High-end Blackwell will be a 512-bit GPU
It looks like the release of the next generation of Nvidia graphics cards got closer again. Major news about its Blackwell architecture GPUs was leaked on the web, or more accurately on (former) Twitter, yesterday – the specs of the two highest performance models that are due out first: the GeForce RTX 5080 and the new high-end, GeForce RTX 5090. And there’s a surprise: the specs of the high-end have improved compared to previous reports.
These specs were revealed by Kopite7kimi, one of the best sources when it comes to internal Nvidia information. So it’s worth looking at them, even if it’s not yet certain that Nvidia has finalized the parameters. Theoretically, there could still be some changes, the company has often developed redundant configurations in recent years, variously canceling or reviving them and changing plans often at the last minute.
GeForce RTX 5080
The GeForce RTX 5080 will probably be the card that is more relevant to most people (although Nvidia may still set the price above 1000 USD again, which used to be the territory of products for extreme enthusiasts several years ago). The internal designation of this model is reportedly PG144/147-SKU45 and the card uses the GB203 chip (GB203-400-A1 variant), which is the second GPU in the Blackwell generation. So like with the RTX 4080, the second model in the line is yet again not using a stripped down version of the top of the line chip, which is what last happened with the GeForce RTX 3080, instead it uses a physically smaller and weaker GPU.
According to Kopite7kimi, this GPU is supposed to have 10,752 shaders active, which is 84 SM blocks – this presumably is the fully-enabled configuration of the chip. The clock speeds are not yet known, neither for the GPU itself nor for the memory. The GPU uses a 256-bit memory bus and GDDR7 memory type. Earlier indications were that Nvidia could run it at 28.0 GHz effective speed (this would give a bandwidth of 896 GB/s), but Kopite7kimi has not yet confirmed this. The bus width pretty much means that the memory capacity will be 16GB for the GeForce RTX 5080.
As we’ve written before, this graphics card is said to have a higher TDP than today’s GeForce RTX 4080 – namely 400W. But still take this as tentative for now. It is not certain whether the value means the TDP or the so-called maximum power limit, which is usually higher than the typical power consumption. These two different values can sometimes cause confusion. The way Kopite7kimi presents it, 400 W is more likely to be the TDP (i.e. the real power consumption of the whole card), but we can’t know that for sure yet.
According to recent rumors, Nvidia is said to be targeting this model to perform roughly 10% above the GeForce RTX 4090 (but in what games and settings, we don’t know).
Tip: New Nvidia GPUs: Even the GeForce RTX 5080 beats the RTX 4090 by 10%
GeForce RTX 5090: 512-bit GPU
The highest model is reportedly internally designated PG144/145-SKU30 and is based on the GB202 chip, the largest GPU of the Blackwell generation, in the GB202-300-A1 version. In this case, it’s a cut down version. The RTX 5090 graphics card is said to have 21,760 shaders active, which is 170 SM blocks (the full version of the GPU may have as many as 192 SMs, according to previous rumors). Again, we don’t know the clock speeds yet. There have been reports that already the base clock speed is expected to go up to 2.9 GHz, but so far they are unconfirmed.
There is a major change in the memory specs – previous information stated that Nvidia would only use 448 bits of the supposedly 512-bit memory bus, so the capacity would be 28GB. But Kopite7kimi is now stating that in the end, the memory will be full-size with 512 bits of width (which was not seen in a gaming graphics card since the days of the Hawaii GPU 10 years ago). This means that at 28.0 GHz, the bandwidth would be 1792 GB/s (1.75 TB/s). And the capacity would be 32 GB GDDR7.
However, the performance is said to be associated with high power consumption. According to Kopite7kimi, this card is supposed to have a TDP(?) of as much as 600W. This is at the very limit of what 12+4pin 12VHPWR cables (or the newer revision called 12V2x6) allow.
And as we know, they do not have the same safety margin as classic 8pins, so getting above this limit when overclocking could be dangerous. Recently, there was a report that GeForce RTX 5090 could in fact use a power supply with two 12+4pin 12V2x6 connectors for this reason.
Tip: New generation of GeForce graphics cards said to be powered by two 12+4pins
We don’t yet know Nvidia’s intended pricing, and there’s no certain release date. Throughout this year, there have been alternating reports that these graphics cards won’t come out until early 2025, or that they will instead come out this fall. The last reports from around the summer break period said that it was more likely to wait until the first quarter of next year. But it’s probably still not out of the question that Nvidia will pull in the release to still make it in time for the busy shopping period ahead of Christmas this year.
Sources: Kopite7kimi (1, 2), VideoCardz
English translation and edit by Jozef Dudáš
⠀
