Has the Nvidia RTX 4090 Ti been discontinued before it was even announced?

[midian182]

Rumor mill: The long-running rumors that Nvidia plans to release a massive RTX 4090 Ti graphics card could stop after a leaker claimed it has been canceled. Assuming Nvidia really did intend to release a Lovelace product above the RTX 4090, killing it off at a time when even the "mid-range" cards cost a fortune was probably a sensible move.

Rumors of a monstrous, quad-slot RTX 4090 Ti arrived in January. The card was alleged to pack 24GB of 24Gbps memory and a slightly cut-down version of the AD102 with 18,176 Cuda cores, 192 ROPs, 568 TMUs, a 96MB L2 cache, and a 384-bit memory bus. It was also said to have a TGP of 600W.

The AD102 configuration is similar to the one used in the RTX 6000 Ada Generation card aimed at the professional and enterprise market, though that $7,000+ product packs 48GB of GDDR6.

But the same prolific leaker who revealed those specs, kopite7kimi, now says that Nvidia has decided to cancel the card. Given that the RTX 4090 carries a $1,600 price tag, it's unlikely that many people would have been lining up to purchase something even more expensive in these economically uncertain times.

While Nvidia is apparently canceling a potential flagship card, kopite7kimi also states that it is preparing some new variants further down the product stack. The leaker says new versions of the RTX 4070 and RTX 4060 are being released with some higher-end GPUs: a cut-down AD103 and an AD106.

The AD103, currently used in the desktop RTX 4080 and RTX 4090 mobile, could give a spec boost to the RTX 4070, which uses the AD104 GPU. We could see an upgrade from a 192-bit bus to 256-bit, and maybe a version with 16GB of VRAM.

As for the RTX 4060, which uses the AD107 GPU, the new variant could also offer a higher configuration using the AD106, found in both the RTX 4060 Ti 8GB and 16GB.

There are a lot of unknowns here, and even if Nvidia does use cut-down versions of more powerful GPUs in new card variants, it's something the company has done before without noticeably improving their performance.

Going back to the RTX 4090 Ti, it does look as if Nvidia is going to keep the RTX 4090 at the top of the consumer product line – we also heard of a 900W Lovelace Titan being canceled last year as it was melting PSUs. Team Green could always release the 4090 Ti at a later date, of course, but then the RTX 5000 cards are rumored to be launching next year with massive performance increases.

Permalink to story.

https://www.techspot.com/news/99569-nvidia-rtx-4090-ti-discontinued-before-even-launched.html

 

[Theinsanegamer]

Maybe they figured out, finally, that their power connector just cant handle this much power draw?

 

[QuaZulu]

"but then the RTX 5000 cards are rumored to be launching next year with massive performance increases."

Whatever. Every generation, whether CPU or GPU, we hear the same hype train: "massive performance increase inbound." The companies seem to love when we buy that smoke-blowing rhetoric so that they can actually present us with massive price increases with minimal performance increases.

 

[takaozo]

"Rumors of a monstrous, quad-slot RTX 4090 Ti arrived in January. The card was alleged to pack 24GB of 24Gbps memory and a slightly cut-down version of the AD102 with 18,176 Cuda cores, 192 ROPs, 568 TMUs, a 96MB L2 cache, and a 384-bit memory bus. It was also said to have a TGP of 600W."

TGP refers to GPU chip power draw and TBP refers to total board power (memory included)?
That means more like 670-700W TBP, right?

 

[punctualturn]

I can see all the Tubers/Influencers, err, sorry, reviewers, crying because they wont receive a new free gpu from their favorite evil corporation.

 

[HardReset]

"Rumors of a monstrous, quad-slot RTX 4090 Ti arrived in January. The card was alleged to pack 24GB of 24Gbps memory and a slightly cut-down version of the AD102 with 18,176 Cuda cores, 192 ROPs, 568 TMUs, a 96MB L2 cache, and a 384-bit memory bus. It was also said to have a TGP of 600W."

TGP refers to GPU chip power draw and TBP refers to total board power (memory included)?
That means more like 670-700W TBP, right?

