Wednesday, November 4th 2020

NVIDIA Reportedly Working on GeForce RTX 3080 Ti Graphics Card with 20 GB GDDR6X VRAM

A leak from renowned (and usually on-point) leaker Kopite7kimi claims that NVIDIA has finally settled on new graphics cards to combat AMD's RX 6800 threat after all. After the company has been reported (and never confirmed) to be working on double-memory configurations for their RTX 3070 and RTX 3080 graphics cards (with 16 GB GDDR6 and 20 GB GDDR6X, respectively), the company is now reported to have settled for a 20 GB RTX 3080 Ti to face a (apparently; pending independent reviews) resurgent AMD.

The RTX 3080 Ti specs paint a card with the same CUDA core count as the RTX 3090, with 10496 FP32 cores over the same 320-bit memory bus as the RTX 3080. Kopite includes board and SKU numbers (PG133 SKU 15) along a new GPU codename: GA102-250. The performance differentiator against the RTX 3090 stands to be the memory amount, bus, and eventually core clockspeed; memory speed and board TGP are reported to mirror those of the RTX 3080, so some reduced clocks compared to that graphics card are expected. That amount of CUDA cores means NVIDIA is essentially divvying-up the same GA-102 die between its RTX 3090 (good luck finding one in stock) and the reported RTX 3080 Ti (so good luck finding one of those in stock as well, should the time come). It is unclear how pricing would work out for this SKU, but pricing comparable to that of the RX 6900 XT is the more sensible speculation. Take this report with the usual amount of NaCl.
Sources: Kopite7kimi @ Twitter, via Videocardz
Add your own comment

136 Comments on NVIDIA Reportedly Working on GeForce RTX 3080 Ti Graphics Card with 20 GB GDDR6X VRAM

#126
AusWolf
purecain
Exactly, we're being forced to buy online due to covid and we have idiots using bots to buy what little stock there is, I'd of set my own bot up just to buy one for myself had i known it was going to be every single new product. This is going to have a terrible affect on the community. This would of been enough to put me off pc's all together when i was first starting. I cant stand dishonesty. Manufacturers should of made sure the public could get their products, we have shops using sister companies which basically prevent hardly any of us being able to buy at the msrp. If one shop had sold them at msrp no one would of been able to profit from scalped cards. That only works if people cant get them anywhere. It feels artificial, like this was planned and i'm sure it was. Call me paranoid. :nutkick:
I'm sure scalpers are part of the problem, but I also think they're only a small part of it. Interestingly, I never heard so much about scalping during any of the previous hardware launches.

Personally, I think the real problem is Samsung's 8 nm node not being able to produce a reasonable number of GPUs for nvidia and board partners. Random fact: there is not a single fully unlocked GPU die either in the current, or rumoured 30-series product stack anywhere. Why is that? Is nvidia making room for a later coming Super series? Maybe, but I think yields are so low that they basically can't produce fully functional dies at all, and even partially functional ones are short on supply - that's probably why we're hearing rumours about the large GA102 seeping down into 3070 Ti territory and the GA104 into 3060 Ti, or even 3060 levels (instead of introducing new, smaller dies as usual).

In short: scalping is only ever an issue when there is a pre-existent shortage of a certain product.
Posted on Reply
#127
purecain
I'm just guttered, usually we would all be overclocking and benchmarking our new cards and others would be making their minds up based on our experiences. This is usually a lot of fun for our community. Lets pray for decent amounts of stock over the next month so we can at least get set up for xmas, although i have a feeling we will all still be hunting parts into February 2021.
Posted on Reply
#128
lexluthermiester
purecain
Lets pray for decent amounts of stock over the next month
They are working on it. Shipments are incoming.
Posted on Reply
#129
Makaveli
RedelZaVedno
That's how it is in duopoly which was illegal not that long ago. We need working antitrust legislation ASAP.
Good luck with that let me know how it goes.
Posted on Reply
#130
ratirt
AusWolf
Personally, I think the real problem is Samsung's 8 nm node not being able to produce a reasonable number of GPUs for nvidia and board partners. Random fact: there is not a single fully unlocked GPU die either in the current, or rumoured 30-series product stack anywhere. Why is that? Is nvidia making room for a later coming Super series? Maybe, but I think yields are so low that they basically can't produce fully functional dies at all, and even partially functional ones are short on supply - that's probably why we're hearing rumours about the large GA102 seeping down into 3070 Ti territory and the GA104 into 3060 Ti, or even 3060 levels (instead of introducing new, smaller dies as usual).
I don't think so. 8nm Samsung is 2 years old news. How is that holding back production I have no idea since the yields are good.
If it really was Samsung's node, and the yields would be low due to defects for the 3090 and 3080, You could salvage those to get decent number of 3070 GPUs and these have problems with stock as well. Problem is elsewhere.
Posted on Reply
#131
fynxer
320bit mem bus must be fake news and Trump won the election if you only count his votes :)

Why 320bit mem bus? If people are going to pay +$300 they would want some upgrade to the mem bus like 352bit.

