Next-gen Nvidia RTX 5000 GPU rumored to use GDDR7 memory on a 384-bit bus

Daniel Sims

Posts: 1,376   +43
Staff
Rumor mill: Recent developments in the DRAM industry suggest GDDR7 will likely be ready for primetime when the next generation of enthusiast-class graphics cards come to market. However, the amount of memory the upcoming flagships will feature remains uncertain, and the information available to trusted leakers keeps on changing.

Well-known leaker Kopite7kimi recently retracted an earlier claim about the memory for the top-end model of Nvidia's upcoming Blackwell lineup, likely to be named GeForce RTX 5090. The latest information suggests it will still be among the first products to shift to GDDR7 VRAM, but it will probably feature the same amount of memory as the current model.

A September leak suggested Blackwell's flagship GPU, the GB202, would feature a 512-bit memory bus, indicating a product with 32GB of VRAM. Kopite7kimi now admits he probably miscalculated the number based on the specifications of the Ada Lovelace architecture. The revised prediction equals the RTX 4090's 384-bit bus, meaning 24GB of video memory.

However, Kopite7kimi maintains that VRAM will be GDDR7, which should provide a significant speed boost compared to the GDDR6X memory that Nvidia's high-end GPUs currently use. Information regarding RAM speeds hasn't emerged yet, but early GDDR7 products could reach 32 Gb/s, a significant boost over GDDR6X's 22-23Gb/s. Further revisions could eventually push GDDR7 to 36 Gb/s – a 50 percent leap over GDDR6.

Prior information on GB202 suggested the RTX 5090 would have 50 percent more cores, 52 percent more memory bandwidth, 78 percent more L2 cache, and a 15 percent higher frequency than the RTX 4090 for a 1.7x performance improvement. This would indicate roughly 24,000 CUDA cores, a 2.9 GHz boost clock, and 128MB of L2 cache.

Kopite7kimi believes GB203 will likely appear in some high-end Blackwell cards and will use a 256-bit memory bus, the same as the 16GB RTX 4080. It's yet unclear if it will be GDDR7, GDDR6X, or possibly the upcoming GDDR6W, which could be another option for next-gen GPUs. Current rumors don't say whether all Blackwell cards will keep the same amount of memory as their Lovelace counterparts, which would be disappointing for the mainstream tiers.

The first GeForce RTX 5000 GPUs are expected to launch sometime in 2024 at the earliest, alongside the first products to arrive with GDDR7. However, we've also seen rumors and a leaked Nvidia roadmap that points to a possible delay into 2025.

Permalink to story.

 
The biggest news about the 5000 series is that the 4000 and rest of 3000 series will get way cheaper and payable. I don’t need to play games on ultra nor at ultra high frame rates, nor I buy unoptimized games. So older generations will be enough for me, I don’t work to be able to play games…
 
The 5090 is going to be a monster. As much as I don't like nVidias mainstream lineup, the 4090's performance is mind boggling. Lets hope AMD can actually make a top teir card that doesn't just trade blows with nVidia's second best next round.
 
The 5090 is going to be a monster. As much as I don't like nVidias mainstream lineup, the 4090's performance is mind boggling. Lets hope AMD can actually make a top teir card that doesn't just trade blows with nVidia's second best next round.
We can all hope, but I dont count on it. Outside of the rumors of rDNA4 not having a high end chip, rDNA2 was the first time since Hawaii in 2013 that AMD contended for the crown. Nearly a decade apart. The 7900xtx is a good GPU but it doesnt touch the 4090.
 
The 5090 is going to be a monster. As much as I don't like nVidias mainstream lineup, the 4090's performance is mind boggling. Lets hope AMD can actually make a top teir card that doesn't just trade blows with nVidia's second best next round.
3090 was mindboggling monster too, and look where it is now.
 
We can all hope, but I dont count on it. Outside of the rumors of rDNA4 not having a high end chip, rDNA2 was the first time since Hawaii in 2013 that AMD contended for the crown. Nearly a decade apart. The 7900xtx is a good GPU but it doesnt touch the 4090.

- In fairness Fiji was actually fairly competitive with GM102 at stock settings.

Its just that the 980Ti overclocked so outrageously well (as well, ironically, as AMD's HD7970 did) that all the aftermarket parts were stacking 20% OCs on the founders edition part and leaving Fiji in the dust.
 
I can understand AMD's point of view - having the fastest card with the highest price tag is'nt where the money is. It's within low to mid / mainstream of consumers - people who seek best bang for the buck.

And designing a high end chip is expensive too. Esp with Nvidia - since it's still working on monolithic dies on which the failure rate on wafers is far higher then AMD is doing with it's computing / graphics chiplets.

It's a interesting race. That's for sure.
 
Back