Nvidia's GeForce RTX 3060 boasts 12GB of VRAM and 3,584 CUDA cores for $329
- Thread starter Polycount
- Start date
defaultluser
Posts: 527 +397
Precisely , back in the days of Kepler and Maxwell, NVIDIA experimented around with mixed memory bus sizes to capacity. The GTX 660 was 1.5GB (6 chips) + 0.5 GB (2 chips) = a standard 2GB VRAM (even if you can only access a fraction of the memory bus at-once), This meant you could use the exact sam e PCB layout for a GTX 670 as you would for the GTX 660 Ti.
But after the lawsuit with the GTX 970 3.5 + 0.5 GB ,they decided to ditch that.
Unfortunately, you're now stuck with a continuous memory space, with means multiples of 6 chips on a 192-bit bus, or multiples of 8 chips on a 256-bit bus. That's going to cause massive VRAM size mismatches just like the the 8800 GTX 768MB on a 384-bit bus versus the 8800 GTS 512 om 256,
You're also going to feel the pain when Using two different memory densities on older versus ewer products, NVIDIA got shoved in this hole by GDDR6X, and we knew it would be painful when they finally jumped
If I had to guess why, the performance of the 3060 is around RTX 2070, so if the dropped that down to 6GB VRAM, it would be a waste pf a new product - so they launched it first over the 3070 with double-density.
My GTX 1060 6GB can't max-out DOOM ETERNAL at 1080p Ultra Nightmare, - even though it's runs at over80 fps at ultra-nightmare effects and nightmare textures. NVIDIA would have been completely stupid to launch a 6GB $300 video card in 2021..but no worries folks, the doubling of densities are eventually coming to the rest of the product line!
Last edited:
128-bit bus = four 32-bit memory controllers
192-bit bus = six 32-bit memory controllers
256-bit bus = eight 32-bit memory controllers
384-bit bus = twelve 32-bit memory controllers
Micron, Samsung, SK Hynix, et al manufacture GDDR6 with 8 or 16 Gb densities, and in the case of GDDR6X, from Micron, just 8 Gb at the moment. There are no other densities available.
GDDR6/6X operates so that a single DRAM module is connected to a single controller, in 32-bit mode, or two modules per controller, each in 16-bit mode. So the memory configurations possible are:
128-bit bus (4 or 8 modules)
8Gb modules = 4 or 8 GB
16Gb modules = 8 GB or 16 GB
192-bit bus (6 or 12 modules)
8Gb modules = 6 or 12 GB
16Gb modules = 12 GB or 24 GB
256-bit bus (8 or 16 modules)
8Gb modules = 8 or 16 GB
16Gb modules = 16 or 32 GB
384-bit bus (12 or 24 modules)
8Gb modules = 12 or 24 GB
16Gb modules = 24 or 48 GB
AIB vendors will want to use as few DRAM modules as they can, to keep costs down. The higher density modules are roughly twice the price of the lower ones, when purchased from retailers, although they'll be cheaper than this when bought direct. AMD have gone with 16Gb for their RX 6800 lineup, Nvidia is stuck with 8Gb for GDDR6X.
There's a small but not insignificant price difference between module speeds too, roughly $5 or so, although it varies as according to supply and demand. AMD were obviously happy to absorb this by choosing the fastest, highest density GDDR6 modules for their 6800 series, but now the consumer is going to expect this to be the norm for all future products.
I still think Nvidia's original plan for the 3060 was for it to have 6 GB of 14Gbps 8Gb GDDR6, but once they got word of what AMD were planning with their new products, they switched to 16Gb modules. That way, when AMD releases their mid-range/budget RDNA 2 cards, with 128-bit memory buses, they'll have less memory than the 3060 and thus offer Nvidia a marketing advantage (unless AMD are going with a 192-bit bus too).
192-bit bus = six 32-bit memory controllers
256-bit bus = eight 32-bit memory controllers
384-bit bus = twelve 32-bit memory controllers
Micron, Samsung, SK Hynix, et al manufacture GDDR6 with 8 or 16 Gb densities, and in the case of GDDR6X, from Micron, just 8 Gb at the moment. There are no other densities available.
