NVIDIA GeForce RTX 4090 Evaluate: Ada Dominates PC Graphics

Although NVIDIA has branded this card the GeForce RTX 4090, like its predecessor the RTX 3090, it matches into the identical class beforehand held by NVIDIA’s Titan RTX and older Titan-branded playing cards. As such, the GeForce RTX 4090 isn’t solely a product focused at players. After all, highly effective GPUs with hundreds of cores, excessive clocks, gobs of reminiscence, and the newest media engines are sometimes well-suited to content material creation and gaming, so we did some testing throughout an array of workloads in an try to color an entire image of GeForce RTX 4090 efficiency. However we’re getting forward of ourselves.

Underneath The Hood Of Ada Lovelace And The AD102 GPU

Earlier than we go any additional, we need to spotlight a few earlier articles:

We’ve already lined a lot of the underlying know-how on the coronary heart of the GeForce RTX 4090, so we received’t be doing so once more right here. If you would like the total scoop on Ada Lovelace and the brand new options it brings to the RTX 40 collection, we suggest testing our protection of the structure and of NVIDIA’s different GTC 2022 bulletins. When you’ve bought all of that meat digested, you’ll higher perceive what the GeForce RTX 4090 we’ll be exhibiting you right here immediately is all about.

The brand new AD102-powered GeForce RTX 4090 is beefed-up and enhanced versus the previous-gen GA102-based GeForce RTX 3090/3090 Ti in virtually each means. The AD102 on the coronary heart of the GeForce RTX 4090 is packing 11 GPCs, 42 TPCs, and 84 SMs, versus 7, 42, and 84 within the AD102, respectively. All instructed, the AD102 has extra CUDA Cores (16,384 vs. 10,752), extra RT Cores (512 vs. 336), and extra Tensor Cores (128 vs. 84) versus GA102, and all of these cores are of a newer-generation as effectively.

The AD102 on the GeForce RTX 4090 additionally has a whopping 12× extra L2 cache than the GA102, in addition to extra texture models and ROPs. Along with having considerably extra sources within the GPU, the AD102 additional operates at a lot increased frequencies than the GA102. Whereas the GeForce RTX 3090 Ti supplied a typical Increase Clock round 1,860MHz, the GeForce RTX 4090 operates above 2,520MHz and better, temperatures and energy allowing. As such, the RTX 4090’s texture fillrate and peak compute efficiency are is greater than double the RTX 3090 Ti’s.

All of these new cores and far bigger cache require transistors to construct, and the AD102 is certainly on one other stage over the GA102 in that regard. NVIDIA’s Ampere-based, previous-gen GA102 is comprised of roughly 28B transistors, nonetheless, the brand new Ada Lovelace AD102 includes a huge leap to over 76 billion. NVIDIA was capable of pack in so many transistors by shifting to TSMC’s superior 4N node, versus the customized Samsung 8nm course of employed for Ampere. As well as, although it’s packing roughly 3X the transistors, the AD102 is definitely smaller than GA102. The AD102 GPU has a die dimension of 608 mm² versus 628.4mm² for the GA102. Bear in mind, the AD102 on the GeForce RTX 4090 isn’t totally enabled – 16 SMs are disabled right here. Think about these 16 further SMs and a theoretical RTX 4090 Ti would have much more CUDA and RT cores.

The GeForce RTX 4090 does, nonetheless, have an analogous reminiscence configuration to its predecessors. 24GB of GDDR6X reminiscence is linked to the GPU through a 384-bit reminiscence interface, working at an efficient information fee of 21.2Gbps. That equates to only over 1TB/s of peak reminiscence bandwidth. Although NVIDIA didn’t speak a lot about any new reminiscence compression applied sciences, we suspect the a lot bigger caches in Ada cut back transactions out to border buffer reminiscence, and in the end enhance effectivity because it pertains to reminiscence bandwidth utilization.

Along with growing cores and cache, these further transistors have been additionally used to allow many new options within the Ada Lovelace structure, like Shader Execution Reordering (SER), Displaced Micro-Meshes, Opacity Micro-Masks, FP8 inferencing, a a lot sooner Optical Move Accelerator, and DLSS 3 with AI Body Era. The AD102 additionally includes a new Eighth-gen media engine which includes twin AV1 encoders together with the entire encoding / decoding capabilities of previous-gen architectures. For particulars on these options, please test our earlier protection right here.

