NVIDIA's RTX 5060 "Blackwell" Laptop GPU Comes with 8 GB of GDDR7 Memory Running at 28 Gbps, 25 W Lower TGP

[R0H1T]

Interesting so this affects only GDDRxx based cards or every GPU? Like this ~

NVIDIA B200 SXM 192 GB Specs

NVIDIA GB100 x2, 1837 MHz, 16896 Cores, 528 TMUs, 24 ROPs, 98304 MB HBM3e, 2000 MHz, 4096 bit

www.techpowerup.com

NVIDIA Blackwell Architecture and B200/B100 Accelerators Announced: Going Bigger With Smaller Data

www.anandtech.com

 

[Onasi]

@R0H1T
I was talking specifically about GDDR. HBM has its own rules and standards. It’s different since there it’s eight 128-bit channels per stack for HBM and HBM2 and sixteen 64-bit ones for HBM3 onward. So it always comes down to 1024-bit memory bus per stack. Total width therefore is based on the number of said stacks.

 

[R0H1T]

I guess it's understandable if that's something which is simply not possible with GDDR7 as of now. Also who knows maybe Nvidia will do a 96/192 bit 24GB 5060/Ti on the desktop just to spite critics like with 3070/3060 just a while back?

 

[Prima.Vera]

Funny how I still can play ANY possible game out there on FULL details and Ultra textures, without any hiccups, with my RTX 3080 with "just" 10GB of VRAM.
Even funnier is that my VRAM usage, as shown by RivaTuner, never goes more than 80% in any of the games...
And yes, I play natively on 3K resolution, or 3440x1440.

 

[watzupken]

While the TGP seems lower on Blackwell, but in reality, the existing RTX 4070 and 4060 are hard limited to about 90 to 100W in games. So I see no improvement here really.

 

[Minus Infinity]

I get this is a laptop but 5060 mobile option won't be cheap, so I think it would be reasonable to expect my expensive laptop to play 1080p with a lot of quality settings maxed. But for that you'll need to spend a lot more for the 12GB 5070 mobile part.

Next year GDDR7 with 3GB/32bit modules comes out, so you will be able to offer 12GB on 128bit bus, 18GB on 192bit bus, 24GB on 256bit bus etc. I wonder if they'll offer those options on both laptop and desktop parts. Wouldn't mind seeing a 5080 with 24GB

 

[Apocalypsee]

Lowering TGP is good but, laptop manufacturer LOVE to abuse this by putting dinky heatsink and fan that is lower than the advertised TGP rating because 'This laptop is 1mm thinner than the old one!!11one' and really, even my 3050 mobile would scales better with higher TGP but alas its limited by cooler and rating that's been set in VBIOS.

 

[igormp]

Next year GDDR7 with 3GB/32bit modules comes out, so you will be able to offer 12GB on 128bit bus, 18GB on 192bit bus, 24GB on 256bit bus etc. I wonder if they'll offer those options on both laptop and desktop parts. Wouldn't mind seeing a 5080 with 24GB

I doubt it. Nvidia has great market segmentation between their consumer em dc/workstation parts by just using different vram amounts because that's the most limiting factor for tons of workloads.

Slap 12 or 16GB of vram in a 5060 and it'll now be called a RTX B2000 that costs 3x more.

 

[Onasi]

@Minus Infinity
We might see more playing around with 3 and 6 gig chips on the inevitable refreshes. The first wave of GDDR7 is all 2 gig dies and I assume the orders have been placed way in advance, so NV is working with what they have.
Hell, they might just again ignore such concerns - GDDR6 also had 1.5 and 3 gig density in spec and those were shown, but apparently nobody was particularly interested and so no real mass production happened.

 

[wolf]

We might see more playing around with 3 and 6 gig chips on the inevitable refreshes

I really hope so, moving from 2GB to 3GB, or even down to 1.5GB can give us 12GB cards on 128-bit bus, 18GB on 192-bit, and 24GB on 256-bit bus, seems like an easy win to increase VRAM by 50% instead of double without needing to go too hard on engineering trickery.

