Nvidia RTX 50 series Blackwell GPUs tipped to use 28Gbps GDDR7 memory with 512-bit interface

[DragonSlayer101]

Rumor mill: The release of Nvidia's RTX 50-series GPUs is still a long way off, with rumors suggesting that it could happen either in late 2024 or early 2025. However, that isn't stopping the rumor mill from churning out reports about their specs, architecture, performance, and power requirements.

According to the latest post by prolific tipster @kopite7kimi, the RTX 50 "Blackwell" graphics cards will be based on the GB202 and GB203 GPUs. The former is expected to be the full fat flagship, while the latter will be the cut-down version with half the number of cores.

In comparison, the current-gen AD102 GPU underpinning the RTX 4090 lineup comes with 80 percent more cores than the AD103 that powers the RTX 4080 and lower. However, the tipster was unsure about whether the GB202 will use a multi-chip package.

I think GB203 is half of GB202, just like GB102 and GB100. But I don't know if GB202 has a multi chip package.

– kopite7kimi (@kopite7kimi) March 11, 2024

Another rumor surrounding the upcoming graphics cards is that they might feature a 512-bit memory interface compared to the 384-bit interface found on the flagship AD102 Ada GPU. That's in direct contrast to a rumor from last week, which claimed that the GB200 gaming GPUs will retain the exact same memory interface as Ada, suggesting they will only pack a 384-bit bus at the maximum.

Finally, the first set of Blackwell gaming cards are also tipped to ship with 28Gbps GDDR7 memory, which is a tad lower than the 32Gbps maximum capacity of the GDDR7 chips. However, this will still be substantially faster than previous generation graphics memory, as GDDR6 only offered up to 18Gbps bandwidth, while GDDR6X went up to 21Gbps. While Nvidia is expected to stick with 16Gbit densities for the GDDR7 memory chips, there's no word on the memory bus widths of the individual SKUs.

28Gbps.

– kopite7kimi (@kopite7kimi) March 11, 2024

If the latest report turns out to be accurate, it could mean a significant performance boost for the RTX 5090, thanks to the extra cores from the GB202 GPU, the increased memory bandwidth, and the 512-bit bus.

As with all rumors, these need to be taken with a pinch of salt. A lot can still change between now and the eventual launch of the RTX 50-series cards.

Permalink to story.

https://www.techspot.com/news/102211-nvidia-rtx-50-series-blackwell-gpus-tipped-use.html

 

[Disiw]

I don't see why Nvidia should max out bus width while using 28 Gbps memory, when 32 or even 36 is available.

 

[dangh]

Good, I'm looking for a change in this timeframe, so knowing nVidia is planning something I can safely bet AMD will propose some new cards as well.

 

[gigantor21]

The pricing is what I’m worried about.

 

[Disiw]

Good, I'm looking for a change in this timeframe, so knowing nVidia is planning something I can safely bet AMD will propose some new cards as well.

They are. Radeon 8000 series, which won't have any high-end SKUs. They will target low to mid-end, like Radeon 5000 series.

Expect "8700XT" to have 7900XT performance at lower watts. Making 7900XT go EOL. Hopefully 8700XT will be 499-599 tops tho.

Lets see if AMD abandons MCM and go back to monolithic.

 

[yRaz]

I don't see why Nvidia should max out bus width while using 28 Gbps memory, when 32 or even 36 is available.

Some workloads benefit from larger bus widths over just bandwidth. Also, AI is pretty bandwidth heavy as it is so the extra buswidth would help even with the extra memory speed.

 

[Disiw]

The pricing is what I’m worried about.

Expect 1999 dollars for 5090 and 1199 for 5080, if not more for 5080 considering it will beat 4090 with ease and 4090 already beats everything from AMD with ease.

Rumours claim 5090 will get 32GB of GDDR7 running at 28 Gbps on a 512 bit bus. Using TSMC 3nm. Meaning it won't be cheap. It will also be made on the same node as Nvidias AI and Enterprise cards, which are in huge demand. I don't think Nvidia will allocate much space to gaming GPUs before sometime in 2025.

