Nvidia will release the Ada Lovelace architecture in October 2022, starting with the GeForce RTX 4090 and following up with additional variants. Prices have increased compared to previous generations and are higher than others, but these are some of the best graphics cards available today and the 4090 takes the top spot in the GPU benchmark tier. occupies.
But we knew from the start that Nvidia was holding back, leaving room for a potential RTX 4090 Ti or the new Titan RTX Ada. Here’s what we know about the future uber-GPUs of the Ada generation.
Nvidia is currently producing three desktop Ada GPUs namely AD102, AD103 and AD104 which are used in the RTX 4090, RTX 4080 and RTX 4070 Ti respectively. There are also Mobile RTX 40 series GPUs currently in production, ranging from the entry-level RTX 4050 Mobile to the RTX 4090, adding smaller AD106 and AD107 chips to the mix, but the Mobile 4090 uses the same chip. Note that As a desktop 4080. The RTX 4090 Ti and Titan RTX Ada would naturally be at the top of the stack, depending on which direction Nvidia goes, so they’ll also use the AD102 chip.
Currently, the RTX 4090 only uses a portion of the full AD102 GPU for all its performance and power. Certainly a big part, with 128 of the 144 possible Streaming Multiprocessors (SM) enabled. That’s 89% of the GPU cores. However, Nvidia disabled 1MB of L2 cache on each of his 12 32-bit memory interfaces and used 75% of the maximum L2 cache. Turn on the missing pieces, perhaps just bump up the clock speeds and power consumption a bit and you have a new, even faster RTX 4090 Ti.
Make no mistake, Nvidia already has a real AD102 chip that is basically fully functional. For example, his Nvidia RTX 6000 Ada Generation for Pros has 142 out of 144 SMs turned on, plus a full 96MB of L2 cache. Aside from the power consumption being capped at a “minor” 300W, this means the card usually won’t run very fast. Not really a bad approach, but with a little more power and voltage, most Ada chips run near 3.0 GHz. This is effectively what we can expect from the future RTX 4090 Ti and/or Titan RTX Ada.
| graphics card | Titan RTX Ada | RTX 4090 Ti | RTX4090 | RTX4080 | RTX 4070 Ti |
|---|---|---|---|---|---|
| architecture | AD102 | AD102 | AD102 | AD103 | AD104 |
| process technology | TSMC 4N | TSMC 4N | TSMC 4N | TSMC 4N | TSMC 4N |
| Transistor (billion) | 76.3 | 76.3 | 76.3 | 45.9 | 35.8 |
| Die size (mm^2) | 608.4 | 608.4 | 608.4 | 378.6 | 294.5 |
| SMS | 142 | 142 | 128 | 76 | 60 |
| GPU core (shader) | 18176 | 18176 | 16384 | 9728 | 7680 |
| Tensor cores | 568 | 568 | 512 | 304 | 240 |
| RT core | 142 | 142 | 128 | 76 | 60 |
| Boost Clock (MHz) | 2520 | 2625 | 2520 | 2505 | 2610 |
| VRAM Speed (Gbps) | twenty four | twenty four | twenty one | 22.4 | twenty one |
| VRAM (GB) | 48 | twenty four | twenty four | 16 | 12 |
| VRAM bus width | 384 | 384 | 384 | 256 | 192 |
| L2 cache | 72 | 72 | 72 | 64 | 48 |
| ROP | 192 | 192 | 176 | 112 | 80 |
| TMU | 568 | 568 | 512 | 304 | 240 |
| TFLOPS FP32 (Boost) | 91.6 | 95.4 | 82.6 | 48.7 | 40.1 |
| TFLOPS FP16 (FP8) | 733 (1466) | 763 (1527) | 661 (1321) | 390 (780) | 321 (641) |
| Bandwidth (GBps) | 1152 | 1152 | 1008 | 717 | 504 |
| TDP (Watts) | 800 | 600 | 450 | 320 | 285 |
| release date | ? | ? | October 2022 | November 2022 | January 2023 |
| launch price | arm + leg + kidney | arm + leg | $1,599 | $1,199 | $799 |
The specs currently rumored are basically similar to what we’ll see on the RTX 6000 Ada Generation, but with a TGP (Total Graphics Power) double, a slightly higher boost clock for the 4090 Ti and 800W of The power limit has been significantly increased. Titan RTX Ada.
Of course, a higher TGP means these cards have a much higher chance of reaching their rated boost clocks. more likely to exceed Testing of the RTX 4090, 4080, and 4070 Ti cards so far has shown that real-world game boost clocks are often nearly 200 MHz higher than official boost clocks on average.
