Discussion Nvidia Blackwell in Q1-2025

[jpiniero]

^^^^

Def too optimistic at the low end.

 

[igor_kavinski]

^^^^

Def too optimistic at the low end.

Yeah. AI texture compression incoming with 5000 series!

 

[tajoh111]

Teclab Makes Its Own GeForce RTX 4090 SUPER GPU With Binned 26 Gbps G6X Memory, NVIDIA AD102 On 3090 Ti HOF OC Lab PCB, 13% Faster Than Stock

Teclab has just created the world's first GeForce RTX 4090 SUPER GPU using binned G6X memory, a 3090 Ti HOF OC Lab PCB & NVIDIA's AD102.

wccftech.com

Blackwell doesn't need much to be 50 percent faster.

Just a memory overclock on a RTX 4090 gets 13% better performance.

If Nvidia released a RTX 4090 ti with fast memory, all the L3 cache enabled, more sm enabled and basically the same clocks, you would have a card 25% faster than a RTX 4090.

Nvidia is competing against themselves at this point. A full RTX 4090 ti might be close to 50% faster than a 7900xtx but for next generation, it makes it more difficult for their next generation to stand out.

 

[coercitiv]

Just a memory overclock on a RTX 4090 gets 13% better performance.

in an 8K benchmark.

 

[MrTeal]

in an 8K benchmark.

With just an easy memory OC of swapping AD102 to a new PCB with binned memory chips. ezpz.

TIL AD102 and GA102 are pin compatible though, that's pretty cool.

 

[amenx]

Interesting bit about GDDR7

The biggest fundamental change with GDDR7 is that it will use PAM3 signaling, where GDDR6 uses NRZ (non return to zero) signaling — and GDDR6X uses PAM4 signaling. PAM3 (3-level pulse-amplitude modulation) reduces energy requirements compared to NRZ, while being less complex to implement than PAM4 (4-level PAM). That should make GDDR7 manufacturing equipment less complex and less expensive, though that doesn't mean it will be inexpensive.

GDDR7 will also support non-power-of-two configurations, and we expect to see 24Gb and, eventually, 48Gb memory chips. GDDR6 may also have 24Gb solutions coming, though no company has shipped a product using such a configuration at present. This means 50% more memory can be put on each 32-bit interface so that a typical 128-bit graphics card, for example, could have 12GB of VRAM instead of only 8GB.

Another change is that a 32-bit GDDR7 memory interface gets subdivided into four 8-bit channels, which helps facilitate fetching larger chunks of data. Where GDDR5 was an 8n prefetch, and GDDR6 was 16n, GDDR7 will have a 32n prefetch architecture. This is a way to pull larger amounts of data from DRAM while still operating at relatively low clocks.

https://www.tomshardware.com/pc-components/gpus/what-is-gddr7-memory

 

[jpiniero]

Interesting bit about GDDR7

Not really. It's been long discussed about there being 3 and 4 GB chips on the roadmap. But it appears that it won't be ready for the first wave of Blackwell.

 

[DavidC1]

Another change is that a 32-bit GDDR7 memory interface gets subdivided into four 8-bit channels, which helps facilitate fetching larger chunks of data. Where GDDR5 was an 8n prefetch, and GDDR6 was 16n, GDDR7 will have a 32n prefetch architecture. This is a way to pull larger amounts of data from DRAM while still operating at relatively low clocks.

The prefetch part just means the total GT/s can be double the previous generation. GDDR6 was double GDDR5, and DDR5 was double DDR4.

 

[poke01]

Rumor: GeForce RTX 5090 base clock nears 2.9 GHz - VideoCardz.com

NVIDIA RTX 5090 with significantly higher base clock Chiphell leaker Panzerlied claims that the RTX 5090 may see much improved base frequency. Discussion on Chiphell forums has led to new details about the RTX 5090 graphics card. Panzerlied, who previously revealed that the RTX 5090 will use a...

videocardz.com

RTX 5090 should be an excellent upgrade for 3090users.

 

[TESKATLIPOKA]

Rumor: GeForce RTX 5090 base clock nears 2.9 GHz - VideoCardz.com

NVIDIA RTX 5090 with significantly higher base clock Chiphell leaker Panzerlied claims that the RTX 5090 may see much improved base frequency. Discussion on Chiphell forums has led to new details about the RTX 5090 graphics card. Panzerlied, who previously revealed that the RTX 5090 will use a...

videocardz.com

RTX 5090 should be an excellent upgrade for 3090users.

Without knowing boost clocks, this near 2.9GHz base clock means nothing much.

 

[MoogleW]

If true, then boost clock is expected to be around 3.2ghz (with the memory bandwidth to feed the GPU and improve clock scaling).

That would also mean that GB207 could aim to perform around AD106 level, the minimum expected improvement gen on gen improvement assuming GB207 remains with 24SMs, and maybe at the same time aim to be around the rumored level of Navi44 (Navi33 < Navi44<Navi32)

 

[jpiniero]

If true, then boost clock is expected to be around 3.2ghz (with the memory bandwidth to feed the GPU and improve clock scaling).

2.9 is likely the stock boost although you should expect that chips should be able to get over 3 in most games.

GB207 has 20 SMs, btw. Also it seems that the desktop version has a decent shot at using GDDR6 only, although I suppose that could change. The 4060 uses 17 gbps, so there's room for higher even with GDDR6.

 

[TESKATLIPOKA]

2.9 is likely the stock boost although you should expect that chips should be able to get over 3 in most games.

