Basics:

Nvidia's GeForce RTX 3090 Founders Edition is the fastest GPU around but at double the price of an RTX 3080. Founders Edition is absolutely massive from price to performance and most gamers don't need it. Here we show you the full review from price to performance and much more in our in-depth review. 

 Specs:

  • NVIDIA CUDA Cores                                                                                            10496 
  • Base / Boost Clock                                                                                                1.40 / 1.70
  • Memory Config                                                                                                 24 GB GDDR6X
  • Memory Interface Width                                                                                             384-bit
  • Ray Tracing Core                                                                                                   2rd Generation
  • Tensor Cores                                                                                                           3rd Generation
  • NVIDIA Architecture                                                                                                   Ampere
  • DirectX 12                                                                                                                       Yes
  • NVIDIA DLSS                                                                                                                 Yes
  • PCI Express Gen4                                                                                                          Yes
  • NVIDIA Ansel                                                                                                                   Yes
  • NVIDIA ShadowPlay                                                                                                         Yes
  • NVIDIA G-SYNC                                                                                                               Yes
  • NVIDIA Studio Drivers                                                                                                       Yes
  • NVIDIA GPU Boost                                                                                                            Yes
  • Vulkan RT API, OpenGL 4.6                                                                                               Yes
  • HDMI 2.1                                                                                                                              Yes
  • DisplayPort                                                                                                                           Yes
  • NVIDIA Encoder                                                                                                                  7th Gen
  • NVIDIA Decoder                                                                                                                  5th Gen
  • VR Ready                                                                                                                                 Yes
  • Maximum  Resolution                                                                                                          7680*4320
  • GPU Temperature                                                                                                                        93
  • GPU Power (W)                                                                                                                          350

The Nvidia GeForce RTX 3090 Founders Edition presently asserts the best position on our GPU benchmarks progression, however where it lands on the best design cards is a more troublesome inquiry to reply to. While the GeForce RTX 3080 kept with the current estimating structure of the RTX 20-arrangement, supplanting the RTX 2080 Super at the $699 value point, the RTX 3090 sees the RTX 2080 Ti's $1200 dispatch cost and raises it another $300. There's an elective perspective, in any case: The GeForce RTX 3090 is likewise a trade for the $2,500 Titan RTX, in which case it's a quicker card that costs $1,000 less. In any case, you will require some profound pockets on the off chance that you need to claim Nvidia's new radiance card. 

We as of now have the Ampere design profound jump that subtleties what makes the GeForce RTX 3090 tick. Start there in the event that you need more data on how the 3090 and GA102 convey new degrees of execution. We'll cover the features here, yet the basic answer is that Nvidia has gone with a more modest cycle hub, more centers, quicker memory … in addition to more power, and a more exorbitant cost. 

We have a couple of extra tests we're actually attempting to run before we render our last decision (taking a gander at you 8K, or if nothing else 4K with DSR copying 8K), so we won't have the last decision at this time. We additionally have a couple of AIB (include load up) accomplice cards that we'll be taking a gander at in the coming days, whenever we've had some an ideal opportunity to run all the tests and accumulate some information. 

The short synopsis of the GeForce RTX 3090 is much as you'd anticipate. It's hypothetically about 20% quicker than the RTX 3080 dependent on specs alone, with over double the memory. The extra VRAM doesn't generally make a difference much for most games except if you're running at 8K (perhaps 5K), yet it can demonstrate help for some expert outstanding burdens. The RTX 3090 likewise requires a bulky CPU to capitalize on the card (we'll be running some extra tests in the coming days to show that, however, you can get some thought of what's in store from our RTX 3080 CPU scaling article). This is a lot of a GPU intended for 4K ultra gaming, and at those settings, it's 10-12% faster than the 3080 on normal in our current test suite, and 14% quicker in our reward test suite. Drop down to 1440p ultra, and the 3090 is just 8% quicker than the 3080. In case you're running at 1080p? Try not to, truly. 

 GeForce RTX 3090 Architecture:

To start with, GA102 utilizes Samsung's 8N cycle innovation, which implies a bigger number of semiconductors in a more modest region than TU102. Taking a gander at generally speaking semiconductor thickness, GA102 packs in 45 million semiconductors for each square millimeter, though TU102 thickness is 'just' 24.7 million semiconductors for every square millimeter. That is acceptable, however obviously not tantamount to TSMC's N7 hub: The bigger GA100 chip utilized in Nvidia's A100 contains 65.4 million semiconductors for every square millimeter. Additionally, AMD's Navi 10 has 41 million semiconductors for each square mm, so Nvidia has, in any event, coordinated that degree of thickness — contrasting across models is certainly taking a gander at apples and oranges, however. 