 

[A&P211]

Nah, it’s alive, just in a different form. It’s basically, and forgive me if I am getting this wrong, is now a low power mode that ALL laptop GPUs possess instead of being a separate GPU SKU. And also a shit-ton of other “power saving” tech. Even NV seems to basically have no idea what it is anymore. But it is “AI powered”, so you know it’s good, right?

Again, the name Max-q thing was stopped being used some years ago.

 

[Onasi]

Again, the name Max-q thing was stopped being used some years ago.

It rebranded, as I mentioned. I admitted already that it isn’t in the GPU name as a separate SKU, but it’s not gone as a concept. There are still low tuned TGP mode models, it just isn’t explicitly marked in the branding. If you disagree - take it up with NVidia, here is their official statement:

Max-Q branding is not going away.
When we originally introduced Max-Q back in 2017, the brand was initially used in GPU naming since Max-Q referred to the GPU TGP only.
Today, 3rd Generation Max-Q is broader, and is a holistic set of platform technologies and design approach to building powerful and thin laptops.
In addition, to be more transparent about a laptop’s exact capabilities, RTX 30 Series laptops now show more information than ever, listing exact TGP, clocks and features supported. You will find this in the control panel which now reports maximum power (TGP+Boost), and support for key features including Dynamic Boost 2, WhisperMode 2, Advanced Optimus, and others, all of which fall under the Max-Q umbrella.
We strongly encourage OEMs to list clocks and other technologies a laptop supports, including Advanced Optimus, Dynamic Boost 2, and more. Ultimately, like all laptop features and specs, it is up to the OEM to market what their particular laptop configuration supports.

Source: HotHardware

 

[Prima.Vera]

Why tf do you need 24GB of VRAM?? For what?? Even if you play on 8K with 8K textures you still won't be able to fill the VRAM.
And I'm talking about actually using the VRAM, not just caching it, like most of the engines are doing it nowadays.

 

[Lewzke]

Even at

They did mention CUDA stuff, if they're running any kind of large-ish ML model, VRAM often is a pretty major limitation compared to raw compute.

Even with 12GB you have a very limited space for LLM models, so he was right 5060 is a significant jump in performance, but I think still that 8GB VRAM is not that promising, I expected 12.

 

[Dimitriman]

I see a lot of cope in these replies about the amount of VRam. If experience taught us anything about Nvidia, it was two things: 1) They will overcharge this to new heights due to no competition from AMD, and 2) if we don't ask for better, they will absolutely not give us better.

We need to demand better from Nvidia. Especially for the amount of money they keep demanding from us.

 

[Quicks]

Nvidia is taking their loyal customers for idiots. They know if they future proof their products then you might skip one generation and they don't want that. They want to keep upgrading every generation.

5060 Laptop or Desktop GPU regardless 12 GB Vram minimum. 5070 must start with 16 GB Vram and 5080 20 GB Vram, and lastly 5090 with 24 GB Vram. That seems fair.

 

[Vayra86]

x60 laptop class GPUs were always, will always be and are today typical 'Fools and Money parted' territory.

Just get an IGP and face the fact you're not buying a 'gaming' laptop here. I've had my share of this class of laptops and it was all unfiltered poop. You can not game on this class of GPUs. Get an IGP and save money, or spend big on something that does run a game proper. Get a Steam Deck alongside a 500 dollar laptop and you have the same 1K dollar spent but then you have two devices, of which the gaming device doesn't drain your battery in a half hour while your non gaming device can be thin and light and excellent for anything that isn't gaming.

/thread really.
And also /discussion over 8GB on this class of cards. People... this is a lying piece of shit product what do you expect. Shouldn't even be called Geforce and frankly since we have half decent gaming APUs, should probably just not even exist.

Nvidia is taking their loyal customers for idiots. They know if they future proof their products then you might skip one generation and they don't want that. They want to keep upgrading every generation.

5060 Laptop or Desktop GPU regardless 12 GB Vram minimum. 5070 must start with 16 GB Vram and 5080 20 GB Vram, and lastly 5090 with 24 GB Vram. That seems fair.

