Test System and Benchmark Methods
|Test System Components|
|Motherboard||Gigabyte Z690 Tachyon ($349.99)|
|CPU||Intel i9-12900K (stock)|
|CPU Cooler||Corsair iCUE H150i|
|Memory||Kingston Fury Beast 2×16 GB 5200 MHz CL40 ($161.99)|
|SSD||Mushkin Helix 1TB NVMe (OS + Applications – $80)|
|Power Supply||EVGA 850 W P6 ($119.99)|
|Video Card||MSI RTX 4090 Suprim Liquid X ($1,749) / 521.90 driver|
Our test system is based on the latest (at the time of publishing) mainstream Intel z690 platform and uses the i9-12900K 8P,8E/24t CPU. The CPU runs stock. The DRAM is in a 2×16 GB configuration at 5200 MHz with CL40 timings, a middle-of-the-road option balancing performance, and cost.
Since the last update, we have made some changes and updated titles. More details can be found in the GPU Testing Procedure article, which we have updated with our latest benchmarks. Below is a quick summary for easy reference.
- UL 3DMark Time Spy – Default settings
- UL 3DMark Fire Strike (Extreme) – Default settings
- UL 3DMark Port Royal – Default Settings (Ray Tracing capable cards only)
- Unigine Superposition – 1080p High, 1080p Extreme
- Shadow of the Tomb Raider – DX12, “Highest” preset
- The Division 2 – DX12, Ultra preset, VSync Off
- Assassin’s Creed: Odyssey – Ultra High preset, VSync Off
- Far Cry 6 – Ultra defaults, HD Textures enabled
- F1 2021 – DX12, Very High defaults, Bahrain track, show FPS counter.
- Metro: Exodus – DX12, Ultra defaults
Our first set of benchmarks hails from Underwriters Laboratories, which acquired Futuremark back in 2014. Earlier in 2018, a rebrand occurred, and since that time, Futuremark is now UL. The benchmarks have not changed, just the name. We chose to stick with 3DMark Fire Strike (Extreme) and 3DMark Time Spy as these tests give users a good idea of performance on modern titles.
3DMark Fire Strike (Extreme) is a DX11-based test that runs at 1080p resolution. UL says the graphics are rendered with detail and complexity far beyond other DX11 benchmarks and games. 3DMark Time Spy is a DX12 benchmark designed for Windows 10 PCs. It supports new API features such as asynchronous compute, explicit multi-adapter, and multi-threading and runs at 2560 x 1440. 3DMark Port Royal is the first Ray Tracing benchmark designed for Windows PCs and graphics cards with Microsoft DirectX Raytracing capabilities.
As we can see from the charts, the RTX 4090 Suprim Liquid X is considerably faster than the flagship RTX 3090 Ti and dominated the competition, even at the more CPU-bound resolution of 1080p. On average, the increase in these scores over the 3090 Ti is 57% (low of 46%, high of 79%) which is an incredible increase at this low resolution. But will it hold in games?
We have updated our testing suite for gaming benchmarks to bring more modern titles into the mix. Gone are Battlefield V, F1 2018, Far Cry 5, AOTS:e, and World of Tanks, which have been replaced with Metro Exodus, Assassin’s Creed: Odyssey, F1 2021, and Far Cry 6. We kept The Division 2 and Shadow of the Tomb Raider. The games should provide a good view of the card’s overall performance with many DX12-based. One thing worth noting with these high-end cards, they aren’t really made for 1080p gaming, so the gaps between cards tend to get minimized.
1920 x 1080 (1080p) Results
The RTX 4090 wins in almost everything here. This time, the average across all titles is 14% with a low of 0.5% (Far Cry 6) and a high of 42% (Assassin’s Creed). If you’re buying this card for 1080p, you’d better own a 360 Hz monitor to make it worth the expense.
2560 x 1440 and 4K UHD Results
Below are the higher resolution results starting with 2560 x 1440 and the gaining-in-popularity 3840 x 2160 (4K UHD) resolution.
Here, at more GPU-board resolutions, the RTX 4090 smokes the rest of the field. At 2560×1440, we see even more impressive gains with an average increase of 32% across the same titles.
Moving up in resolution, our RTX 4090 is much faster than the competition again. In fact, we’d go as far as saying it could handle [email protected] for high res, high FPS gaming. The difference here between it and the 3090 Ti averages a whopping 60%. You really get the most for your money the higher resolution you play… but let’s bring in some ray tracing and DLSS results and see how DLSS 3.0 fares in testing.
Ray Tracing and DLSS Testing
Below, we tested Shadow of the Tomb Raider with ray tracing at 1440p and 4K. Even with the frame-reducing ray tracing enabled (without DLSS), the RTX 4090 easily handled this title.
