RTX 5070 with an RTX 5080 chip: what Nvidia might really be preparing

RTX 5070 GB203 Nvidia Blackwell RTX 5080

A leak announces a RTX 5070 equipped with the large die GB203 from the RTX 5070 Ti and RTX 5080. Performance, heat, **GPU-Z**: here’s what that would change.

A RTX 5070 with the silicon of the RTX 5080?

Nvidia would be preparing a second version of the GeForce RTX 5070. Same name. Same memory. But under the cooling, a different GPU: the small GB205 would make way for the GB203, the die found in the RTX 5070 Ti and RTX 5080.

The info comes from the leaker Hongxing2020, who posted a terse sentence on X: "The 5070 uses the GB203 core," followed by an announcement of a new driver for the next day. VideoCardz, TweakTown, and Guru3D picked up the leak in the wake. As of September 10, 2026, Nvidia has announced nothing. Its official sheets do not mention this variant. We are therefore facing a leak, not to a press release .

Click to enlarge

If you want to refresh your memory on the entire range, our file on the RTX 50 Blackwell details the characteristics and positioning of the RTX 5070. And to understand where this famous GB203 comes from, our article on the RTX 5080 reviews its technical sheet and its role as a high-end 4K card.

One point to keep in mind right now: a GB203 in a RTX 5070 does not turn it into a RTX 5080. The number of active units would remain, according to the leak, that of a RTX 5070. Neither more, nor less.

What the leak says, and only the leak

The original message provides no release date, no partner brand, and no PCB photo. No promise of additional memory or increased performance either. The only claim: a RTX 5070 would use the GB203, and this change would be linked to a driver expected on September 9.

The GeForce Game Ready driver 616.92 has indeed arrived at this time. Its public notes mention new games and DLSS updates. Nothing in it about a revised RTX 5070. The connection between this driver and the new GPU remains very speculative.

And this caution is important. On graphics cards, a new PCI identifier or a line added to a driver can correspond to many things: support for a new sub-model, preparation of a card reserved for certain manufacturers, simple recognition of a hardware revision. As long as a card has not been photographed, tested, and identified by a tool like GPU-Z, it is impossible to speak of a confirmed launch.

That said, the idea holds up. Nvidia already sells several cards from the same large die with some units disabled. The GB203 is used for the RTX 5070 Ti and RTX 5080. Dropping it down a notch in the lineup would be nothing new.

Why would Nvidia put a big GB203 in a RTX 5070

Manufacturing GPUs is not about printing perfectly flawless chips. On a wafer, some dies come out with some CUDA cores, cache, or controllers out of service. A GB203 chip that does not meet the specifications of an RTX 5080 can still become an RTX 5070 Ti. And if it has too many defects for this second model but enough functional blocks to reach 6 144 CUDA cores , Nvidia can still sell it as RTX 5070.

This is what is called the die harvesting , the recovery of dies. The faulty units are disabled, the necessary blocks remain active, and the chip is classified in a slower reference. The principle has existed for decades. It is even the most logical way to avoid throwing away a piece of silicon that was expensive to produce.

On the Nvidia side, the calculation is clear: improve the yield of the GB203, sell imperfect chips, and feed the stock of RTX 5070 without relying solely on the GB205. On the buyer's side, the benefit is much less clear-cut. You are still buying an RTX 5070 with the limitations of an RTX 5070.

This practice can also help card manufacturers maintain their volumes when the availability of dies varies. A commercial reference can then exist with two different physical GPUs, without a new box or visible suffix. And this is where the subject deserves attention: two cards bearing the same name could feature different silicon.

GB205 against GB203: the real differences

The current RTX 5070 is based on a GB205 and activates 6,144 CUDA cores. 12 GB of GDDR7, 192-bit bus, thermal envelope announced at 250 W. The RTX 5080, on the other hand, goes up to 10,752 CUDA cores, 16 GB on 256 bits and 360 W. The RTX 5070 Ti sits between the two with the GB203, 8,960 active CUDA cores and 300 W.

The variant mentioned by the leak would keep the commercial specs of the RTX 5070: 6,144 cores, 12 GB, 192-bit bus, around 250 W. The change would be located only under the hood. A large portion of the GB203 would be disabled to replicate the configuration of the GB205.

Click to enlarge

The GB203 is physically larger. More surface to dissipate heat to the cold plate, in theory. But this is not a guarantee of lower temperatures. The cooling of the board, the quality of the thermal contact, the applied voltage, and the BIOS matter much more than the size of the die alone. A well-assembled RTX 5070 GB203 could be a bit easier to cool. An entry-level model with a medium heatsink will remain... an entry-level model.

Card / variant Die CUDA Cores Memory Bus Power
Current RTX 5070 GB205 6 144 12 GB GDDR7 192 bits 250 W
Revised RTX 5070 (leak) GB203 disabled 6,144 expected 12 GB GDDR7 expected 192 bits expected Approximately 250 W
RTX 5070 Ti GB203 8 960 16 GB GDDR7 256 bits 300 W
RTX 5080 GB203 10 752 16 GB GDDR7 256 bits 360 W
Swipe to view more

No, it will not be a RTX 5080 hidden at half price

Seeing "GB203" in GPU-Z will probably trigger some pretty wild comparisons on the forums. The die name does not define performance, however. A RTX 5080 exploits many more compute units, a wider memory bus, and 16 GB of GDDR7. The supposed RTX 5070 would keep its 12 GB and 192 bits. No magic driver can reactivate physically defective or locked blocks.