Everything else follows the core count and boost clock. As for the raw specs, if the rumors are correct, the RTX 4090 Ti will achieve peak throughput of 95.4 teraflops for FP32 work and up to 1.5 petaflops for FP8 compute on Tensor Cores. In both cases, it’s 15% faster than the RTX 4090.
GDDR6X memory is also increased up to 24 Gbps to allow the card to sustain maximum throughput. That’s 14% faster than the 21 Gbps memory used in the RTX 4090. Combined with 33% more L2 cache, the memory subsystem should be well equipped. Continue feeding data to the rest of the GPU.
Titan RTX Ada rumors aren’t as solid and currently point to the same 2520 MHz boost clock as the RTX 4090. However, we would expect such a card given that it also features dual 16-pin power connectors and his 800W TGP. Push the clock even higher (if it really comes out). What we’ve seen with Micron’s latest GDDR6X 24 Gbps module is that even with chips on both sides of the PCB, power and heat from memory aren’t as much of an issue as they were with his previous RTX 3090. suggests.
However, there are still competing powers. Even if the card doesn’t hit 600W/800W all the time, when demanding games he runs at 4K at max settings, it can well exceed 500W. We’ve already seen it happen when manually overclocking on the RTX 4090 card we reviewed, but the 4090 Ti/Titan Ada makes it a standard, not an OC.
You need a good cooling subsystem to handle this much power. In general, we found the RTX 4090’s triple-slot cooling to be decent, but there are images and leaks floating around the web showing quad-slot gold coolers. The gold livery suggests that we likely won’t see the new Titan RTX Ada over the RTX 4090 Ti. I have a silver cooler. It’s too early to say for sure how everything will unfold.
The image of the 4-slot Nvidia RTX 4090 Ti / Titan RTX Ada card looks comically large. Given that even the existing RTX 4090, in many games he’s starting to hit CPU limits below 4K (at least if the game doesn’t have ray tracing). , this definitely feels like Nvidia getting ready to jump sharks, hopefully Nvidia has a very sturdy IO bracket that covers all 4 slots to prevent sagging (cooler’s fixed at the top and bottom). Alternatively, Nvidia may just include some form of kickstand, as many AIB partners have opted to use with his RTX 4090.
All that remains is the price and release date, both of which we don’t know yet. Alternatively, Nvidia could hold out for the summer or fall launch window and let the card serve as a mid-cycle update at the top of the cycle. RTX 40 Series stack. It’s certainly possible that it will be available at GDC (in March) or at the next GTC (also in March). Especially if Nvidia chooses the “prosumers and creators” target market with the new Titan RTX Ada and can keep the RTX 4090 Ti. for later.
The pricing is understandably ridiculous. The hope that future RTX 4090 Ti or Titan RTX Ada cards will be affordable is basically that RTX 4090 cards regularly sell out in the first three months after launch or prices rise above $2,000. And it went out of the window. , we suspect the 4090 Ti or Titan RTX Ada will serve the same users.
If Nvidia adopts the GeForce RTX 4090 Ti branding, the price will likely be set at $1,999. $2,499 or $2,999 seems likely for the new Titan RTX Ada. The Titan offers most of what the RTX 6000 Ada already does (depending on how much VRAM it has), plus TGP increases for higher clocks and less than half the price of true professional GPUs even at 3 Grand Those who are only interested in bragging rights at any price seem to have enough money to pay for what Nvidia demands.
For the rest of us, we can stick with more reasonable GPUs, such as the RTX 3060 Ti replacement. Or perhaps he can wait for the inevitable RTX 50 series coming in late 2024. No matter how fast the future RTX 4090 Ti and/or Titan RTX Ada will be, at some point Nvidia will release his RTX 5080 and RTX 5090 (or possibly other names) with even higher performance than the fastest provide. Ada Lovelace graphics cards can be expected to deliver.
With any luck, the current slump in GPU and PC sales may even result in more aggressive pricing for the change. That’s what happened with his RTX 3080 in 2020, at least in theory. Cryptocurrency mining is disappointing and the pandemic has ruined prices for the next two years and has impacted current price increases.
At least we have a codename for Ada’s successor. Blackwell is probably named after his David Harold Blackwell, a mathematician and statistician specializing in game theory and information theory. However, it’s unclear if Blackwell is just a future Hopper replacement (for the data center market) or if he, like Ampere, spans both data center and consumer markets. Considering Ada Lovelace is only a few months old, it seems a little premature to set our sights on her Blackwell at this stage.