What ever! If I feel like Nvidia is screwing me I will jump to AMD
Posted on Reply
#132
AusWolf
ratirt
I don't think so. 8nm Samsung is 2 years old news. How is that holding back production I have no idea since the yields are good.
If it really was Samsung's node, and the yields would be low due to defects for the 3090 and 3080, You could salvage those to get decent number of 3070 GPUs and these have problems with stock as well. Problem is elsewhere.
My theory is that the originally planned 3090 and 3080 produced defects, that's why we have those awkward numbers of cuda cores in those GPUs, and basically no fully unlocked chip anywhere in the product stack. Even the 3070 uses a GPU with disabled parts, and I'm pretty sure there isn't so much difference between the performance of 46 and 48 SMs that would make any sense for nvidia to have plans with the unlocked chip in the future. Even the 3070 Ti rumours suggest a highly watered-down version of the GA102 instead of a full GA104. If it is financially beneficial to manufacture such a large chip and disable parts for a card several tiers below the 3090, I can only assume that something is awfully wrong with the node or the design (or both).
Posted on Reply
#133
ratirt
AusWolf
My theory is that the originally planned 3090 and 3080 produced defects, that's why we have those awkward numbers of cuda cores in those GPUs, and basically no fully unlocked chip anywhere in the product stack. Even the 3070 uses a GPU with disabled parts, and I'm pretty sure there isn't so much difference between the performance of 46 and 48 SMs that would make any sense for nvidia to have plans with the unlocked chip in the future. Even the 3070 Ti rumours suggest a highly watered-down version of the GA102 instead of a full GA104. If it is financially beneficial to manufacture such a large chip and disable parts for a card several tiers below the 3090, I can only assume that something is awfully wrong with the node or the design (or both).
Sure but you need to realize that the bigger the chip is, the harder it is to make it. Of course there will be defects and you can use that chip as a lower tier. The fact that NV chips are quite big is no fantasy and that would contribute to defective chips. So the quantity might not be as impressive as a smaller chip but that's the way it is. The node is matured and it does well with all the improvements it had. (To be clear it is 10nm Samsung improved becoming 8nm) So I don't think the problem is the node. Maybe memory if you consider 3080 and 3090.
The other thing is, maybe NV was rushing the release so bad, they didnt have the stock to begin with. Maybe these cards were going to be released later not August but much later. They didnt have the stock but wanted to release it before AMD. Hard to tell but the node problem or yields problem (they are there due to large chips) is not convincing me.
Posted on Reply
#134
r9
What a ground breaking news ... for the two people that will be able to buy it.
Posted on Reply
#135
londiste
AusWolf
My theory is that the originally planned 3090 and 3080 produced defects, that's why we have those awkward numbers of cuda cores in those GPUs, and basically no fully unlocked chip anywhere in the product stack. Even the 3070 uses a GPU with disabled parts, and I'm pretty sure there isn't so much difference between the performance of 46 and 48 SMs that would make any sense for nvidia to have plans with the unlocked chip in the future. Even the 3070 Ti rumours suggest a highly watered-down version of the GA102 instead of a full GA104. If it is financially beneficial to manufacture such a large chip and disable parts for a card several tiers below the 3090, I can only assume that something is awfully wrong with the node or the design (or both).
Has Big Navi die size been confirmed?
Rumored 536mm² is only 15% smaller than Nvidia's 628mm². TSMC N7 is a smaller, newer and more complex node compared to Samsung's 8N.
Posted on Reply
#136
Zach_01
wolf
But, lets for arguments say you're totally right and they actually "wipe the floor" product vs product, by basically offering what is at best a respectable price to performance benefit, with some extra placebo VRAM and a minor perf;watt advantage. Nvidia adjust prices, actually officially announces SKU's that fit the gaps, and life goes on with both happily selling GPU's to happy consumers.
Its not placebo... they need it for SmartAccessMemory to increase performance even higher for their own platform. How good is it or not, or how this will be exploited by game devs in future is yet to be seen.
Posted on Reply
Add your own comment