We may earn a commission when you buy through links in our articles. Learn more.

Nvidia’s new Blackwell GPU specs just leaked, including RTX 5090

With TPCs, GPCs, and more divulged in this leak, we're getting an ever surer picture of how GeForce RTX 5000 series cards will look.

nvidia blackwell gb202 gpus specs leak

A significant Nvidia Blackwell specs leak purportedly shows the exact number of TPCs and GPCs for each new Nvidia gaming GPU, as well as the memory bus and GDDR configuration attached to each model. This doesn’t give us the exact specs of future RTX 50 graphics cards yet, but provides our closest look yet at what sort of power Nvidia’s upcoming graphics chips should be able to muster.

We fully expect the high-end of our best graphics card guide to be dominated by Nvidia‘s upcoming 5000 series graphics cards, as AMD isn’t expected to compete with a rival RDNA 4 graphics card of this power. However, there’s still plenty to find out about Nvidia’s next offerings before those sort of recommendations can be made.

nvidia blackwell gb202 gpus specs leak tweet

The new Nvidia Blackwell spec leak comes via regular X/Twitter-based leaker @kopite7kimi who simply put forth the following list of Nvidia GPU specs:

  • GB202 12*8 512-bit GDDR7
  • GB203 7*6 256-bit GDDR7
  • GB205 5*5 192-bit GDDR7
  • GB206 3*6 128-bit GDDR7
  • GB207 2*5 128-bit GDDR6

These GB numbers refer to the codenames for Nvidia’s GPUs, with the B of the codenames referencing the Blackwell architecture of the GPUs. In a similar vein, current-generation GPUs based on the Ada Lovelace architecture have codenames ranging from AD102 to AD107. As such, the list provided here seemingly corresponds to the full range of chips, from a flagship GB202 GPU that we’d expect to be used an in RTX 5090 class of card, all the way down to the GB207 that we’d expect to be used in RTX 5060-class cards.

The next two numbers on this list are the GPC and TPC configurations of the GPU. GPCs, or graphics processing clusters, are the largest overarching subsections that Nvidia defines in its GPUs. Within each GPC is a whole host of repeated graphics processing parts.

One of those parts is called a TPC or texture processing cluster, which again defines a smaller subsection of the GPU that contains several of the same smaller processing components. You can see the GPC/TPC relationship for the current AD102 GPU in the image below.

So, from the above list we can see that GB202 will reportedly have 12 GPCs, each containing eight TPCs. Now, at this point, we have to make an assumption about what is contained in those TPCs. If we assume that Blackwell will have the same configuration of components inside each TPC as is used in Ada Lovelace then that means the GB202 will contain a total of 192 streaming multiprocessors (SMs). That compares to 144 SMs for AD102.

Extrapolating the specs of the other GPUs in a similar manner gives us the following specs from this leak:

GPU Architecture Graphics cards GPCs TPCs SMs Mem bus Mem type
GB202 Blackwell TBD 12 8 192 512 GDDR7
GB203 Blackwell TBD 7 6 84 256 GDDR7
GB205 Blackwell TBD 5 5 50 192 GDDR7
GB206 Blackwell TBD 3 6 36 128 GDDR7
GB207 Blackwell TBD 2 5 20 128 GDDR7
AD102 Ada Lovelace RTX 4090 12 6 144 384 GDDR6X
AD103 Ada Lovelace RTX 4080 Super, RTX 4080, RTX 4070 Ti Super 7 6+4 84 256 GDDR6X
AD104 Ada Lovelace RTX 4070 Ti , RTX 4070 5 6 60 192 GDDR6X
AD106 Ada Lovelace RTX 4060 Ti 3 6 38 128 GDDR6
AD107 Ada Lovelace RTX 4060 2 4 24 128 GDDR6

From this we can see that, as we’d expect, Blackwell is set to offer increases in the sheer number of raw graphics processing components for some GPUs, but not others. That means the step up from lower-tier cards to high-tier cards could be even bigger with Blackwell than Ada.

Also notable is that this leak suggests Nvidia will use an exact number of TPCs per GPC for each GPU, whereas this sometimes varies with Ada Lovelace GPUs. For AD103, for instance, one GPC contains only four TPCs, not six like the rest of the GPCs. As such, we must doubly emphasize that the above are estimated SM numbers, so only limited conclusions can be drawn right now.

It’s also worth keeping in mind that these GPU numbers often don’t correspond with the configurations we see used in graphics cards. That’s because the GPUs are defined according to the maximum potential features they’re designed to include.

However, when it comes to manufacturing, errors in the process mean that parts of the GPU often have to be deactivated to make the chip work. That’s why the RTX 4090, for instance, has only 128 SMs when its AD102 GPU can theoretically have up to 144.

So, it’s clearly too early to say exactly what the specs of a future RTX 5080 might be, but we’re at least one step closer if we can believe @kopite7kimi.

For more of our thoughts on what we definitely do know about the latest graphics cards, see why we think the RTX 4070 Super and AMD Radeon RX 7900 GRE are the best graphics cards around right now for most gamers.