More like customers fooling themselves. If you buy the right GPU for your needs you can easily last upwards of 5 years on one. I consider myself a devout gamer and I played 6 years on a GTX 1080. Yes, it means you will not have the latest greatest nor the highest FPS you can possibly achieve. So what?

But what do we see every time Nvidia drops a new gen? Tech enthusiasts jumping on it for their 30% performance gain and in recent generations quite possibly even less, or just a 'new feature' like DLSS3. OF COURSE Nvidia is going to bleed those idiots dry. They literally ask for it. The price/perf $ figure did not, or barely moved between Ampere and Ada, but lots of people didn't wait for Supers to improve it one bit. They saw a marginal performance gain and jumped on upwards of 600 dollar worth GPUs. Never mind the fact they practically stood still in the large scheme of things. Must buy buy buy

 

[wolf]

x60 laptop class GPUs were always, will always be and are today typical 'Fools and Money parted' territory.

Just get an IGP and face the fact you're not buying a 'gaming' laptop here.

You can not game on this class of GPUs.

I think that's a very broad brush to paint all this way. In 2018 I bought a GTX 1060, 1080p 144hz laptop and you bet your ass it was a better gaming laptop than igp at the time. The 5060 will be at least as fast as a current mobile 4060 is which significantly outperforms any igpu even at low wattages. A butt-tonne of people will play on devices that run it far worse.... but "You can not game on this class of GPUs.?" - Hard disagree.

 

[igormp]

Even at

Even with 12GB you have a very limited space for LLM models, so he was right 5060 is a significant jump in performance, but I think still that 8GB VRAM is not that promising, I expected 12.

You can run way more models with 12GB rather than 6 or 8gb. I did the jump from my 2060 Super to a 3060 back then exactly for the extra 4gb, while having pretty much the same performance. Allowed me to run bigger models for cheap. (but since then did the jump to 2x3090 for that sweet 48GB, and that's still not comfortable enough for larger models).

And reminder that LLMs and not everything out there, you can run many other models as well (in my case it was CNNs and ViTs).

 

[Chrispy_]

If we're talking

I mean they gave the 4070 laptop 8GB also lol....

The 4070 laptop isn't even 4070 AD104 silicon, it's the same bullshit AD106 128-bit rubbish that the xx60 series gets on desktop.

For sure the lack of progress at the 60/60ti range has been troubling.

One of Nvidia's greatest tricks this generation has been to pass off the smallest, most crippled AD107 silicon as a 4060. It's the 4050 in all but name.

xx60 series GPUs have been silicon configurations like xx106 for ages. 10-series, 16-series, 20-series, 30-series....

 

[Vayra86]

I think that's a very broad brush to paint all this way. In 2018 I bought a GTX 1060, 1080p 144hz laptop and you bet your ass it was a better gaming laptop than igp at the time. The 5060 will be at least as fast as a current mobile 4060 is which significantly outperforms any igpu even at low wattages. A butt-tonne of people will play on devices that run it far worse.... but "You can not game on this class of GPUs.?" - Hard disagree.

Yeah perhaps it got slightly better from Pascal onwards, prior to that at least it was a major clusterfuck in that segment of dGPUs. A big part of that cluster was also the fact the GPUs got stuck into laptops with throttling CPUs, or a throttling system the moment you started up a game. We've gained a bit of (CPU) headroom in that sense over the last years. So yeah I think the prospects of x60 mobile have improved slightly then.

Regardless, they're bottom tier, I read some here expecting 12GB on this class of GPUs and that just doesn't make sense. Says the VRAM Herald of TPU, no less... And honestly - would you NOT prefer the alternative I spoke of? If you are spending 1k... spend 500 on a thin laptop with IGP and spend the other half on a Steam Deck. Best of both worlds, all portable, separate batteries and both devices will be actually good at what they're doing.

 

[Chrispy_]

Regardless, they're bottom tier, I read some here expecting 12GB on this class of GPUs and that just doesn't make sense. Says the VRAM Herald of TPU, no less...

8GB may even be fine, as the GB207 rumoured (not guaranteed) to become the vanilla 5060 is a smaller, less impressive config than the 4060, which is already a pretty underwhelming GPU. There's a good chance the 5060 isn't much of an improvement on the 4060, so potentially like the 40-series, it'll be about new features rather than more performance.

