RUMORS: NVIDIA Ampere GPU Massive Die Size, Specifications, Architecture and More! – Wccftech

Notice those big bold words at the start of the article? If completely unverified leaks aren’t your piece of cake – it would be best to stop reading now. For the rest of us, however, here is an article full of rumors galore – most of which come courtesy of KittyCorgi over at twitter. The reason why this is tagged a rumor is because this particular account has no history of leaks and has only posted a couple of tweets so far – all of which haven’t had the chance to be verified. While all this could turn out to be utter tosh, we did think this was worth covering for the connoisseurs among us. At this point, it might be best to treat this article as an exercise in wish-fulfillment rather than an actual leak.

NVIDIA Ampere GPU ‘GA100’ flagship will have massive 826mm² die

NVIDIA’s upcoming Ampere GPUs have been the subject of their fair share of hype and anticipation. If this particular rumor is to be believed, however, NVIDIA’s flagship GPU will have a massive 826mm² die. While the process node was not mentioned, a die this large does slightly increase the probability of it being an older, more mature node like the 12NFF. Depending on how mature TSMC 7nm is, an 826mm² large die isn’t totally out of the question for limited quantity runs as well. Previous leaks have pointed out that the company is going to shift to 7nm with Ampere so I would assume that this entire architecture is based on 7nm for the duration of the article.

An 826mm² die on 7nm would be absolutely enormous and insanely dense. This wouldn’t be entirely uncharacteristic of NVIDIA considering the company has pushed the power and thermal envelopes before with their high-end parts (that usually end up in workstations and rarely the TITAN series). A GPU of this magnitude would easily decimate anything AMD has planned for the remainder of 2020 considering we haven’t heard of anything that could match this.

NVIDIA Ampere GPUs will have 2x faster RTX and a re-engineered Tensor core

Here is where things get interesting, however, the rumor includes a block diagram (made by KittyCorgi I assume) which indicates improvements in the architecture and NVIDIA doubling down on their RTX strategy. The last part isn’t particularly surprising as the RTX portion on Turing was relatively tiny and limited the scale of application for RTX games.

Here is the list of improvements suggested by the rumor:

• INT32 Unit remains unchanged.
• Double the FP32 Unit for shader proportion.
• The performance of the new Tensor Core is doubled.
• Enhanced L1 Data Cache for more comprehensive functions.
• True architecture for RTX GAMING with all-new design RT CORE ADVANCED.
  
  

The last bit is something that sounds straight out of NVIDIA marketing material and is something that makes me think that this just might turn out to be true. RT Core Advanced is also a very plausible naming scheme coming from what appears to be a non-native English speaker. Consider me intrigued at the very least although I would still advise copious amounts of salt to go with this.

NVIDIA RTX 3080 Ti ‘GA103’ GPU will have 60 SMs and 320-bit 10GB/20GB Graphics Memory

The rumor doesn’t end there, the tweeter has also made block diagrams of the GA103 and GA104 GPUs – which would be the graphics processors that the mainstream client will actually get after launch. These are the parts that a gamer will be able to buy without going completely bankrupt. For the higher end GA103 part, which will end up being the RTX 3080 Ti or whatever NV decides to call it, you are looking at 60 SMs for a total of 3840 CUDA cores assuming the old proportion or 7680 CUDA cores assuming a new 128 cores per SM ratio (thanks for the flag, Boro and PCMasterRace!) For reference, the RTX 2080 Ti has 72 SMs with 4608 CUDA Cores.

Coupled with the doubled RTX core performance (and the resulting increase in DLSS throughput) you are easily looking at real-world performance increase in the range of 50% over Turing. If NVIDIA shifts to a 7nm process, they would be able to manage all this while staying well within acceptable power thresholds. The GA103 Ampere GPU will be coupled with 10GB/20GB vRAM.

NVIDIA RTX 3080 ‘GA104’ GPU will have 48 SMs and 256-bit 8GB/16GB Graphics Memory

Up next, you have the GA104 GPU which should end up in the RTX 2080 replacement (the RTX 3080?) with 48 SMs (3072 CUDA cores based on old proportion, 6144 new). This is very slightly more than the RTX 2080 at 46 SMs. Coupled with higher performance throughput and the doubled RTX cores you are once again looking at significant performance increases if this turns out to be true. According to the rumor, the RTX 3080 GPU will be coupled with 8GB/16GB of vRAM and a 256-bit bus width.

While I don’t know if this rumor has any truth to it (as a gamer, I wish it does), it does make for a very interesting read with a cup of tea. One thing is for sure, the hype for NVIDIA’s Ampere GPU is at an all-time high and everything points to the company getting ready to launch those at GTC 2020 this year with Jensen personally inviting the press to attend.

Share
Tweet
Submit