Proceeding onward, the GeForce RTX 3090 utilizations an almost complete GA102 chip. Of the 84 expected SMs (streaming multiprocessors), just two are impaired. That proposes either the yields are awesome … or Nvidia isn't wanting to sell close to the same number of 3090 chips as 3080 chips. We presume the subsequent choice is nearer to reality, and real yields are a firmly watched mystery nowadays. The 3090 likewise has seven GPCs (illustrations handling bunches), and the ROPs (render yields) are currently essential for the GPC rather than the memory regulator, giving the 3090 112 ROPS. 

Maybe a greater change is that the GeForce RTX 3090 comes outfitted with 24GB of GDDR6X memory, this time timed at 19.5 Gbps (contrasted with 19 Gbps on the 3080). This is refined by empowering the last two 32-digit memory regulators on GA102, and afterward running 24 chips into equal parts width 16-bit interface mode. That additionally implies the GDDR6X chips are situated on the two sides of the PCB (printed circuit board), while the RTX 3080 just has memory on a similar side as the GPU. There's absolutely an issue of how much this influences GDDR6X memory temperatures, especially when a large portion of the VRAM isn't effectively cooled. Notwithstanding, as of now, we don't have an approach to gauge the GDDR6X chip temperatures. 

We've examined the GPU CUDA center changes a lot in past Ampere articles since it's quite possibly the main contrast between Ampere and Turing. Turing included 64 FP32-proficient CUDA centers per SM, with an extra 64 INT32-competent CUDA centers (and two FP64 centers for similarity purposes). The FP32 and INT32 centers were discrete datapaths and could be utilized simultaneously. The FP32 centers could likewise run FP16 math at twofold the exhibition — 'quick math' mode. 

For Ampere, the FP32 centers lose uphold for FP16 quick math, which gets moved over to the Tensor centers — it's still double the FP16 execution of the devoted FP32 centers, however just equivalent FP16 execution generally speaking. In the interim, the INT32 centers pick up help for FP32 counts, which means absolute FP32 execution per SM has multiplied. Simultaneously, INT32 execution per SM has remained the equivalent. This makes them premium repercussions for general execution, however fundamentally, around 30-35% of the CUDA center remaining task at hand in games is INT32 (for address pointer queries, surface computations, and other comparable work). That implies a decent piece of the second datapath will be occupied with INT32, so this present reality execution lift will regularly be not exactly the crude FP32 TFLOPS number would propose.

Somewhere else, the tensor centers are another significant update comparative with Turing, with twice the throughput percenter. Besides, there is half the same number of tensor centers per SM. The third era tensor centers do add uphold for fine-grained sparsity, which can twofold throughput again comparative with Turing when utilized. Additionally important is that the third-gen tensor centers add uphold for INT8 and INT4 information types at 2x and 4x the base FP16 throughput. 

The video ports on GeForce RTX 3090 FE are equivalent to on RTX 3080: one HDMI 2.1 port and three DisplayPort 1.4a yields. There's no VirtualLink port this round, as VirtualLink is essentially dead. Every one of the four yields is fit for 8K60 utilizing DSC (Display Stream Compression), an "outwardly lossless" procedure that is really not generally outwardly lossless (however you presumably won't see at 8K). We've seen different cards with five presentation yields, many picking double HDMI 2.1, so remember that in case you're wanting to utilize a multi-screen arrangement. 

Different changes (and we're disregarding things here as we've just shrouded this in the RTX 3080 survey, just as the Ampere engineering piece) incorporate RT centers that are up to twice as quick — or considerably more in specific outstanding tasks at hand (like movement obscure). The L1 reserve/shared memory limit and transfer speed have been expanded. The L2 reserve is additionally bigger than previously, and the L1 store can be arranged as changing measures of L1 versus shared memory, contingent upon the necessities of the application. Register document size is likewise expanded, and GA102 can do simultaneous RT + designs + DLSS (beforehand, utilizing the RT centers would stop the CUDA centers). 

The raster administrators (ROPS) have been moved out of the memory regulators and into the GPCs. Each GPC has two ROP segments of eight ROP units each. This gives greater adaptability in execution, giving the GA102 and RTX 3090 an aggregate of 112 ROPS. Each GPC additionally incorporates six TPCs (Texture Processing Clusters) with eight TMUs (Texture Mapping Units) and a polymorph motor, however, these come two by two and Nvidia empowers 41 TPCs for the 3090. 

At long last, as talked about previously, the RTX 3090's memory regulator has another component called EDR: Error Detection and Replay. At the point when the memory identifies a bombed transmission, instead of smashing or adulterating information, it essentially attempts once more. It will endeavor to do this until effective, however, it's as yet conceivable to cause an accident with memory overclocking. EDR permits the potential for higher memory timekeepers that may not really perform better as you approach the restrictions of the memory. That is on the grounds that EDR winds up decreasing absolute memory throughput when bombed transmissions happen.

Previous Post Next Post