5070 and 5060 series sometime in 2025 is what AMD have to compete with. Nvidia is not going to rush these lower tier series (which is why 4000 series got refreshed) - I don't see AMD touching 5080 before 2026+ (and then 6090/6080 launch)


AMD had a huge oppotunity of regrabbing marketshare, with Nvidias AI focus, and what do they do, release mid-end stuff at best. If Radeon 8000 series don't come with very aggressive pricing, it will be an utter fail.

Also AMD have to improve FSR alot. It is still far worse than DLSS/DLAA with way more shimmering, artifacts and less support in new AAA games.

This is why AMD users hate upscaling, because they only can use FSR, which is mediocre. Even XeSS beats FSR in most games at this point. DLSS/DLAA rules supreme and RTX users don't even have to bother testing other upscalers.

 

[godrilla]

These rumors change on a daily basis.

I don't see why Nvidia should max out bus width while using 28 Gbps memory, when 32 or even 36 is available.

Probably mass production and cost.

 

[Disiw]

These rumors change on a daily basis.
Probably mass production and cost.

Yeah but Kopite said 512 bit today. Most trustworthy leaker.

Usually its cheaper going with smaller bus and faster mem.

However if Nvidia uses 384 bit again, then 32 Gbps should be minimum. 28 Gbps is almost below spec (OC headroom might be huge if 28 Gbps is used)

Nvidia is not forced to do anything tho, they could hold performance crown while using 256 bit at this point, maybe even 192 bit.

192 bit flagship is not happening, but it would be enough to win and would be cheap to make

Medium sized Blackwell chip made at 3nm TSMC + 192-256 bit bus + GDDR7 at 32 Gbps would easily whoop 4090, hence beating everything from AMD which won't have any high-end stuff anytime soon ... RDNA5 is 2026+ and will fight RTX 6000 series.

GDDR7 allows for way more VRAM on the smaller buses, while retaining high bandwidth

It does not bode well for AMD in the high-end space but their focus should be to re-grab marketshare and improve FSR to match DLSS/DLAA.

 

[envirovore]

Yeah but Kopite said 512 bit today. Most trustworthy leaker.

Usually its cheaper going with smaller bus and faster mem.

However if Nvidia uses 384 bit again, then 32 Gbps should be minimum.

From two days ago, leaked by Kopite as well...saying that 5000 will use the same bus width across skus as Ada.
https://wccftech.com/nvidia-geforce-rtx-50-blackwell-gb200-gaming-gpus-same-memory-ada-384-bit-max/

I'm aware they tend to get a lot correct as far as rumors and speculation go, but as stated up thread that's still all it is currently - rumors and speculation that changes (nearly) daily.

Edit: and I've seen recent rumors claiming either GDDR7 and GDDR6x for 5000 series.

 

[Disiw]

From two days ago, leaked by Kopite as well...saying that 5000 will use the same bus width across skus as Ada.
https://wccftech.com/nvidia-geforce-rtx-50-blackwell-gb200-gaming-gpus-same-memory-ada-384-bit-max/

I'm aware they tend to get a lot correct as far as rumors and speculation go, but as stated up thread that's still all it is currently - rumors and speculation that changes (nearly) daily.

Edit: and I've seen recent rumors claiming either GDDR7 and GDDR6x for 5000 series.

Yeah he corrected that, Nvidia might not even be sure themself yet, things are changing all the time, until official, we are still 6-9 months from release, 12 months could happen as well, since gaming GPU market is not a priority for Nvidia right now and 4000 series just got refreshed.

Nvidia has no problems at all right now. Sales are good. Marketshare is high. They is not FORCED to release 5000 series.

I highly doubt GDDR6X will be used again, not on highest end stuff, maybe on lower end stuff like 5070 eventually

 

[Achaios]

5090 ----->$ 2699 msrp (Whales have proven time and time again they will gladly pay any price for a 90 series GPU)
5080------> $1499 msrp ("GDDR7 is very expensive")
5070------> $1199 msrp ("Inflation & GDDR7 & new process" etc)
5060------> $799 msrp ("The Era of Cheap NVIDIA Gpus Is Over")
5050------> $599 msrp ("We don't make NVIDIA GPUS for poor people").

