The dust has hardly settled from the launch of Nvidia's RTX 3080, but already we are back with a look at the next GPU in the stack. Sitting at the head of the Ampere family, the RTX 3090 is nothing short of a monster in terms of its spec, boasting 82 Streaming Multiprocessors, 10496 CUDA Cores, 24GB (!) of GDDR6X memory, and total memory bandwidth approaching 1TB per second. Here we look at Gigabyte's implementation of this GPU in its Eagle OC graphics card, and see just how fast this GPU is.
The Gigabyte RTX 3090 Eagle OC will look very familiar. Just last week, we reviewed the RTX 3080 Eagle OC, finding it to be a very capable card that also offers excellent value. ‘Value' is not something that applies to the RTX 3090, and this model is priced at £1449 here in the UK. For that price, we get a triple-fan cooler, metal backplate, factory overclocked core and a hint of RGB lighting – on top of the performance of the RTX 3090 GPU itself.
GPU | RTX 3090 | RTX 3080 | RTX 2080 Ti (FE) | RTX 2080 SUPER | RTX 2080 (FE) |
SMs | 82 | 68 | 68 | 48 | 46 |
CUDA Cores | 10496 | 8704 | 4352 | 3072 | 2944 |
Tensor Cores | 328 | 272 | 544 | 384 | 368 |
RT Cores | 82 | 68 | 68 | 48 | 46 |
Texture Units | 328 | 272 | 272 | 192 | 184 |
ROPs | 112 | 96 | 88 | 64 | 64 |
GPU Boost Clock | 1695 MHz | 1710 MHz | 1635 MHz | 1815 MHz | 1800 MHz |
Memory Data Rate | 19.5 Gbps | 19 Gbps | 14 Gbps | 15.5 Gbps | 14 Gbps |
Total Video Memory | 24GB GDDR6X | 10GB GDDR6X | 11GB GDDR6 | 8GB GDDR6 | 8GB GDDR6 |
Memory Interface | 384-bit | 320-bit | 352-bit | 256-bit | 256-bit |
Memory Bandwidth | 936 GB/Sec | 760 GB/Sec | 616 GB/sec | 496.1 GB/sec | 448 GB/sec |
TGP | 350W | 320W | 260W | 250W | 225W |
RTX 3090 uses the GA102 GPU, which also forms the basis for RTX 3080. Even with its 82 SMs, however, this is not actually a full implementation of GA102, as a fully populated die houses 84 SMs, so just two are missing from the RTX 3090. Ampere has doubled the number of FP32 shaders per SM, so there's a total of 128 CUDA Cores in each of those 82 SMs, for a grand total of 10496.
Each SM also houses a single RT Core, as well as four Tensor cores, giving us 328 with the RTX 3090. Additionally, there's 328 texture units and 112 ROPs.
As for the memory, this is a significant step up from the RTX 3080. RTX 3090 of course still uses GDDR6X memory, but it has over double the amount with 24GB. Additionally, the modules are clocked slightly faster at 19.5Gbps, and the memory interface is also slightly wider than RTX 3080, with a 384-bit interface. Total memory bandwidth comes in just shy of 1TB/s, hitting 936 GB/s.
Finally, the one are where Gigabyte has tweaked things is with the clock speed. RTX 3090 defaults to a boost clock of 1695MHz, but Gigabyte has added 30MHz to this, bringing it up to 1725MHz. Later in the review we test exactly how fast this GPU actually operates, as we would expect GPU Boost to bring clock speed well above the rated speed.