GeForce RTX 6090: what Nvidia's Rubin leak claims (192 SMs, 32 GB GDDR7)
A GeForce RTX 6090 leak points to 192 SMs, a 512-bit bus and 32 GB of GDDR7 on Rubin. Nvidia confirms nothing, and the spec sheet is smaller than it sounds.
Vlad Makarovreviewed and published
Nvidia has not announced a GeForce RTX 6090. There is no product page, no price, no board partner model, no first-party benchmark and no confirmed specification for the generation that would contain one. What exists instead is a Guru3D report assembling a rumor: up to 192 streaming multiprocessors, a 512-bit memory interface and 32 GB of GDDR7, on a Rubin architecture the outlet says may originate in data-center GPU designs before being adapted for gaming. The silicon is described as GR20x-class, the process as TSMC's 3 nm, and the mid-range as 20 GB and 16 GB of GDDR7. Guru3D is explicit that these are "rumors (and they are just that)." That caveat is the most reliable sentence in the story.
A spec sheet that is smaller than it sounds
Set the leaked RTX 6090 next to the RTX 5090 a buyer can actually order, and it does not describe a wider or dramatically fatter GPU. It describes a nearly identical memory configuration and a modest ceiling on compute.
| RTX 5090 (shipping) | RTX 6090 (leaked) | |
|---|---|---|
| Streaming multiprocessors | 170 active of 192 physical | up to 192; enabled count unstated |
| CUDA cores per SM | 128 | not specified in the leak |
| Memory interface | 512-bit | 512-bit |
| Memory | 32 GB GDDR7 | 32 GB GDDR7 |
| Process | not disclosed by Nvidia | rumored TSMC 3 nm |
| Price | $1,999 at launch | none |
Two rows are the same. The 5090's bus is already 512-bit and its capacity is already 32 GB of GDDR7, so the leak enlarges neither. A future card repeating a shipping card's memory configuration is not a regression, but it is not the generational leap the "6090" label implies before a single benchmark exists. The one row that moves is the SM count, and it moves the way Nvidia's halo parts usually do. OC3D's reading of the same leak is blunt: 192 SMs against the 5090's 170 works out to a ceiling roughly 13% higher, and Nvidia "almost never" enables a full die. The promised gain therefore lives mostly in architecture and node efficiency rather than raw size — harder to verify from a spec sheet and easier to overpromise.
The launch window keeps moving
Timing is the second soft number. Early chatter pointed to the second half of 2027; Tom's Hardware reported a prominent leaker placing the consumer Rubin parts, tracked as GR20x, no earlier than 2028, with gaming taking a back seat to data-center allocation. Guru3D's piece reaches the same place from the other direction, noting that memory supply constraints could push availability toward 2028 and that a planned RTX 5000-series refresh appears to have been cancelled. A rumored product with a sliding date is nothing new. What is new is how rarely the coverage says so out loud.
Why 32 GB is not the good news it sounds like
Memory is where the rumor meets a market working against it. TrendForce's latest memory research projects conventional DRAM contract prices rising 10-15% quarter-over-quarter in the fourth quarter of 2026, with graphics DRAM constrained by capacity being reallocated to server products and PC DRAM supply possibly declining into 2027. A 32 GB GDDR7 flagship is a large, expensive order of exactly the component whose price curve is bending the wrong way. With the current top card already gone from first-party retail and trading from $4,399 to a $10,849.96 bundle, the useful question is not whether the 6090 gets 32 GB. It is whether 32 GB at these memory prices leaves any room for a sane launch price at all.
The card this rumor is really measured against
Much of the RTX 6090 traffic arrives attached to a competitor: AMD's rumored Radeon RX 10800 XT, whose leaked performance claims are stated relative to Nvidia's next flagship. GameGPU's figures put the Rubin card 15-20% ahead in raster, at parity in hybrid ray tracing and 35-50% clear in path tracing, which means the AMD leak's own window of advantage closes the moment this unannounced card arrives. We worked through that comparison in full; the point here is narrower. The 6090's leaked specs are not being offered as a product description. They are being used as a benchmark target to make a different rumor look better or worse, and neither claim is well served by that.
What would settle it
Three things, and none of them is another leak. A launch with a price and board partner models a buyer can order. Independent 4K reviews measuring raster, hybrid ray tracing and path tracing on shipping silicon rather than on an SM count. And a memory bill of materials a vendor is willing to state, given how much of this card's cost sits in its GDDR7. Until then, the honest description of the RTX 6090 is a plausible direction — Rubin-derived, 3 nm, wide bus, large frame buffer — and nothing more specific than that. The 192 SMs and 32 GB are the shape of a rumor, not the shape of a card.