Bow down before your Chinese overlord.

 

[MarcusNumb]

AI hype + new GDDR7 = excellent excuse to increase prices.

This time imo if AMD play their "value focus" card right, they will gain big. Maybe I'm dumb but I prefer paying 500-600 for a 7900xtx performance like GPU than 1200 for a 4090 performance like GPU. Not everyone has a 240hz 4k monitor and a 1200w PSU to max out a 4090.

 

[ScottSoapbox]

The memory interface will be driven by the amount of VRAM more than anything else.

AI likes VRAM and Nvidia likes selling those cards for far more so expect as like VRAM growth as they can get away with.

The 5090 is the only possible exception because they can price it high enough that they don't care if it is used as a baby AI card.

 

[erickmendes]

But will it run Crysis?

 

[pcnthuziast]

MFW the 5090 has an MSRP of 2999.99

 

[godrilla]

5090 ----->$ 2699 msrp (Whales have proven time and time again they will gladly pay any price for a 90 series GPU)
5080------> $1499 msrp ("GDDR7 is very expensive")
5070------> $1199 msrp ("Inflation & GDDR7 & new process" etc)
5060------> $799 msrp ("The Era of Cheap NVIDIA Gpus Is Over")
5050------> $599 msrp ("We don't make NVIDIA GPUS for poor people").

Bow down before your Chinese overlord.

If the rumors are true of the 2.5x raytracing performance gains over the 4090, then the 4090 owners might just make an even swap for 5080.
$1200 premium from the 5080 to the 5090 for barely 20 to 25% ( current delta between 4080/4090) . Compared to $400 difference in price from the launch price 4080/4090. Currently the difference in price between the lowest 4080super ($1029) and 4090 ($1829) is $800. Via pcpartpicker website. Nvidia would be testing brand loyalty more than ever at those prices.

 

[Branch]

The pricing is what I’m worried about.

I'm preparing myself to be completely disappointed.

 

[daffy duck]

Good, I'm looking for a change in this timeframe, so knowing nVidia is planning something I can safely bet AMD will propose some new cards as well.

Depends, do you want a mid-tier RDNA4 this year in late Q3 or early Q4, or a high-end RDNA5 in very late 2025 at best.

 

[Disiw]

Depends, do you want a mid-tier RDNA4 this year in late Q3 or early Q4, or a high-end RDNA5 in very late 2025 at best.

True. RDNA4 is very late 2025 or most likely 2026

Meaning Nvidia can do whatever they want in the high-end space for 2 years

AMD is sleeping

 

[Disiw]

If the rumors are true of the 2.5x raytracing performance gains over the 4090, then the 4090 owners might just make an even swap for 5080.
$1200 premium from the 5080 to the 5090 for barely 20 to 25% ( current delta between 4080/4090) . Compared to $400 difference in price from the launch price 4080/4090. Currently the difference in price between the lowest 4080super ($1029) and 4090 ($1829) is $800. Via pcpartpicker website. Nvidia would be testing brand loyalty more than ever at those prices.

Brand loyalty? AMD will have nothing in this segment before 2026 when RDNA5 hopefully don't disappoint

RDNA4 won't have any high-end SKU at all. It will be 7900XT performance at lower watts at best

Nvidia is competing with themselves at this point, AMD is MIA

 

[OortCloud]

The barriers to entry on the GPU market mean that nVidious can pretty much charge what they want. It seems AMD price-fix their GPUs to whatever ludicrous prices their 'competition' set.

Hopefully Intel will have finally caught up with all the huge hurdles a new discrete GPU maker has to jump through with per-game/per-engine/per-dx-version optimisations that need to be put in place to compete. They seem to be getting there these days. Hopefully they won't join the GPU cartel.

 

[Disiw]

The barriers to entry on the GPU market mean that nVidious can pretty much charge what they want. It seems AMD price-fix their GPUs to whatever ludicrous prices their 'competition' set.

Hopefully Intel will have finally caught up with all the huge hurdles a new discrete GPU maker has to jump through with per-game/per-engine/per-dx-version optimisations that need to be put in place to compete. They seem to be getting there these days. Hopefully they won't join the GPU cartel.