17% fewer resources on the same crippled 128-bit GDDR6 memory bus means that clockspeed and IPC gains are going to need to be significant to even just match a 4060's performance:

This is all rumour though.

• Nvidia might make these RTX 5050 instead of 5060 and sell them at a lower price, meaning that the 5060 is actually using the 206 die, similar to the current 4060Ti.
• This leak could be completely wrong and 12GB variants of the 5060 get launched after all.
• A combination of more cache and higher-density GDDR6 means these could be viable 16GB cards.

(that last one is particularly unlikely - but at this stage everything is possible because it's all just speculation and unverified leaks)

EDIT:
One thing I will say taking a second look at the leak info is that GB202 (5090) with even more SMs and an even wider 512-bit bus than the 4090 is going to be an absolute monster in every way: Insane price, 32GB of very expensive GDDR7, and likely a 600W TDP on the most conservative reference/founders edition! Building a smaller GPU on a simpler PCB for 'just' a 384-bit memory bus like a 4090 is practically child's play by comparison....

 

[oxrufiioxo]

8GB may even be fine, as the GB207 rumoured (not guaranteed) to become the vanilla 5060 is a smaller, less impressive config than the 4060, which is already a pretty underwhelming GPU. There's a good chance the 5060 isn't much of an improvement on the 4060, so potentially like the 40-series, it'll be about new features rather than more performance.

17% fewer resources on the same crippled 128-bit GDDR6 memory bus means that clockspeed and IPC gains are going to need to be significant to even just match a 4060's performance:

View attachment 360861

This is all rumour though.

• Nvidia might make these RTX 5050 instead of 5060 and sell them at a lower price.
• This leak could be wrong.
• A combination of more cache and higher-density GDDR6 means these could be viable 16GB cards

(that last one is particularly unlikely - but at this stage everything is possible because it's all just speculation and unverified leaks)

Considering they are using only a slightly better node they are gonna have to pull a pascal out of their hat for this generation to be really good or make the Desktop chips much larger tier for tier.

I'm expecting a decent buy expensive 5090 with everything else being meh AF price to performance. Would love to be wrong though.

 

[igormp]

One thing I will say taking a second look at the leak info is that GB202 (5090) with even more SMs and an even wider 512-bit bus than the 4090 is going to be an absolute monster in every way: Insane price, 32GB of very expensive GDDR7, and likely a 600W TDP on the most conservative reference/founders edition! Building a smaller GPU on a simpler PCB for 'just' a 384-bit memory bus like a 4090 is practically child's play by comparison....

You're assuming nvidia will use the full GB202 die for the 5090, which I really doubt will be the case. The 4090 was a die cut of the AD102 already, I can see the 5090 being an even rougher die cut with some memory controllers disabled, maybe with a 416-bit or 448-bit bus so it can have 26GB or 28GB to be considered an upgrade over the 4090.

The 4090 already had less than 90% of the SMs of the full AD102, I guess the 5090 would be in a similar ballpark with that, or even a bit less.

Rumours do point to 448-bit with 28GB being the case, with only 160 out of the 192 SMs being available.

 

[ARF]

You're assuming nvidia will use the full GB202 die for the 5090, which I really doubt will be the case. The 4090 was a die cut of the AD102 already, I can see the 5090 being an even rougher die cut with some memory controllers disabled, maybe with a 416-bit or 448-bit bus so it can have 26GB or 28GB to be considered an upgrade over the 4090.

The 4090 already had less than 90% of the SMs of the full AD102, I guess the 5090 would be in a similar ballpark with that, or even a bit less.

Rumours do point to 448-bit with 28GB being the case, with only 160 out of the 192 SMs being available.

At this point, even GB203 will be enough to beat substantially the old Radeon RX 7900 XTX, while the poor small Navi 48 will directly compete against GB205.
Expect AMD's GPU shipments to fall off the cliff even further, from the current 9% to 4%.
Unless Navi 48 top product retails for 199$. It won't.