Following on from Nvidia’s announcement, we’re now starting to see more detailed specifications for the three initial RTX 40 Series graphics cards. Key differences between the two RTX 4080 siblings have already been detailed right here at Club386, and we now know that Nvidia’s top duo falls short of realising the full power of Ada Lovelace architecture.
GeForce RTX 4090, RTX 4080 16GB and RTX 4080 12GB are all carved from unique dies – AD102, AD103 and AD104, respectively – yet it is only the weakest of the trio that serves as a complete implementation. RTX 4080 12GB carries a full 60 SM units and 7,680 CUDA cores, representing the best that the near-36-billion transistor chip has to offer.
| AD102 | AD103 | AD104 | |
|---|---|---|---|
| Die size | 608mm2 | 379mm2 | 295mm2 |
| Transistors | 76.3bn | 45.9bn | 35.8bn |
| SM Units | 144 | 80 | 60 |
| CUDA Cores | 18,432 | 10,240 | 7,680 |
| RT Cores | 144 | 80 | 60 |
| Tensor Cores | 576 | 320 | 240 |
| ROPs | 192 | 112 | 80 |
| L2 Cache | 96MB | 64MB | 48MB |
| Memory Bus | 384-bit | 256-bit | 192-bit |
Things get interesting at the top-end of the scale, where Ryan Smith has filled in some of the gaps. RTX 4080 16GB makes use of 76 out of 80 SMs available to the 379mm2 AD103 die. As yields improve on TSMC’s 5nm process, it is logical to assume a GeForce RTX 4080 Ti may one day surface with all 10,240 cores ready to rock.
Even more headroom exists on AD102. At present, RTX 4090 employs 128 of 144 possible SMs. If Nvidia feels the need, there’s potential to boost core count by 12.5 per cent. Enough, one would think, for an RTX 4090 Ti, or possibly even an all-new Titan. No surprise – chip designers often leave room for such positioning – and whether or not more powerful cards see the light of day could well depend on what rival RDNA 3 brings to the table.
To recap, here’s how the first wave of GeForce RTX 40 Series graphics cards take shape:
| GeForce | RTX 4090 | RTX 4080 (16GB) | RTX 4080 (12GB) |
|---|---|---|---|
| Die | AD102 | AD103 | AD104 |
| SM Units | 128 | 76 | 60 |
| CUDA Cores | 16,384 | 9,728 | 7,680 |
| RT Cores | 128 | 76 | 60 |
| Tensor Cores | 512 | 304 | 240 |
| ROPs | 192 | 112 | 80 |
| FP32 Teraflops | 83 | 49 | 40 |
| RT Teraflops | 191 | 113 | 82 |
| Tensor Teraflops | 1,321 | 780 | 641 |
| Boost Clock | 2.52GHz | 2.51GHz | 2.61GHz |
| Base Clock | 2.23GHz | 2.21GHz | 2.31GHz |
| Memory Config | 24GB GDDR6X | 16GB GDDR6X | 12GB GDDR6X |
| Memory Bus | 384-bit | 256-bit | 192-bit |
| Graphics Power | 450W | 320W | 285W |
| MSRP | $1,599 | $1,199 | $899 |