The main problem AMD has, is that being aggressive on GPUs, means less CPU/APU output. Same process nodes are used. It makes little sense for AMD to sell cheap GPUs on prime-nodes, which is needed to compete in the high end market.

AMD relies 100% on TSMC at this point. TSMC raised prices alot too over the years. And AMD mostly can't afford using TSMC's best nodes like Apple (especially) and Nvidia, soon Intel too with Arrow Lake as well (high-end Arrow Lake chips are confirmed to be using 3nm TSMC)

Nvidia don't rely on TSMC to the same extent. They went Samsung with 3000 series and already secured Intel 20A/18A. Samsung 2nm too probably.

I would not be surprised if RTX 6000 series is made on Intel 18A (or Samsung again, if their 2nm node is decent with good yields)

This means Nvidia can go full on AI/Enterprise on TSMC 2nm or better at the time.

RTX 5000 series and Blackwell AI/Enterprise will share same 3nm process. Or, 5090 and 5080 will for sure. Driving prices up.

5070 and 5060 will launch much later, maybe 2H 2025, and might not use 3nm or GDDR7 for that matter. We don't know and Nvidia is not under pressure at all. 4000 SUPER refresh happend for a reason, 5070 and below is not even close.

Nvidia don't need to use peak nodes for gaming to beat AMD. Especially not when AMD said Radeon 8000 series won't have any high-end offerings.

 

[OortCloud]

The main problem AMD has, is that being aggressive on GPUs, means less CPU/APU output. Same process nodes are used. It makes little sense for AMD to sell cheap GPUs on prime-nodes, which is needed to compete in the high end market.

AMD relies 100% on TSMC at this point. TSMC raised prices alot too over the years. And AMD mostly can't afford using TSMC's best nodes like Apple (especially) and Nvidia, soon Intel too with Arrow Lake as well (high-end Arrow Lake chips are confirmed to be using 3nm TSMC)

Nvidia don't rely on TSMC to the same extent. They went Samsung with 3000 series and already secured Intel 20A/18A. Samsung 2nm too probably.

I would not be surprised if RTX 6000 series is made on Intel 18A (or Samsung again, if their 2nm node is decent with good yields)

This means Nvidia can go full on AI/Enterprise on TSMC 2nm or better at the time.

RTX 5000 series and Blackwell AI/Enterprise will share same 3nm process. Or, 5090 and 5080 will for sure. Driving prices up.

5070 and 5060 will launch much later, maybe 2H 2025, and might not use 3nm or GDDR7 for that matter. We don't know and Nvidia is not under pressure at all. 4000 SUPER refresh happend for a reason, 5070 and below is not even close.

Nvidia don't need to use peak nodes for gaming to beat AMD. Especially not when AMD said Radeon 8000 series won't have any high-end offerings.

I agree with everything you said, but this does mean AMD basically is happy to sell their GPUs poorly but with really healthy margins and therefore not over-compete on price with nVidia. Sure now and then they give slightly better value especially on low end parts but it's nothing remarkable. It is a unspoken cartel between the two of them. Hence even really poor low end cards like the 4060 get $300 launch prices.

 

[godrilla]

Brand loyalty? AMD will have nothing in this segment before 2026 when RDNA5 hopefully don't disappoint

RDNA4 won't have any high-end SKU at all. It will be 7900XT performance at lower watts at best

Nvidia is competing with themselves at this point, AMD is MIA

Yes brand loyalty especially at those prices quoted. I mean AMD will also win in the end by allowing Nvidia to swell prices it can fill in the gap accordingly with perceived hero like underdog. By focusing in the midrange with rdna 4 in the near term they can improve market share. MIA? no rdna 5 is part of roadmap and should make a nice comeback. If 5090 is indeed north of $2500 how much do you think AMD will charge for a competitive rdna 5 card? $1000? Is Nvidia competitive with itself is half true because at those prices it would be offering similar performance per dollar as the 2022 lineup in 2025. Nvidia needs to constantly fortify its position of dominance to justify its price premium. 80% market share is Nvidia's to lose otherwise.