Nvidia’s GeForce RTX 4090 would possibly look extremely sturdy, and will definitely rank because the quickest choice on our listing of the finest graphics playing cards when it debuts (at the least till AMD’s RDNA 3 GPUs arrive), however the shaved down AD102 die within the RTX 4090 is not near exhibiting off the complete potential of AD102 with all of its cores and cache enabled. This mixed with extra enhancements may trace at a future RTX 4090 Ti that shall be a lot quicker — and even perhaps dearer.
The specs for the Nvidia RTX 40-series and Ada Lovelace GPUs, however these solely present the introduced and rumored playing cards. Nvidia’s full AD102 die comes outfitted with 144 SMs, 18,432 CUDA cores, 96MB of L2 cache, and 192 ROPs. This interprets to 12% extra CUDA cores and a whopping 33% extra L2 cache capability in comparison with the RTX 4090 we now have right now. The totally enabled AD102 die additionally packs 9% extra ROPS and 12% extra Texture Mapping Models as properly, because of the extra SMs.
However that is not all that could possibly be performed for the long run 4090 Ti. Micron has new 24Gbps GDDR6X reminiscence modules within the works, one other 14% enhance over the RTX 4090’s 21Gbps modules, and nonetheless quicker than the RTX 4080 16GB’s 22.4 Gbps modules that Nvidia claims are the quickest on this planet proper now. That will push the hypothetical (however very possible) RTX 4090 Ti as much as 1152 MB/s of bandwidth.
However quicker reminiscence would include greater energy consumption, and we suspect that Nvidia is critically holding again AD102’s full clock pace and energy potential as properly. All these rumors of 600W RTX 40-series graphics playing cards? We all know Nvidia has efficiently overclocked RTX 4090 to greater than 3.0GHz, and that might undoubtedly push up energy use.
It seems just like the Ada structure and TSMC’s 4N course of have loads of headroom remaining past the RTX 4090’s 2520 MHz enhance frequency. As soon as the method matures a bit extra, and if Nvidia is prepared to extend the ability limits, we would not be stunned to see a RTX 4090 Ti clock at nearer to 2800 MHz.
The theoretical efficiency of AD102 with all these bells and whistles enabled may attain a whopping 103 teraflops in FP32 workloads, and 826 teraflops in FP16 workloads with the Tensor cores, and 1652 teraflops with the Tensor cores in FP8 mode. That will be an enormous 25% efficiency leap compared to the RTX 4090.
These good points would solely be realized in GPU restricted eventualities, after all, so most likely not 1080p or 1440p gaming. Heavy compute functions would additionally possible profit. The mix of extra L2 cache capability, extra GDDR6X bandwidth, and extra cores and clocks may lead to tangible enhancements.
| RTX 4090 Ti (Full AD102) | RTX 4090 | RTX 3090 Ti | |
| Course of | TSMC 4N | TSMC 4N | Samsung 8N |
| Transistors | 76.3B | 76.3B | 28.3 |
| SMs | 144 | 128 | 84 |
| GPU Cores | 18432 | 16384 | 10752 |
| Tensor Cores | 576 | 512 | 336 |
| Ray Tracing Cores | 144 | 128 | 84 |
| Increase Clock | 2800MHz??? | 2520MHz | 1860MHz |
| VRAM Velocity | 24 Gbps? | 21 Gbps | 21Gbps |
| VRAM | 24GB | 24GB | 24GB |
| Bus Width | 384 | 384 | 384 |
| Reminiscence Bandwidth | 1152GB/s | 1008GB/s | 1008GB/s |
| L2 Cache Capability | 96MB | 72MB | None |
| ROPs | 192 | 176 | 112 |
| TMU | 576 | 512 | 336 |
| TFLOPS FP32 | 103.2 | 82.6 | 40 |
| TFLOPS FP16 | 826 | 661 | N/A |
| TDP | 600W?? | 450W | 450W |
When Will We See an RTX 3090 Ti?
It seems Nvidia has numerous efficiency headroom remaining with its GA102 die, with the potential to create a RTX 4090 Ti that might theoretically smoke the RTX 4090. It will definitely price much more cash, and devour far more energy than a RTX 4090, however it may be performed.
All of it will rely on how exhausting Nvidia needs to push its GA102 die, and that can nearly definitely rely on how shut AMD can come to matching Nvidia’s efficiency with the upcoming RDNA 3 chips. Yields on totally purposeful AD102 GPUs would additionally play a task, although it is uncertain these could be excessive quantity components.
Nvidia may add some or all of those enhancements to an RTX 4090 Ti any time it feels the necessity. We did not get the RTX 3090 Ti till 18 months after the RTX 3090 debut, however there have been numerous compounding elements in play. Extra possible is we’ll see a 2023 refresh of the RTX 40-series a while round 9 months to 12 months after the preliminary salvo.
There’s additionally the uncommon probability Nvidia may skip the RTX 4090 Ti utterly in favor of a brand new Titan variant, however we doubt that would be the case. Titan playing cards have a tendency to chop into the profitable RTX A-series skilled card income an excessive amount of.