GDDR6/6X operates so that a single DRAM module is connected to a single controller, in 32-bit mode, or two modules per controller, each in 16-bit mode. So the memory configurations possible are:
128-bit bus (4 or 8 modules)
8Gb modules = 4 or 8 GB
16Gb modules = 8 GB or 16 GB
192-bit bus (6 or 12 modules)
8Gb modules = 6 or 12 GB
16Gb modules = 12 GB or 24 GB
256-bit bus (8 or 16 modules)
8Gb modules = 8 or 16 GB
16Gb modules = 16 or 32 GB
384-bit bus (12 or 24 modules)
8Gb modules = 12 or 24 GB
16Gb modules = 24 or 48 GB
AIB vendors will want to use as few DRAM modules as they can, to keep costs down. The higher density modules are roughly twice the price of the lower ones, when purchased from retailers, although they'll be cheaper than this when bought direct. AMD have gone with 16Gb for their RX 6800 lineup, Nvidia is stuck with 8Gb for GDDR6X.
There's a small but not insignificant price difference between module speeds too, roughly $5 or so, although it varies as according to supply and demand. AMD were obviously happy to absorb this by choosing the fastest, highest density GDDR6 modules for their 6800 series, but now the consumer is going to expect this to be the norm for all future products.
I still think Nvidia's original plan for the 3060 was for it to have 6 GB of 14Gbps 8Gb GDDR6, but once they got word of what AMD were planning with their new products, they switched to 16Gb modules. That way, when AMD releases their mid-range/budget RDNA 2 cards, with 128-bit memory buses, they'll have less memory than the 3060 and thus offer Nvidia a marketing advantage (unless AMD are going with a 192-bit bus too).
It’s looks slower than a 2070 super. Maybe the ray tracing will be quicker but the rasterisation looks between the 2070 and 2070 super. I felt an improvement with the 3060ti over the 2070 super but that card is like a 2080 super.
AMD has forced Nvidia to add more RAM to their cards. Now 3070 looks really silly with the same amount of RAM as the old Radeon 580 
Wait... the ancient Radeon R9 390 had 8 GB RAM as well...
That was like 5 generations ago
Wait... the ancient Radeon R9 390 had 8 GB RAM as well...
That was like 5 generations ago
defaultluser
Posts: 527 +397
AMD has forced Nvidia to add more RAM to their cards. Now 3070 looks really silly with the same amount of RAM as the old Radeon 580
Wait... the ancient Radeon R9 390 had 8 GB RAM as well...
That was like 5 generations ago
And then all their gaming-focused high-end video cards have been holding the phone at 8GB VRAM for the last five years
Vega 64 = 8GB ram, RX 5700 XT = 8 GB ram,
The reason AMD could get away with riding that wave for.half a decade, is because they kept improving memory compression. They were also forced to by memory manufacturers, who have hit the wall on DRAM process shrinks. around 2016:
https://blocksandfiles.com/2020/04/13/dram-is-stuck-in-a-10nm-process-trap/
Without DRAM shrinks, it's harder to double the density of all these GDDR memory chips, so be amazed that we managed 16 Gb in just 5 years.
AMD just had good timing on mass-availability pf these chips, so of course NVIDIA has to react while they must wait a few more months for the 20GB 3080. I mean, The PS5 console's 256-bit bus is using 16 GDDR6 chips , because these are so new!
Last edited:
Tom Sunday
Posts: 81 +13
The Mumbai boys at the local stripcenter PC shop are waiting for another load this March of used 980ti, 1080ti and 2080's all from their network of 'Miners.' I was told that the 'cryptos' will be selling like hotcakes and mostly to the man on the street. A used Miner 1080ti will probably go for $65. For people like me with rather shallow pockets, a NVIDA dream may come true after all. I may go for a stretch and get a RTX 2080 for hopefully a little over a $100, but that if the company allows me some overtime driving a forklift. I am most definitely ready to experience Metro Exodus and meet the Aurora Team at Novosibirsk station. Life is good and interesting times ahead.
Tom Sunday
Posts: 81 +13
Yes I will need luck as it has always been in my illustrious life. But "Crypto Cards" have always pulled quite a bit more 24/7 heavy duty online then the average gamer sitting at home. Thus less pricey. The Mumbai boys and their shop is in LA but not exactly in a preferred drive too area. But for deals many come anyway. Madhu from Varanasi said the 'perfect storm' for used 1080's and 2080's (non-cryptos) will be around late April/May and once Asus, MSI, Gigabyte, EVGA, Zotac and many others bringing their custom '3000' series to bear and start flooding the now stale market. A base 3080 will run $699 and let's see what that will do for a used 2080ti? That will be the test of time for many and the people I know who initially shelled-out $1,450 for a 2080ti MSI Trio Gaming. I am on the other hand find myself on the lower end of the food chain and "Crypto" will unfortunately stay and remain my middle name.
Last edited:
Similar threads