Folding at Home
For all of our folders out there, and especially those here at Team 32, we took the card out for a spin in Windows 11 22H2 to see what kind of points this beast could gather. After three full days, our RTX 4090 averaged just over 23 million points. The AIO kept the card to a maximum of 51C (the hotspot was 59C). Power use was around 300W with that particular WU (will vary by WU). Of course, those true to the cause will run Linux and likely squeeze out even more PPD, but I wanted to show what a quick and dirty installation looked like.
Overclocking the MSI RTX 4090 Suprim Liquid X
For overclocking, we used the latest version of MSI Afterburner without issue. The OC Scanner also did the job of finding a stable overclock. In the end, we settled on +78 Mhz for the core and +200 for the memory. While the scores didn’t go up much (a couple of percent) on these mid-resolution benchmarks, there was still some improvement to be had. This card also allows for a 10% increase of the power limit (3090 Ti didn’t allow any) in case you find yourself limited in that respect. With these settings, the card peaked at 2,865 MHz and ran around that speed consistently through most of our benchmarks.
Temperatures and Power Use
We test power consumption by running through the game benchmarks of Shadow of the Tomb Raider and F1 2021 at stock speeds and while overclocked. We monitor temperatures throughout this testing, with the peak temperature listed in the data below. The benchmarks are extended (time) to allow the card to settle to simulate real gaming conditions more accurately.
Temperatures on the liquid-cooled MSI RTX 4090 Suprim Liquid X reached 64°C in F1 2021 and 63°C SOTR at stock settings. After overclocking the card, temperatures increased slightly in each game, peaking at 66°C in F1 2021 and SOTR. Longer gaming sessions will increase the temperatures, but this gives you a general idea of the capabilities. For cooling up to 450W, the temperatures are well in order and much better than any air-cooled card. Keeping the temperatures in this range also allows the card to use all the boost bins and isn’t limited by temperatures.
Power use for the RTX 4090 Suprim Liquid X peaked at 640 W (system total power) at stock and reached 646 W while overclocked (both in F1). At minimum, a high-quality 850 W power supply will be adequate for your needs, but a 1,000 W model is recommended. Power use didn’t change much from the 3090 Ti Suprim X, however, the performance increase at the same power was significant.
Wrapping things up, Nvidia hit a home run with performance. From the raw raster improvements to DLSS 3.0, FPS hasn’t been higher, and graphics cards incredibly capable. However, that comes at a price. On the front end, an RTX 4090 costs you around $1499 dollars for an FE version, more for custom high-end models like the Suprim, Strix, and Aorus from card partners. The other side where it can bite you is with power consumption and heat. With power use reaching 450W (and more) the cards get thicker, or you need a liquid-cooled solution like the Suprim Liquid X we have from MSI.
The card we have takes the monster Nvidia created and takes it to a whole different level. With a 240mm radiator to control the room heater, temperatures are well under control, and it does so while being relatively quiet. To get the most out of these cards and get all of the boost bins you can, watercooling is the way to go. If this card is left to sit in the 80s as the 3090 Ti’s did (at least, the Suprim X I have), you’ll lose a few bins to keep the temperature at the default limit.
The most painful part to swallow is the price. People, it doesn’t seem like they are coming down anytime soon. At almost $1600, it’s ‘only’ $100 more than the RTX 3090. From a performance per dollar perspective, it’s a lot better, but tell that to our bank account that doesn’t count frames, just dollars. On top of the expense, it’s not easy to mitigate 450W worth of heat, but the Suprim Liquid X’s AIO does a great job of it. In order to run this card, you’ll need a beefy power supply with four PCIe 6+2 connectors and a minimum of 850W. Most (all?) cards that require the new PCIe 5.0 connector will come with a high-quality adapter to use, or if you’re connector averse, you’ll need to purchase a new power supply that has the connector attached natively (will be ATX 3.0 spec).
Price notwithstanding, our overall take on these cards is that they are one hell of an improvement over the previous generation. All of the updated hardware and DLSS 3.0 combine to form, by far, the fastest graphics card we’ve tested. Even without DLSS in use the card is 14% faster at 1080p, 32% faster at 2560×1440, and 60% faster at 4K UHD in our tests which is one of the more significant jumps we’ve seen (in higher resolution) in generations. Add the DLSS 3.0 in and ray tracing improvements, and it makes AMD’s job to beat it on a performance front incredibly difficult. While I wouldn’t rush out to get one of these if I had a 3000 series and it was performing well for my needs, those with older cards have a real reason to upgrade if the price points for all of these cards don’t get in the way.
– Joe Shields (Earthdog)