Click to enlarge

In gaming, one should expect almost identical results between a RTX 5070 GB205 and a RTX 5070 GB203. The frequencies may vary slightly from one partner model to another, as is already the case, but not to the point of creating a new tier of performance. The cooling and power limits of the specific model will have more effect than the die change at equal specifications.

There is still a small gray area on the side of computing uses. At the time, some RTX 2060 equipped with a large TU104 had shown a measurable advantage in Blender, while games remained at the same level as the TU106 versions. This is not a prediction for the GB203. It is a precedent that justifies testing the new variant instead of concluding too quickly that it will be strictly identical in all scenarios.

What this changes if you buy a RTX 5070

In the short term, not much. Don't pay more. a card only because a seller's listing, a forum, or a screenshot claims it contains a GB203. Without independent benchmarks, this mention is not worth a penny extra.

The choice must continue to be made based on concrete criteria: final price, cooling quality, noise level, dimensions, warranty, connectivity. The 12 GB of memory remain the most debatable compromise of the RTX 5070. Switching to GB203 does not turn them into 16 GB and does not widen the memory bus.

If two models cost the same price and one is certainly identified as a GB203 version, the info may be fun for hardware enthusiasts. A larger die could offer slightly different thermal behavior or a small bonus in a specific application. But for gaming at 1440p, first take the best card overall, not the name of the silicon on it.

Another thing: Nvidia and its partners could mix revisions according to the batches. The same Asus, MSI, or Gigabyte reference could change GPU during production. This scenario has been seen in the past. The commercial number printed on the box may not necessarily be enough to decide.

How to check the GPU with GPU-Z

The most reliable way remains to install GPU-Z after mounting the card. The software displays the name of the GPU, its revision, its hardware ID, the number of active cores, and the memory bus width. If the variant becomes available on the market, its database will likely be updated in the following days.

Do not rely on a single field during the first days. An old version of GPU-Z may misidentify a recent revision, and a driver may expose an identifier without displaying the detailed name of the die. Update the software, keep a screenshot of the main tab and compare the information with tests published on the same model.

Under Windows, the Device Manager also allows you to retrieve the hardware ID in the properties of the card. It's less readable, but it works to confirm the existence of two variants. Disassembling the cooling would provide the definitive answer thanks to the chip marking, except that it complicates the warranty and makes no sense for an average buyer. GPU-Z will suffice.

Nvidia has already pulled the stunt with the RTX 5060

This rumor does not come from nowhere. RTX 5060 have already been spotted with a larger GB205 instead of the expected GB206. MSI has notably distinguished some cards with a "V1" revision, while other manufacturers have modified their production more discreetly.

We detailed this previous one in our survey on the RTX 5060 equipped with GB205 . The principle is the same: keep the name, the memory and the targeted performance, but use a higher die that is largely disabled. The RTX 5070 with GB203 would be the logical continuation of this recovery strategy.

In the case of the RTX 5060, the chip change did not create a miraculously faster card. This is probably what awaits the RTX 5070. Another piece of silicon, yes. A discounted RTX 5080 in a 5070 box, no.

Should we wait for this new RTX 5070?

No, not if you need a card now and you find a good price. The leak does not provide any availability date in France, no manufacturer has announced a specific model. Waiting several weeks for a probably invisible difference in gameplay is not worth it.

If your purchase is not urgent, monitoring the upcoming batches may make sense. The initial tests will need to measure temperatures, noise, consumption, and computing performance. Not just launching three games and concluding. That's where a possible difference could manifest.

Our reading: this revision is above all a good operation for Nvidia's yields. It can also secure the supply of the RTX 5070 in the medium term. For the player, it do not correct the most common complaint about this card in 2026: its 12 GB of VRAM. As long as the price and memory do not change, the choice remains the same as before.

FAQ on the RTX 5070 GB203

Is the RTX 5070 GB203 official?

No. As of September 10, 2026, Nvidia has not announced anything on this subject. The information comes from a leaker and has been picked up by several hardware media.

Will it have the performance of a RTX 5080?

No. The die would be the same, but with many more disabled units. The memory would remain limited to 12 GB on 192 bits, compared to 16 GB on 256 bits for the RTX 5080.

Will the GB203 version heat less?

Possible, but far from guaranteed. A larger die distributes heat better, but the radiator, ventilation, voltage, and assembly quality remain the most determining factors.

How to know which version I received?

GPU-Z should allow identifying the GPU and its revision once its database is updated. The Windows hardware identifier can serve as a second verification.

Should we pay more for the GB203 variant?

No. Without gain measured by independent tests, the GB203 does not justify any extra cost. Rather compare the price, cooling, noise, and warranty of the complete model.

Main sources: publication X of Hongxing2020 , taken over by VideoCardz and TweakTown ; official technical sheets Nvidia for GeForce RTX 5070 and RTX 5070 Ti and of the RTX 5080 ; GeForce Game Ready driver 616.92. All information related to the GB203 variant remains unconfirmed by Nvidia as of September 10, 2026.