NVIDIA GeForce RTX 3080 FE Review

NVIDIA GeForce RTX 3080 FE: Introduction

Hello and welcome to another review! Today we will be looking at the newest generation of GPUs from NVIDIA under the Ampere architecture. With this comes a slew of new performance improvements over NVIDIA's previous Turing Generation (20XX) series. NVIDIA has a lot to live up to based on the hype the company has created, touting a 1.5x performance increase over the RTX 2080 Super and built upon Samsung's 8nm technology. On top of this, NVIDIA Founder Editions brings with it a new cooler for whisper-quiet operations based on the newly designed dual fan air cooler to keep the temperatures low during those long gaming sessions.

NVIDIA is set to release the RTX 3090, RTX 3080, and RTX 3070 this year. In this review, I will be covering the RTX 3080 Founders Edition in detail followed by a future review of the other products in this lineup. We will be covering multiple systems, games, resolutions and ray-tracing, temperatures, and power consumption. This is a monstrous review with lots of topics, so let's get started!

NVIDIA GeForce RTX 3080 FE: Closer Look

With this Ampere generation (2nd Gen RTX) a lot of minor architectural improvements have been made over the Turing GPUs. On top of the massive 8704 CUDA Cores, which is exactly 4352 more than NVIDIA last year's flagship 2080 Ti, a large portion of the Streaming Multiprocessors (SM), RT, and Tensor cores have been reworked to maximize the improvements in this Ampere GPU generation. With a massive 28.3 billion transistors (GA102), all these combined make for some major uplifts in the performance that fundamentally changes how game developers can utilize these optimizations fully in future AAA games. NVIDIA pointed out a 1.9x performance increase per watt over the Turing series in Ray-Tracing applications. A deeper investigation into this claim is covered later in the review.

Each SM contains 128 CUDA Cores, four third-generation Tensor Cores, a 256KB register file, four texture units, one second-generation Ray Tracing Core, and finally 128KB of L1 Shared Memory. The memory section of the GA102 GPU consists of twelve 32-bit memory controllers (creating 384-bit total for the RTX 3090) 512KB of L2 cache that is paired with each 32-bit memory controller. This brings the total to 6MB of L2 cache for the full GA102 GPU.

The Ampere generation includes a 5th generation NVDEC (NVIDIA hardware-based decoder) and also includes new improvements. With it comes AV1 hardware decoding and brings H.265 8K encoding up to 12-bit Chroma 4:4:4. This is a big improvement over the previous generation and HEVC (H.265) is finally reaching mass adoption now that streamers and content creators can share video high-quality videos without waiting hours to encode. Encoding will be quicker and YouTube will soon see an increase in 8K content as well as more 4K content thanks to NVIDIA GPUs off-loading most of the encoding tasks from Professional applications like BlackMagic DaVinci Resolve and Adobe Premiere.

Now let's talk about architectural improvement in detail. First up is the improvement in the SM core with Programmable Shader being increased to two shaders calculations per clock. Next is the 32-bit floating-point (FP32) throughput which now has double data paths. When optimized it can be best described as double the performance with the same amount of CUDA Cores. In the Turing generation, each partition had two primary datapaths. However only one could process FP32, while the other was limited to integer maths. With Ampere, the FP32 now flows down both datapaths. Ampere now executes 128 FP32 operations per clock (32 per partition). In practice, the total uplift will be less due to the nature of this improvement.

It seems Integer math has been left unchanged, meaning that some applications will only see the increase purely from the extra CUDA cores, as INT32 does not have dual data paths in the Ampere GPU. As game engines better optimize for this change, the performance gained from doubling the FP32 paths and combining them will only increase over time. Next, let's talk about the RT Core, which is used for off-loading the math involved in bounding boxes and triangle intersection for ray-tracing. This Ampere generation RT Core has also been revamped with an improved triangle solver dealing with the bounding boxes and now is considered in its second iteration (Gen 2). This intersection throughput for bounding boxes has been doubled bringing the total RT-TLFOPS to 58. On top of this improvement, new accelerations have been introduced to handle blurred objects. This leads to an 8x performance increase for dealing with motion blurred ray-traced objects.

The very underappreciated Tensor core among gamers has been overhauled also. As most are aware the Tensor core is design for handling AI Deep Learning among things. The 3rd generation Tensor Core has also been reworked to be more efficient with the biggest architectural change having less per SM, but far more powerful in the instruction set. This turns out to be double the performance over the Turing GPU for the RTX 3080. Upgrade to the GA100 and it's four times in certain deep learning scenarios. While this Tensor core isn't used to great effect in games, in the right talented hands, this can be a revolution in off-loading functions from the CPU to the GPU. It also allows for better DLSS as more background noise reduction operations can be run at the same time, giving a clear imagine overall without an impact on framerate.

