A Deeper Look At VRAM On GeForce RTX 40 Series Graphics Cards
What is VRAM? How much do I need? Does bus width matter? These questions and more are answered in our new explainer article.
www.nvidia.com
- Status
Yeah and normally when this happens they would just focus on the performance increase over 2 generations which actually makes sense for a lot of buyers.
I don’t know how many people on older GPUs who didn't deem Ampere fast enough are going to think an extra 15% changes things.
Maybe everyone who wants to upgrade hasn’t upgraded yet? Point is it’s better to market at those people than pitching frame gen nonsense to Ampere owners.
The 16GB vram made it interesting for ML
DavidGraham
Veteran
Tokyo Institute of Technology selected H100 for it's supercomputer TSUBAME4.
There is no 2060Ti. But there was a $380 2060 Super, which was nearly a 2070 in performance, which really doesn't make for that favorable a comparison, either. 50-60% performance improvement after four years, with zero increase in VRAM unless you want to spend an extra $100 on top of the existing price hike.
In reality, like everything with this Lovelace lineup - the actual processor on its own is fantastic. It's a proper, sub 200mm² lower end part that's performing better than last generation's upper midrange part. The problem is that they are again renaming and upcharging a lower end part to a higher tier. If this was a 4050/4050Ti for like $270, few people would complain. Nvidia's ridiculous, insane greed is literally the only problem here.
The 4060 is equally bad, being a ~150mm² genuine bottom level part, sold for midrange prices.
And the amount of people proclaiming that 'oh that's not bad' or even 'that's actually good' around the web is properly scary to me. It's crazy how easily people are duped.
DegustatoR
Legend
Die size have no relation to market positioning.
The thing that literally dictates the cost of the processor has no relation on what 'market position' a GPU is put on?
I know you're not dumb enough to actually believe that, so as usual, you're just resorting to desperate, dishonest arguments to try and defend Nvidia(or your own 40 series purchase).
By your reasoning, there's literally no reason a 4090 should cost more than a 4060, right? Come the frick on dude. Who are you trying to fool here? What is wrong with you?
DegustatoR
Legend
The thing dictates relative cost of production of a chip on the same process and nothing else. It doesn't say anything about it's market positioning or correlate in any way to another chip made on a different process.
You'd think that people here would be aware of such simple truths.
The rest of your personal attacks just shows how insecure you're in your arguments.
This sounds like a language barrier to me. DegustatoR, "It doesn't say anything" doesn't mean "It's not the main factor". It means "has no impact whatsoever, directly or indirectly, on the matter". I think it's obvious that die size isn't the only thing, or even the main thing, affecting the final price of a graphics card and I think that's what you are trying to say. You can't possibly mean that there is no *indirect* connection between die size and pricing... no company would sell a $400 chip for $200, they'd just cancel the product.
When you say "[doesn't] correlate in any way to another chip made on a different process", the production cost is very roughly "die size * yields * wafer cost", so die size isn't enough to say anything without knowing the other factors, but there *is* a correlation to the cost. And similarly, the cost isn't the only factor or even the main factor affecting positioning in some cases, but it is *one* of the factors.
You know what they say... Only a Sith deals in absolutes!
EDIT: And obviously TSMC 4nm wafers are more expensive now than TSMC 12nm wafers, even if you look at their relative cost at the same point in the process lifetime. But some of what is making these GPUs bigger is larger caches that are allowing them to use cheaper memories, and DDR4/DDR5 prices have gone down a lot due to oversupply, so I suspect GDDR6 prices have gone down a bit too. Therefore, I'm not convinced the bill of materials has gone up as much as the wafer cost would imply, but interestingly if the larger L2 cache reduces the bill of material overall (so in terms of raw profits it's a win) but replaces "low margin DRAM" with "high margin GPU" then you'd expect the board price to go up rather than down for the same gross margins despite higher gross profits... this kind of thing is really complicated and it's good to avoid absolute statements unless you have a lot of insider knowledge.
When you say "[doesn't] correlate in any way to another chip made on a different process", the production cost is very roughly "die size * yields * wafer cost", so die size isn't enough to say anything without knowing the other factors, but there *is* a correlation to the cost. And similarly, the cost isn't the only factor or even the main factor affecting positioning in some cases, but it is *one* of the factors.
You know what they say... Only a Sith deals in absolutes!
EDIT: And obviously TSMC 4nm wafers are more expensive now than TSMC 12nm wafers, even if you look at their relative cost at the same point in the process lifetime. But some of what is making these GPUs bigger is larger caches that are allowing them to use cheaper memories, and DDR4/DDR5 prices have gone down a lot due to oversupply, so I suspect GDDR6 prices have gone down a bit too. Therefore, I'm not convinced the bill of materials has gone up as much as the wafer cost would imply, but interestingly if the larger L2 cache reduces the bill of material overall (so in terms of raw profits it's a win) but replaces "low margin DRAM" with "high margin GPU" then you'd expect the board price to go up rather than down for the same gross margins despite higher gross profits... this kind of thing is really complicated and it's good to avoid absolute statements unless you have a lot of insider knowledge.
Last edited:
DegustatoR
Legend
There is always an indirect (or very direct sometimes) connection between a cost of production and pricing.
There is no connection whatsoever between a die size and a market segmentation though. A certain die size does not imply anything on the level of price of a product which is using this die. Partially because the die isn't the only component of a product but mostly because the actual cost of a die differ wildly between not only production processes used but also companies producing them sometimes (prices can be different for different customers).
Saying that a die is somehow "low end" because it's just 200mm^2 is completely disingenuous to the whole discussion.
It is a long known fact now that these new advanced processes are rarely providing any price per transistor improvements and are on fact more expensive on that metric more often than not.
AD106 will probably end up being about 24B of Ts? This puts it closer to GA102's 28.3B than GA104's 17.4B. Do any of these dies look "low end" to anyone?
True if a company is willing to lose money on a product.
Otherwise, bold is used to point out the obvious connection.
DegustatoR
Legend
For the third time: there's a connection between a cost of production and a product's price. But there are no connections between a die size and a product price.
In other words you cannot make any assumptions about a product's price from a chip die size which is used in that product.
Die size is direct part of cost of production and thus connected to product price.
DegustatoR
Legend
Sure. But it doesn't tell us anything about how high or low that price "should" be.
- Status
