Recent rumours declare Nvidia’s next-gen Blackwell playing cards will not have a wider reminiscence bus or extra VRAM—aside from the RTX 5090

Nvidia’s keynote speech at Computex 2024 contained nothing about its subsequent technology of GeForce graphics playing cards. So, in the intervening time, we’re left to flick thru the standard sources of leaks and rumours to construct an image of what is coming. The newest of which means that the RTX 5090 might be utilizing a 512-bit extensive reminiscence bus, however the remaining would be the similar as within the present RTX 40-series.

The supply of stated hearsay is Kopite7kimi on X, who has a reasonably good fame for making correct predictions and statements about future developments in GPUs. In a current publish, the leaker set down the reminiscence configurations for the 5 Blackwell GPU variants anticipated to be launched later this yr (although a few of them might not be introduced till 2025).

First up is the GB202, which is able to undoubtedly be used within the RTX 5090 and a raft of professional-grade graphics playing cards. The largest GPUs all the time have the widest reminiscence bus, so that every one these shaders could be saved busy with knowledge—and within the case of the Blackwell monster, it is being claimed that it’s going to sport a 512-bit extensive reminiscence bus and GDDR7 VRAM chips.

If one paired that with Micron’s slowest GDDR7 chips, which run at 28 MT/s, you are taking a look at an mixture bandwidth of 1.8 TB/s or so—roughly 77% extra bandwidth than the RTX 4090. Even when the RTX 5090 ‘solely’ sports activities a 384-bit bus, it will nonetheless have 33% extra bandwidth due to the usage of quicker GDDR7 (the RTX 4090 makes use of 21 MT/s GDDR6X).

Kopite7kimi suggests the opposite GPU variants stay unchanged in regards to the reminiscence bus width, although. The GB203 is 256-bits, the GB205 is 192-bits, and the bottom-end GB206 and GB207 are each 128-bits. That is the identical because the AD103, AD104, AD106, and AD107. Nevertheless, the usage of GDDR7 throughout many of the graphics playing cards that may use these GPUs ought to see a substantial uplift in bandwidth.

It is value noting that the width of a reminiscence bus would not simply impression VRAM bandwidth—it additionally determines how a lot VRAM could be added to the graphics card. In the mean time, all of Micron’s GDDR7 modules are 32-bit extensive and have densities of 16Gb or 8GB, so a 256-bit bus would high out at 16GB.

So apart from the RTX 5090, not one of the forthcoming Blackwell playing cards might be sporting extra VRAM than the present Ada Lovelace fashions, assuming these specs are appropriate. If the successor to the RTX 4090 does have a 512-bit reminiscence bus, we might be taking a look at a graphics card with 32GB of VRAM. Yeah, finest begin saving now…

One thing else that Kopite7kimi recommended within the publish was the interior configuration of the shader blocks in every chip. For instance, the 12*6 for the GB202 refers back to the variety of GPCs (Graphics Processing Clusters) and what number of TPCs (Texture Processing Clusters) are in every GPC.

The AD102 can be a 12*6 configuration, so does this imply the RTX 5090 will not have extra shaders than the RTX 4090? That is a risk, as Nvidia might be trying to enhance general efficiency by simply utilizing larger clock speeds. Nevertheless, the GPC*TPC determine would not inform you what number of SMs (Streaming Multiprocessors) are in every TPC, nor what number of shaders are in every SM.

In Ada Lovelace chips, there are two SMs per TPC and a complete of 128 shaders per SM. Nvidia might be utilizing extra SMs per TPC, extra shaders per SM, or a mixture of each, in Blackwell GPUs. However now we’re within the realm of rampant guessing, so it is best to simply ignore all that till we all know extra.

Nvidia’s market share of discrete GPUs, each add-in playing cards and laptop computer chips, is so massive that it may launch a brand new spherical of graphics processors that are not actually that a lot quicker than their predecessors, and nonetheless promote a bucket load of them.

It could prove that Blackwell GPUs aren’t essentially a lot quicker than Ada Lovelace ones, however due to extra VRAM bandwidth and maybe extra cache, higher AI options, and so forth, the RTX 50-series may nonetheless be notably higher than RTX 40-series playing cards.

Time will inform, after all, however for now, all we will do is speculate on the rumours.