Question 8GB seems to gimp the crap out of the 4060 (Guru3D article inside)

[moonbogg]

NVIDIA GeForce RTX 4060: 3DMark scores leak

The NVIDIA GeForce RTX 4060, equipped with the AD107 GPU, has been subjected to extensive analysis by numerous evaluators and contrasts them with the RTX 3060 12GB and 3060 8GB models....

www.guru3d.com

From the article: "In 3DMark evaluations, the RTX 4060 outperforms the RTX 3060 12GB by an average of 22.8%, and the 8GB Ampere SKU by 47.9%."

I mean, lol, what the crap? Look at that difference from the 3060 12GB version just from the extra vram. Oh my god this 4060 is such a disgusting piece of insulting garbage. I'm convinced Nvidia actually HATES gamers' guts. I mean they just keep kicking us right in the damn throat with releases like this. If it wasn't basically a monopoly, they'd be getting destroyed in the market by releasing smelly, hot, dirty diaper infested garbage like this. This entire generation is rancid trash juice splashed all over your face, like getting in your mouth and everything.

 

[kondziowy]

This really is rtx4050 in disguise, so it performs like a 4050 50 series was never running Ultra presets. Die size 146mm2... maybe its actually RTX4040Ti. Doesn't matter. People are lining up to buy it already.

 

[MrTeal]

The 3060 8GB also has a 128bit bus and doesn't have the large L2 of Ada to help hide it. The bandwidth issue is probably more of a problem for that card than the 8GB.

"While a direct comparison between the RTX 4060 and its 8GB precursor would be justifiable, NVIDIA's official comparison uses the original 12GB SKU."
A comparison to the 8GB 3060 really isn't justified since that card has bad availability, a terrible price, and more importantly the vast, vast majority of 3060 owners have the original 12GB model.

 

[Thunder 57]

Can't even wait for reviews before white knighting Nvidia? 3DMark hasn't been relevant in like 20 years.

 

[A///]

Can't even wait for reviews before white knighting Nvidia? 3DMark hasn't been relevant in like 20 years.

given that stupid song and dance jensen did at computex with the cruddy ai song he white knighted himself later that night I'm sure. if it isn't clear enough I'm not sure what white knighting means but I assume it's foul and dirty.

 

[Thunder 57]

given that stupid song and dance jensen did at computex with the cruddy ai song he white knighted himself later that night I'm sure. if it isn't clear enough I'm not sure what white knighting means but I assume it's foul and dirty.

Not at all actually.

What does the slang white knight mean?
white knight. noun. a champion or rescuer, esp a person or organization that rescues a company from financial difficulties, an unwelcome takeover bid, etc.

 

[Leeea]

I'm not sure what white knighting means but I assume it's foul and dirty.

A knight's job is to murder common people who annoy the lord or happen to be on the road at the wrong time. The common people are typically unarmed, no means to defend themselves, and are murdered without regard to age or gender. It is called the dark ages for a reason.

It is by definition a foul and dirty job. It is wrapped in concepts like and honor and chivalry, but those only apply to royalty. Commoners are akin to live stock or cattle.

White knighting is a reference to bleaching the knights uniform to get the blood of all the common people he has murdered lately out of it. Leaving his uniform nice and white!

 

[Thunder 57]

I should have just said simping or shilling.

 

[moonbogg]

I should have just said simping or shilling.

Shimping

 

[GodisanAtheist]

I should have just said simping or shilling.

-Trying to get the blood of the common folk out of that tabard eh?

 

[Mopetar]

Shimping

Sean Connery is that you?

 

[Saylick]

Shimping

Where's Anand when you need him?

 

[amenx]

The 3060 8GB also has a 128bit bus and doesn't have the large L2 of Ada to help hide it. The bandwidth issue is probably more of a problem for that card than the 8GB.

"While a direct comparison between the RTX 4060 and its 8GB precursor would be justifiable, NVIDIA's official comparison uses the original 12GB SKU."
A comparison to the 8GB 3060 really isn't justified since that card has bad availability, a terrible price, and more importantly the vast, vast majority of 3060 owners have the original 12GB model.

All 3060 variants = 192 bit bus.

 

[MrTeal]

All 3060 variants = 192 bit bus.

I don't believe that's accurate. The 8GB version launched at the end of last year has a 128 bit bus and 2MB L2 vs 192bit and 3MB L2.

Der ultimative GeForce-Grafikkartenvergleich

Hier siehst du, was sich mit den letzten vier RTX-und GTX-Grafikkartenserien verändert hat.

www.nvidia.com

 

[blckgrffn]

I think there is an argument to be made that the 4060 die should be smaller - cutting out much of the L2 cache - and that the 192 bit memory bus should have remained for an easy 12GB of memory card and zero regressions compared to the 3060. Even halving the L2 cache to 16MB (like Navi 24) the with the 192 bit bus should have netted a solid uptick in performance at 1080p with a relatively huge cache hit chance (4 mb is super low, 16MB flirts with 30%, so cache hits might have gone up 300-400%).

