What appear to be final specs leaked recently for the upcoming NVIDIA GeForce RTX 3090 and NVDIA GeForce RTX 3080. The two Ampere GPUs look positively monstrous, with specifications that align with a number of earlier leaks.
At the top of the food chain, the GeForce RTX 3090 sports 5,248 CUDA cores, 24GB of GDDR6X VRAM at 19.5 GB/s, and an eye-watering 350W TGP. You might need to upgrade that PSU if you haven’t already done so. The GeForce RTX 3080 tones things down a bit, with 4352 CUDA cores, 10GB of GDDR6X at 19 GB/s, and a milder 320W TGP. Boost clocks on the RTX 3090 top out at 1695 MHz, while boost clocks on the RTX 3080 are a little higher at 1710 MHz. While both cards feature high-clocked GDDR6X, the RTX 3090 delivers nearly 1 TB/s of bandwidth thanks to a 384-bit bus, while the RTX 3080 offers a still-formidable 760 GB/s.
At the lower end of the spectrum, we have some input about the GeForce RTX 3070, too. This card will sport a 220W TGP, in line with the GeForce RTX 2070 Super. We don’t have a clear idea about CUDA core counts and clock speeds. However, the GeForce RTX 3070 will feature 8GB of 16 GB/s GDDR6 across a 256-bit memory bus for 512 GB/s of bandwidth. All three of these parts look like they could deliver a decent 4K experience. We’ll be sure to let you know how they perform when we get our hands on them.
| RTX 3090 | RTX 3080 | RTX 3070 | |
|---|---|---|---|
| GPU | 7nm GA102-300 | 7nm GA102-200 | ?nn GA104-300 |
| Board | PG132 SKU 30 | PG132 SKU 10 | PG142 SKU 10 |
| CUDA Cores | 5248 | 4352 | TBC |
| Boost Clock | 1695 MHz | 1710 MHz | TBC |
| Memory | 24GB G6X | 10GB G6X | 8GB G6 |
| Memory Clock | 19.5 Gbps | 19 Gbps | 16 Gbps |
| Memory Bus | 384-bit | 320-bit | 256-bit |
| Bandwidth | 936 GB/s | 760 GB/s | 512 GB/s |
| TGP | 350W | 320W | 220W |
Strangely, contrary to popular rumors, it seems that NVIDIA’s Ampere lineup will use the 7nm process, rather than Samsung’s 8nm LPP process. It’s unclear whether this is TSMC’s or Samsung’s 7nm EUV node, but regardless, both are a notable step over the 12nm process that powered Turing.