Lastly is new GDDR6X memory has major improvements over GDDR6 as it now carries four bits per clock cycle instead of the traditional two bits used in every other memory on the market today. Before going further, to clear up the future commentary on this subject, it is still DDR technology but instead of one-bit per raise and fall, two-bit are encoded per signal. This allows NVIDIA to avoid using the extremely expensive HBM2 memory in favor of a proven cheaper method with higher production yields. This isn't without some trade-offs. This mainly has to deal with this GDDR6X being very sensitive to temperature and frequency. As such, the methods of just sending four signals via a different set of voltages would not work on its own. The signal is first encoded before entering the memory pipeline, allows a clear and clean signal to come out the other end. This also means memory overclocking will be underwhelming due to being less forgiving as the frequency is raised past the specifications provided by NVIDIA. I would go as far as to say that memory overclocking with GDDR6X is not a real viable option for many end-users.

NVIDIA GeForce RTX 3080 FE: Closer Look Continued

Alright now we have the improvements covered for this Ampere GPU generation, we can get to the GTX 3080 actual video card itself. NVIDIA Founders Edition has been steadily improving its presentation and design over the years. With this series, the biggest change to the packaging is not just a darker box, but actually how the video card is packaged inside. To open you will be laying the box down with the logo facing upwards. This is a 180 from the previous Turning (20XX) Founders Edition that opened by pulling up on the box. This previous method didn't fair very well for everyone because while your hands are on the box, the video card is free to fall out onto the ground. I had a few close calls personally with that design.

NVIDIA noticed this could be a problem and this time around once you pull off the outer box, the video card is firmly placed without any chance of it moving around. If you are into box collecting, this doesn't look like an expensive item until you get a closer look. It also smells like a new pair of shoes!

Included with every Founders Edition is the Microfit 3.0 to dual 8-pin PEG dongle. While NVIDIA has made the 12-pin connector publicly available for anyone to use, it might take a while before mass adoption can occur, if ever. The jury is still out on if this connector is a smart design choice, but at least it is included. One concern is the warranty card. It states that the warranty may become void if any other adapter is used. I understand the concern that many people have. Well, for one thing, the only way NVIDIA can tell a different adapter was used is if the cable melts from being overloaded. After all, a lot of currents are passing through this connector. My only advice for those who want to use something else is just making sure the wires are 18AWG or you might be in for a surprise.

This might be the first time I spent the majority of my time during photographing just staring at the card instead of doing any actual work. The design choice here comes from the necessity to remove a large amount of heat while keeping the sound and temperatures in check. The cooler uses a two fan design, one to push air out the I/O bracket and the other upwards into the case itself. This design relies on a case with high airflow, especially for the rear fan exhaust. While the sides do have fins, these are more passive due to the frame blocking the most direct airflow from the fans.

The main concern I have been hearing about across the Internet is the concern of the heat generated by the video card passing over the memory. Well, I bring the news that it may be an issue for cases with low air-flow. This cooler is designed to exhaust a large amount of heat and it is up to the end-user to have a good amount of airflow in the case. This is no different than any of the high-end video cards in the past. The difference here is that instead of that heat lingering down below, the hot air is pushed right up in the hopes the case exhaust fans will wisp it away quicker than if it was allowed to passively exit the case. This cooler design allows for a smaller footprint at the expense of other computer components. But let it be known, no matter the cooler solution, the heat generated from the video card has to go somewhere.

I am happy to report that NVIDIA has delivered on all fronts. The RTX 3080 operates at 320 watts under full load. While the card gets a bit toasty after hours of gaming, it doesn't thermal throttle under these conditions. It stays around 72 °C. The uni-body design gives the video card a prestige look and speaks volumes to the quality. Dare I say a sort of elegance that showcases the NVIDIA engineering team's ability to come out with something revolutionary in cooling that is not just a form of beauty, but serves a purpose in function as while in design.

The I/O section has gone under a radical change and resembles a bit of the older blower fan design. While technically a fan is pushing air out the I/O bracket, it isn't the only way the video card expels the heat generated. Below the large open area is the monitor display connections and this time around are a bit of a mixed bag in my opinion. It is no surprise the push for 8K content is upon us as game consoles, TV manufacturers, and now NVIDIA as well is pushing for this image fidelity revolution. But consumers are going to have to wait a bit longer. Besides the lack of content for TVs, HDR has been making strives to stand out and not become the next 3DTVs and perish due to lack of adaption.

