Nvidia’s GeForce RTX 4090 may look incredibly powerful, but when it debuts (at least until AMD’s RDNA 3 GPUs come along) it ranks as the fastest option on our list of best graphics cards. definitely not, but the RTX 4090’s stripped-down AD102 die isn’t. It’s far from reaching the full potential of the AD102 with all cores and cache enabled. Combine this with additional enhancements and you’ll see future RTX 4090 Tis much faster and possibly even more expensive It may become
Nvidia RTX 40 series and Ada Lovelace GPU specs, but these only point to announced and rumored cards. Nvidia’s full AD102 die features 144 SMs, 18,432 CUDA cores, 96MB of L2 cache, and 192 ROPs. This equates to 12% more CUDA cores and a whopping 33% more L2 cache capacity compared to the current RTX 4090. A fully enabled AD102 die has 9% more ROPS and 12% more texture mapping units thanks to the extra SM.
But that’s not all the future 4090 Ti can do. Micron is working on a new 24Gbps GDDR6X memory module that will be 14% faster than the RTX 4090’s 21Gbps module and even faster than the RTX 4080 16GB 22.4Gbps module that Nvidia claims is currently the fastest in the world. is. This pushes the hypothetical (but very likely) RTX 4090 Ti to a bandwidth of up to 1152 MB/s.
However, faster memory comes with higher power consumption, so we suspect Nvidia is seriously constraining the AD102’s full clock speed and power potential. Rumors of a 600W RTX 40 series graphics card? We do know that Nvidia managed to overclock his RTX 4090 above his 3.0GHz and this will definitely increase the power consumption.
It looks like the Ada architecture and TSMC’s 4N process still have plenty of headroom left over the RTX 4090’s 2520 MHz boost frequency. If the process matures a bit more and Nvidia is willing to raise the power limit, I wouldn’t be surprised if the RTX 4090 Ti clocks closer to 2800 MHz.
With all these features enabled, AD102 theoretical performance can reach 103 teraflops with FP32 workloads, 826 teraflops with FP16 workloads with Tensor Cores, and 1652 teraflops with Tensor Cores in FP8 mode. That’s a massive 25% performance boost compared to the RTX 4090.
Of course, these benefits are only realized in GPU-limited scenarios, so probably not in 1080p or 1440p games. Heavy computational applications may also benefit. Combining more L2 cache capacity, additional he GDDR6X bandwidth, and more cores and clocks could result in visible improvements.
| RTX 4090 Ti (Full AD102) | RTX4090 | RTX 3090 Ti | |
| process | TSMC 4N | TSMC 4N | Samsung 8N |
| transistor | 76.3B | 76.3B | 28.3 |
| SMS | 144 | 128 | 84 |
| GPU core | 18432 | 16384 | 10752 |
| Tensor cores | 576 | 512 | 336 |
| Ray Tracing Core | 144 | 128 | 84 |
| boost clock | 2800MHz??? | 2520MHz | 1860MHz |
| VRAM speed | 24 Gbps? | 21Gbps | 21Gbps |
| VRAM | 24GB | 24GB | 24GB |
| bus width | 384 | 384 | 384 |
| memory bandwidth | 1152GB/s | 1008GB/s | 1008GB/s |
| L2 cache capacity | 96MB | 72MB | none |
| ROP | 192 | 176 | 112 |
| TMU | 576 | 512 | 336 |
| TFLOPS FP32 | 103.2 | 82.6 | 40 |
| TFLOPS FP16 | 826 | 661 | none |
| TDP | 600W?? | 450W | 450W |
When is the RTX 3090 Ti Coming?
It looks like Nvidia has a lot of performance headroom left on their GA102 die and could theoretically make an RTX 4090 Ti that could smoke an RTX 4090. It definitely costs a lot more than the RTX 4090 and consumes more power. , but it can be done.
This all depends on how far Nvidia wants to push the GA102 die, and almost certainly depends on how close AMD can get to Nvidia’s performance with its upcoming RDNA 3 chips. Yields also play a role, but I doubt if these will ever become mass-produced parts.
Nvidia could always add some or all of these enhancements to the RTX 4090 Ti if desired. We didn’t get our hands on the RTX 3090 Ti until 18 months after the RTX 3090’s debut, but there were a number of complicating factors at work. His 2023 update for the RTX 40 series will likely come about nine to 12 months after the initial salvo.
There’s also the rare chance that Nvidia will skip the RTX 4090 Ti entirely in favor of the new Titan variant, but I doubt it will. tends to be significantly reduced.