They must have made a decision that lower power (due to the efficiency of the L3 cache) and TSMC transistors was the more profitable path than having a wider bus and more memory chips.

 

[jpiniero]

They must have made a decision that lower power (due to the efficiency of the L3 cache) and TSMC transistors was the more profitable path than having a wider bus and more memory chips.

Actually I suspect AD107 was intended to be 96-bit (hence why it and AD106 have the same bus width).

 

[coercitiv]

I think there is an argument to be made that the 4060 die should be smaller - cutting out much of the L2 cache - and that the 192 bit memory bus should have remained for an easy 12GB of memory card and zero regressions compared to the 3060. Even halving the L2 cache to 16MB (like Navi 24) the with the 192 bit bus should have netted a solid uptick in performance at 1080p with a relatively huge cache hit chance (4 mb is super low, 16MB flirts with 30%, so cache hits might have gone up 300-400%).

Both the 4060 and the 4060Ti would have greatly benefited from another 32 bit lane, with the rest of the setup remaining as is. This would have granted them a 10GB VRAM buffer and a less obvious drop in 1440p perf. Alternatively they could have gone the route you just described, though maybe a much lower L2 cache may have influenced RT performance? Nvidia really really wants us to enable RT even on the low-end cards.

 

[jpiniero]

Both the 4060 and the 4060Ti would have greatly benefited from another 32 bit lane, with the rest of the setup remaining as is. This would have granted them a 10GB VRAM buffer and a less obvious drop in 1440p perf. Alternatively they could have gone the route you just described, though maybe a much lower L2 cache may have influenced RT performance? Nvidia really really wants us to enable RT even on the low-end cards.

Or they realize this is a throwaway generation and it doesn't matter. Maybe the refresh will have GDDR7 with more than 2 GB/chip.

 

[moonbogg]

Or they realize this is a throwaway generation and it doesn't matter. Maybe the refresh will have GDDR7 with more than 2 GB/chip.

I agree that Nvidia is certainly not stupid or ignorant about their products or the markets they operate in. They know these products aren't attractive and won't sell well. We can speculate about what their reasons are. I think it's prioritizing AI sales and maybe clearing out old stock, but I think it's mostly prioritizing AI stuff. However, does that explain AMD's similar behavior? Does AMD sell a bunch of AI stuff too? I don't think AMD is having a huge AI hardware boom like Nvidia, so why is AMD shuffling products names and jacking up prices too?

 

[Saylick]

[snip] so why is AMD shuffling products names and jacking up prices too?

Because it likely makes them more money than not shuffling product names and jacking up prices.

 

[moonbogg]

Because it likely makes them more money than not shuffling product names and jacking up prices.

Hmm. Money as the motivation you say. I'll have to think on that one.

 

[coercitiv]

Maybe the refresh will have GDDR7 with more than 2 GB/chip.

The much more expensive GDDR7?!

(not to mention this will probably require new chips with updated memory controllers)

Spoiler

Earlier this year, Cadence also revealed a few details about GDDR7. According to the company, the new memory technology will use PAM3 signaling alongside NRZ. The former offers a more efficient data transmission rate per cycle compared to the latter, although it transmits only 1.5 bits of data per clock edge as opposed to 2 bits by NRZ.

It is worth noting that the use of a new encoding mechanism in GDDR7 means it will need new memory controllers and graphics processors for widespread adoption. So, while the early-2024 timeline for GDDR7 introduction is certainly exciting, it might take some time for the technology to be available in commercial products.

 

[KompuKare]

The much more expensive GDDR7?!

(not to mention this will probably require new chips with updated memory controllers)

Spoiler

Earlier this year, Cadence also revealed a few details about GDDR7. According to the company, the new memory technology will use PAM3 signaling alongside NRZ. The former offers a more efficient data transmission rate per cycle compared to the latter, although it transmits only 1.5 bits of data per clock edge as opposed to 2 bits by NRZ.

It is worth noting that the use of a new encoding mechanism in GDDR7 means it will need new memory controllers and graphics processors for widespread adoption. So, while the early-2024 timeline for GDDR7 introduction is certainly exciting, it might take some time for the technology to be available in commercial products.

It does say alongside the old style signalling... So you could have a card using GDDR7 but using GDDR6 signalling.
Pay extra but not used the full potential, but maybe something which could be marketed as using GDDR7. Crazy but never underestimate the power of marketing buzzwords

 

[DeathReborn]

Or they realize this is a throwaway generation and it doesn't matter. Maybe the refresh will have GDDR7 with more than 2 GB/chip.

The memory makers could make 3GB/4GB GDDR6 if they saw a way to cash in on it.

 

[Kaluan]

4060 is this generation's 6500XT

That being said, unlike the 6500XT, I don't think we'll EVER see a 16GB version.