GB207 has 20 SMs, btw. Also it seems that the desktop version has a decent shot at using GDDR6 only, although I suppose that could change. The 4060 uses 17 gbps, so there's room for higher even with GDDR6.

GB207 is not interesting with even less SM than AD107.
GB206 looks a lot better in comparison, still 36SM and GDDR7.

 

[jpiniero]

GB207 is not interesting with even less SM than AD107.
GB206 looks a lot better in comparison, still 36SM and GDDR7.

I suspect the GB206 prices are going to be decently higher and/or the L2 cache is going to be slashed.

Edit: Also likely only 4070 NS performance best case. And 8 GB.

 

[TESKATLIPOKA]

I suspect the GB206 prices are going to be decently higher and/or the L2 cache is going to be slashed.

Edit: Also likely only 4070 NS performance best case. And 8 GB.

Of course higher price, but I expect 8 and 16GB variants.

 

[jpiniero]

Of course higher price, but I expect 8 and 16GB variants.

I am not sure there is room for it, assuming the price gap would be more than the $100 that was the difference between the 4060 Tis. I am thinking if 3 GB chips arrive sooner than thought, perhaps they would rush that out as a pricier option instead of 16.

Now that I thought about it... Realistic best case?

GB203 - A tad slower than the 4090 D.
GB205 - Between 4070 Ti Super and 4080, maybe.
GB206 - A bit faster than the 4070 NS but that might be a stretch. Plus would likely require them to not cut the L2 which I bet they do given GDDR7.
GB207 - Maybe it could eek out a bit over the 4060?

 

[TESKATLIPOKA]

I am not sure there is room for it, assuming the price gap would be more than the $100 that was the difference between the 4060 Tis. I am thinking if 3 GB chips arrive sooner than thought, perhaps they would rush that out as a pricier option instead of 16.

If there will be 24gbit modules at that time, then they will go with It for 12GB Vram, but I don't think It will be available at that time.

Now that I thought about it... Realistic best case?


GB203 - A tad slower than the 4090 D.
GB205 - Between 4070 Ti Super and 4080, maybe.
GB206 - A bit faster than the 4070 NS but that might be a stretch. Plus would likely require them to not cut the L2 which I bet they do given GDDR7.
GB207 - Maybe it could eek out a bit over the 4060?

GB206 would need to be 27-33-38% faster to be on par with 4070, that's not very likely, but I don't see a reason for cutting down 32MB L2 even If GDDR7 provides a big increase in BW .

GB207 can't be only a bit faster than 4060 as a new gen unless the price is competitive.

 

[igor_kavinski]

I predict the vanilla 5060 will come with 6GB VRAM that will offer more average performance in even texture heavy scenarios than older 4060 due to neural texture compression.

No bets if Jensen decides to lower that to 4GB to make it just good enough to charge $100 above 4060.

 

[jpiniero]

GB206 would need to be 27-33-38% faster to be on par with 4070, that's not very likely

True. But unlike 5 and 7, 6 keeps the same amount of SMs. Even 26 gbps GDDR7 would be a 40% increase in memory bandwidth.

, but I don't see a reason for cutting down 32MB L2 even If GDDR7

Saves die space.

GB207 can't be only a bit faster than 4060 as a new gen unless the price is competitive.

Pricing will be interesting for sure.

 

[TESKATLIPOKA]

True. But unlike 5 and 7, 6 keeps the same amount of SMs. Even 26 gbps GDDR7 would be a 40% increase in memory bandwidth.

AD106 was bottlenecked by BW, so just GDDR7 will increase performance, but I don't know If architectural changes(IPC, clocks) will provide at least 20% performance on top of that. But I forgot 4060Ti had 2SMs disabled, that's another few percent of extra performance. I think 25-30% is possible If I am optimistic, but not more.

Saves die space.

You will save only a few mm2 by cutting It down from 32->24MB. Don't think It's worth It, but Jensen knows best.

 

[TESKATLIPOKA]

Seasonic lists GeForce RTX 5090 with 500W TDP and RTX 5080 with 350W, 16-pin connector for all RTX 50 models



RTX 5060 has the biggest TDP increase, It's basically comparable to 4060Ti, but that's not surprising considering It's most likely GB206.

RTX 4060 8GB -> 115W TDP
RTX 4060 Ti 8/16GB -> 160/165W TDP
RTX 4070 12GB -> 200W TDP
RTX 4070 Super 12GB -> 220W TDP
RTX 4070 Ti 12GB -> 285W
RTX 4070 Ti Super 16GB -> 285W
RTX 4080 16GB -> 320W TDP
RTX 4080 Super 16GB -> 320W TDP
RTX 4090 24GB -> 450W

 

[jpiniero]

Seasonic lists GeForce RTX 5090 with 500W TDP and RTX 5080 with 350W, 16-pin connector for all RTX 50 models

View attachment 103053

Can't imagine Seasonic would be given that info, esp this early.

 

[Ghostsonplanets]

I'm more surprised at the RTX 5050 existing given 4050 skipped Desktop.

Not surprised at increased power draw. Same node + vastly increased clockspeed.

 

[reb0rn]

16 pin power cables are burning as cheap fries, are they insane to push 50+W on that insane crappy disaster!??
Is NV run by only monkeys now?

 

[TESKATLIPOKA]

Can't imagine Seasonic would be given that info, esp this early.

Why not? Those cards need PSUs capable of feeding them without any issues.
I believe It's true, but those values could still change depending on final specs.

I'm more surprised at the RTX 5050 existing given 4050 skipped Desktop.

Not surprised at increased power draw. Same node + vastly increased clockspeed.

In my opinion just a more sane naming without any Ti.