Both 8K and HDR are the biggest marketing strategies now and in the coming years. But because the DisplayPorts are all version 1.4a on this card instead of 2.0, it means PC gamers are going to be left behind from the highest possible refresh rate and image quality. This also means that those who have been waiting for higher 4K refresh rates above 144Hz or 8K 60Hz gaming must use the single HDMI 2.1 connector. But don't be fooled by those numbers. HDMI 2.1 is still limited in that aspect with a Chroma Sampling of 4:2:0 at those higher resolutions and refresh rates due to having to rely on Display Stream Compression (DSC) and limited cable bandwidth. The reason for DP 2.0 being absent is most likely due to 2.0 not being finalized until late 2019, long past the development and design stages of Ampere. Maybe we will see a revision next year during the cycle refresh. HDMI 2.1 is not superior for PC gaming, it does however play a pivotal role in Home Theaters setups, just not in the eye-candy PC gaming scene.

The rear section has two holes designed for OEM system builders to attach a bracket. Many companies like Dell, HP, and Alienware will do this for shipping procedures and it also serves to keep the card stiff and in place.

Okay, this is the point where the elegant design falls apart a bit. For those who care about the looks of the case interior and cable placements, there is nothing you can do. The adapter is bulky and sticks out the middle of the card. The reason for the Microfit 3.0 connector in the first place is to allow a much higher power draw with a smaller footprint. The Founders Edition PCB was designed around the idea of a cooler that can pass air through the back. Therefore the PCB is about 2/3rd the length of previous generations and the connector is actually at the end of the PCB. Hence why it looks like it comes out of the middle of the card. Now companies like Corsair and Seasonic are making adapters that go directly to the PSU. This will skip the 12-pin to dual 8-pin adapter and be able to deliver even more current. So for those who like a neat interior case, extra purchases will be in order.

Let's talk about the maximum power delivery that the MicroFit 3.0 can provide. I was waiting for NVIDIA's official comment but at the time of publishing, I have not received any response to my query about this power connector capacity. Therefore this next bit of information is speculation due to not having all the information needed for a proper conclusion. First, I can squash the notation that two 8-pin PEG connectors are not enough. While it is true the official ATX specification states that a single 8-pin PEG connector should not exceed 150 watts, it has been noted that this is based on an old standard using 23AWG wire. Any decent power supply will use 18AWG wire allowing for up to 10 amps per 12V pin or 360 watts in total.

Now, this isn't recommended without some prior investigation of your hardware, but it is possible. Now with this Microfit 3.0 connector, it seems I have two bits of conflicting information on Molex's website. One white paper says a maximum of five amps per 12V pin, which would result in 360 watts. The other more plausible option is 8.5 amps per 12V pin resulting in 612 watts. So it's safe to say at least at stock speeds the RTX 3080 FE is okay to handle its power rating. This miss-match of information will be even more of an issue as third parties release their adapters. Once again just make sure it is 18AWG wire to avoid any potential chances of the wires melting.

Due to time restraints, I did not take apart the sample I was provided for review. However, NVIDIA generously provided PCB pictures and heatsink separately. This design is brand new for NVIDIA as it looks as though the PCB was built around thermals as the heatsink is an integral part of the video card and not an afterthought. This becomes even more evident when NVIDIA stated that the Founder Editions will not be providing this PCB design as the standard reference. Meaning that this is an NVIDIA custom PCB design and AIB partners will need to either use a different PCB layout provided by NIVIDA or create a different one entirely.

This also will cause a problem for those who want to water cool this beast of a video card. The water blocks will most likely cost a bit more this time around as the "Reference Design" is no longer interchangeable with the Founder Editions name. Every AIB partner video card will need a close inspection for water block compatibility. Chances are that limited quantities will be produced for each vendor.

Both the RTX 3080 and RTX 3090 use the same PCB and it can be seen that missing MOSFETs and memory is reversed for the RTX 3090. I have asked NVIDIA to provide the parts used and while it is subject to change due to future supply. Currently, as it is today the RTX 3080 Founders Edition is a 15+3 phase design from Monolithic Power Systems. Fifteen phases for the GPU using the MP2888BGU-1228 and three phases for the memory using MP2888BGU-1228-Z. This is all driven by the MP86957GMJ VRM Controller. The total current output is not available at this time. As more information becomes available this section will be updated post-launch. Considering this video card pulls 320 watts under load, it safe to assume, just like all the other Founders Editions in the past, this is built to last.