TGP is supposed to be total power for entire card, memory included. Of course, that 600W might be an estimate.

 

[EdmondRC]

According to Nvidia's documents, 5000 isn't coming until 2025.

 

[neeyik]

Of course, that 600W might be an estimate.

It also might be complete nonsense. The RTX 6000 Ada Generation has the same specs as the purported 4090 Ti, bar clocks, and its TGP is 300W. Of course, its base clock is below the normal 4090's and by quite a margin, (0.915 vs 2.235 GHz) but the boost is virtually the same (2.505 vs 2.520 GHz).

The use of a dual-slot blower for cooling forces the need for a low base clock and TGP, but would higher clocks push the power draw up by a further 300W? Possible, of course, but It seems unlikely.

 

[Thanthan]

It also might be complete nonsense. The RTX 6000 Ada Generation has the same specs as the purported 4090 Ti, bar clocks, and its TGP is 300W. Of course, its base clock is below the normal 4090's and by quite a margin, (0.915 vs 2.235 GHz) but the boost is virtually the same (2.505 vs 2.520 GHz).

The use of a dual-slot blower for cooling forces the need for a low base clock and TGP, but would higher clocks push the power draw up by a further 300W? Possible, of course, but It seems unlikely.

Until you consider that it’s not unusual for the pro cards to consume much less power, since they’re binned for exactly that, and also kept much closer to the ‘good’ end of the voltage/frequency curve.

Power draw increases geometrically, not linearly, so even a 10 or 20% increase to clocks could double a cards power draw depending on how close to the tipping point on the curve you are. And until someone actually has both cards available to them and checks their in benchmark frequencies, there’s no way of telling how actual card behaviour compares.

 

[kiwigraeme]

I can see all the Tubers/Influencers, err, sorry, reviewers, crying because they wont receive a new free gpu from their favorite evil corporation.

Was it last year Steve called out Nvidia shenanigans and other serious reviewers backed him.
Apparently they didn't send out the 4060ti to reviewers - so not all rosy
Think even Steve got hammered on in 4060 qualified review - so had to qualify it more

People watching these reviews want them to be stood up for - Don't watch twitch , streamers so can not comment on that

Given that would be interested to know - how many discrete buyers of GPUs are well informed - as oppose to listening to salesperson in best buy

 

[ScottSoapbox]

Whatever. I need 16 GB VRAM and +30% performance from my 3080 for under $800 or +50% performance for under $1000. Until then I’ll just wait with the 3080 or go outside and ride motorcycles.

 

[godrilla]

According to Nvidia's documents, 5000 isn't coming until 2025.

Will AMD take advantage of this situation or will they continue to give free advertisements to Nvidia until then!


This move of canceling the Titan/4090ti is likely because Nvidia doesn't want to disrupt the fragile pricing structure too much. Pressure from the top can cause the whole house to tumble down. If Nvidia makes great margins across the stack by releasing the niche of the niche 4090ti, they might not be able to offset their margins lowered from people selling their 4090 to upgrade to titan, which will put pressure on the already lackluster sale of the rest of the lineup.

 

[Eldritch]

"but then the RTX 5000 cards are rumored to be launching next year with massive performance increases."

Whatever. Every generation, whether CPU or GPU, we hear the same hype train: "massive performance increase inbound." The companies seem to love when we buy that smoke-blowing rhetoric so that they can actually present us with massive price increases with minimal performance increases.

Correct. Would like to add : performance increase in which segment? Top tier? What about rest of the lineup esp the one that sells most? How much of "massive performance increase" is there in 4060 as compared to 3060?
All hype. No substance. Don't even bother. If you want to buy, then buy. Wait for offers and sales rather than magical next generation that will likely be a damp squib except for top tier.

 

[neeyik]

Power draw increases geometrically, not linearly, so even a 10 or 20% increase to clocks could double a cards power draw depending on how close to the tipping point on the curve you are. And until someone actually has both cards available to them and checks their in benchmark frequencies, there’s no way of telling how actual card behaviour compares.