Regardless of the entire new know-how at play right here, we also needs to point out that GeForce RTX 40 collection playing cards just like the RTX 4090 are native PCIe 4.0 – not PCIe 5. Nonetheless, we’re nonetheless not saturating the present PCIe 4 interface with these graphics playing cards, so the extra bandwidth that will have been afforded by PCIe 5 most likely wouldn’t imply a lot at this level.

One thing new on the playing cards that’s considerably much like the RTX 30 collection is a small PCIe Gen 5 energy connector. This single 12-pin connection is ready to scale as much as 600 watts for energy supply, and shrinks the general connector footprint versus current 6/8-pin PCIe energy connectors. NVIDIA consists of an adapter with the RTX 4090 that converts a quartet of conventional 8-pin PCIe connectors to the brand new mini-12-pin design, do you have to want one, although ATX 3.0 energy provides are additionally coming down the pipeline that can natively sport the mini PCIe 5.0 connector and are capable of provide sufficient energy to beasts just like the RTX 4090 by way of a single cable. We should always level out, nonetheless, not all of NVIDIA’s board companions have adopted the mini-12-pin connector on their designs; some will stick to full-sized 8-pin connectors on their playing cards.

Along with having the ability to feed the playing cards with copious energy, NVIDIA has additionally optimized the VRM on the GeForce RTX 4090 for higher efficiency, and the playing cards’ cooling answer has been optimized to raised handle the warmth that comes with utilizing all that energy.

The GeForce RTX 4090’s PCB is outfitted with 23 energy phases (20 for the GPU, 3 for the reminiscence). And full PID transient management is supported, for what NVIDIA claims is a 10x enhance in energy administration response occasions. These developments permit GeForce RTX 40 collection playing cards to answer workload calls for a lot sooner than previous-gen playing cards, and general energy supply needs to be extra dependable and constant.

The NVIDIA GeForce RTX 4090 Card And Cooler

Though the cooling answer on the GeForce RTX 4090 appears much like the RTX 3090’s, it has been internally redesigned. The GeForce RTX 4090 has barely bigger, higher-performing followers, a brand new vapor chamber design, and a brand new heatsink structure. The brand new design leads to 20% extra airflow and it pulls warmth away from the GPU extra effectively, with out producing extra noise. By utilizing increased density, extra energy environment friendly reminiscence (constructed by Micron on a extra superior course of) NVIDIA was capable of situate the entire RAM on one aspect of the PCB. The brand new cooler design additionally makes higher contact with the reminiscence. All instructed, reminiscence temperatures ought to stay roughly 10°C cooler than previous-gen playing cards and, as you’ll see slightly later, the GPU doesn’t get very hot both, even after hours of sustained load.

Much like earlier high-end Founder’s Version playing cards, the GeForce RTX 4090’s cooler is outfitted with twin axial followers, and it includes a cut up, pass-through heatsink design. The pass-through cooler on the GeForce RTX 4090 is partially enabled by its dense, quick PCB. One finish of the heatsink is hooked up to a vapor chamber, that’s mounted on to the GPU and reminiscence. The fan above that part directs air by way of the heatsink and instantly funnels it out of the chassis by way of giant vents within the RTX 4090’s case bracket. The heatsink on the again half of the cardboard, which is linked to the entrance vapor chamber through a number of heat-pipes, permits air from the second fan to move all over the cardboard, the place it’s propelled upward and is ultimately exhausted from a system, assuming it’s bought respectable air flow.

Just like the RTX 3090, the brand new GeForce RTX 4090 is one heck of a giant graphics card. In reality, it’s barely wider than a 3090 and measures in at 304mm (L) x 137mm (H) and requires 3-slots (61mm). As we’re positive you’ve seen within the many memes floating across the final couple of weeks, many companion boards are even bigger.

The GeForce RTX 4090 has triple full-sized DisplayPorts (1.4a) and a single HDMI 2.1 output, additionally much like previous-gen playing cards. Ada’s show engine hasn’t been up to date considerably versus Ampere, principally attributable to timing based on NVIDIA. The Ada Lovelace design was locked in earlier than newer show output requirements had been finalized.

When it comes to function assist, there’s much more to say like, NVIDIA Reflex latency discount know-how, NVIDIA Broadcast, Ansel, G-SYNC, Omniverse, optionally available Studio drivers, and GeForce Expertise, to call only a few. Virtually every thing supported by previous-gen playing cards can be supported by Ada Lovelace and the RTX 40 collection – besides NVLink. You’ll discover there’s no NVLink connector on the RTX 4090, so conventional SLI setups are not any extra.

Whenever you see how briskly the GeForce RTX 4090 is, nonetheless, you most likely received’t care…