Unless there are changes to the core voltage across the clock domains, the relationship between power consumption and the clock frequency is logarithmic, not geometric.



This is some 3DMark Speed Way 4K testing (which has a good utilization of almost all aspects of a GPU) I did a while back on a 4070 Ti, with the Base and Boost clocks changed over as wide a range as possible. The total board power was logged via GPU-z and while it's not the most accurate way of gleaning such figures, they're not wildly out.

The TDP limit was set to 110% (308 W) and all three fans were set to 100% to ensure that the PerfCap Reason was always the core voltage (fixed to the default 1.1 volts), rather than the power or temperature limits. Here one can see that a 65% increase in clock frequency (1840 to 3030 MHz) resulted in a 20% rise in total board power consumption (229 to 274 W).

Base clock values just determine the lowest figure the GPU will drop to when voltage, power, and/or temperature limited -- when not in such conditions, it will run at its boost clock (which itself can go higher than set to in software). Thus, when not restricted by those three metrics, an RTX 6000 Ada Gen card will run at 2505 MHz or higher -- there's nothing special about the chips being used or certainly nothing more out of the ordinary compared to those used in the different AIB 4090 SKUs.

Edit: Apologies for the overly long comment, but my curiosity was piqued regarding how the power limit actually affects clocks in Ada Lovelace. So I repeated the Speed Way testing again, but this time changing the power limit from 40% through to 100% (while locking the temperature limit to 88C).



At all but 110%, the PerfCap Reason was the power limit (voltage limit was the reason at 110%). Now while the AD102 and AD104 are different chips, the former is essentially nothing more than two of the latter, in every respect. The RTX 6000 Ada Gen has a TGP limit of 300W -- around 67% that of the RTX 4090.

And if one looks at that region for the AD104 above, it marks the point where the clocks don't notably increase, as the power limit goes higher. If the AD102 behaves in the same manner, then it would explain that choice of TGP -- it marks the point that's enough power to achieve the desired clock speed. By design, of course, as none of this occurs through circumstance.

 

[Dimitrios]

What's amazing is that in 5-7 or so years we will look back and laugh while he have APU's that match the performance.

I still can't believe the size, weight and power it consumes.

 

[Theinsanegamer]

What's amazing is that in 5-7 or so years we will look back and laugh while he have APU's that match the performance.

I still can't believe the size, weight and power it consumes.

And we'll be laughing at people who dont understand why their APUs still only get 20-30 FPS int he latest games.

 

[Mr Majestyk]

Only the biggest Nvid!ot would even care. I mean who wasn't looking forward to a 600W $4000AU gpu.

 

[DaveBG]

Only the biggest Nvid!ot would even care. I mean who wasn't looking forward to a 600W $4000AU gpu.

LOL wattage is not a problem but yeah this does not cost even half of that... insanely overpriced.

 

[godrilla]

Only the biggest Nvid!ot would even care. I mean who wasn't looking forward to a 600W $4000AU gpu.

I knew it wasn't going to be cheap but I didn't realize they had publicized the price on a canceled product. Outside of gaming the 3090ti/3090 scaled really well from what I remembered in 4 way sli nvlinked which if you bought when they were $999 would have given you almost 2x the performance in perfect scaling (of 4090ti) in compute Cuda optimized applications. 4x 3090tis last year sold for around that $4k mark as well. Even 2 3090tis tweaked for efficiency in sli would probably have gotten you similar performance to this 4090ti but for $2000 last year this time. Take home message Nvidia likely didn't want to compete with itself, dedicate more silicon to ai, and possibly cause a price adjustment in the current lineup from pressure from the top; which is probably the last thing they want. Nvidia likely realized they can't solely rely on brand loyalists in this market by looking at the bigger picture.
While the 3090ti did launch at $2000 and within months dropped by half the current market ( which was probably caused by 2 things . 1) end of crypto mining craze 2) rumors of 4090 at the time haulted potential buyers.

4 way 3090s in sli nvlinked 300 watt limit