Friday, September 23rd 2022
NVIDIA AD103 and AD104 Chips Powering RTX 4080 Series Detailed
Here's our first look at the "AD103" and "AD104" chips powering the GeForce RTX 4080 16 GB and RTX 4080 12 GB, respectively, thanks to Ryan Smith from Anandtech. These are the second- and third-largest implementations of the GeForce "Ada" graphics architecture, with the "AD102" powering the RTX 4090 being the largest. Both chips are built on the same TSMC 4N (4 nm EUV) silicon fabrication process as the AD102, but are significantly distant from it in specifications. For example, the AD102 has a staggering 80 percent more number-crunching machinery than the AD103, and a 50 percent wider memory interface. The sheer numbers at play here, enable NVIDIA to carve out dozens of SKUs based on the three chips alone, before we're shown the mid-range "AD106" in the future.
The AD103 die measures 378.6 mm², significantly smaller than the 608 mm² of the AD102, and it reflects in a much lower transistor count of 45.9 billion. The chip physically features 80 streaming multiprocessors (SM), which work out to 10,240 CUDA cores, 320 Tensor cores, 80 RT cores, and 320 TMUs. The chip is endowed with a healthy ROP count of 112, and has a 256-bit wide GDDR6X memory interface. The AD104 is smaller still, with a die-size of 294.5 mm², a transistor count of 35.8 billion, 60 SM, 7,680 CUDA cores, 240 Tensor cores, 60 RT cores, 240 TMUs, and 80 ROPs. Ryan Smith says that the RTX 4080 12 GB maxes out the AD104, which means its memory interface is physically just 192-bit wide.
Sources:
Ryan Smith (Twitter), VideoCardz
The AD103 die measures 378.6 mm², significantly smaller than the 608 mm² of the AD102, and it reflects in a much lower transistor count of 45.9 billion. The chip physically features 80 streaming multiprocessors (SM), which work out to 10,240 CUDA cores, 320 Tensor cores, 80 RT cores, and 320 TMUs. The chip is endowed with a healthy ROP count of 112, and has a 256-bit wide GDDR6X memory interface. The AD104 is smaller still, with a die-size of 294.5 mm², a transistor count of 35.8 billion, 60 SM, 7,680 CUDA cores, 240 Tensor cores, 60 RT cores, 240 TMUs, and 80 ROPs. Ryan Smith says that the RTX 4080 12 GB maxes out the AD104, which means its memory interface is physically just 192-bit wide.
152 Comments on NVIDIA AD103 and AD104 Chips Powering RTX 4080 Series Detailed
Samsung's 8LPP has similar frequency potential as N10 but even better density (around +17%)
16nm GP104 has 7.2b transistors at 314mm² while GA104 has 17.4b transistors at 392mm² (1.9X more dense not in logic only but on overall design)
---------------------------------
New info regarding Ada Lovelace features
"DLSS 3 combines Super Resolution, Frame Generation, and NVIDIA Reflex!"
This is something that had never been seen before.
Compare AL to AMP:
chip | die size | pricing
AD102 608mm^2 1600
AD103 379mm^2 1200
AD104 295mm^2 900
GA102 628mm^2 2000
GA102 628mm^2 1500
GA102 628mm^2 1200
GA102 628mm^2 700
GA104 392mm^2 600
GA104 392mm^2 500
GA104 392mm^2 400
GA106 276mm^2 330
GA106 276mm^2 250
GA107 200mm^2 200
JHH literally priced something that should cost around 350 max 2.6 times higher up to 900!!
Like I said in previous post, 4080 12GB IS overpriced, but its only like $200 give or take from some of the older x80 G104 cards with full enabled dies (Adjusted inflation).
The only problem I can see from here is how the future 4070/4060 will be priced. Those won't scale as linearly with inflation metrics... Not that the 4080 12GB does, but theres compensation on card "upgrades" and EE requirements at least.
My assumption? RTX30 die cost/Samsung 8 was insanely cheap. Nvidia priced and positioned it to be a saving grace to how horrible RTX20 was for general rasterization improvement from Pascal.
Turing had HUGE dies and Nvidia ate so much well deserved crap because RTX/DLSS was essentially vaporware. Generational gains were also weak relative to 10/9 series. Of course most logical people skipped it.
and the 4080's not cheap either
Why didn't nGreedia used the proper naming convention for their GPUs, such as:
AD102 = RTX 4080 (NOT RTX 4090)
AD103 = RTX 4070 (NOT RTX 4080 16 GB)
AD104 = RTX 4060 (NOT RTX 4080 12 GB)
To be honest those could also be the Ti variants, comparing to previous iterations, or just some bump up specs.
Their naming convention is totally retarded and BS tbh.
TIME TO ACCEPT THEY ARE GREEDY, buy AMD, as far as multi-billion corps that just want your cash go, they are your only choice.
Sad times, but hey ho.
Remember, both 3090 and 3080 10G were announced September 1, 2020. 3080 Ti was announced on May 31, 2021. For nine months, there was 3090 or 3080 10G and nothing in between. The 3080 12G didn't arrive until January of this year.
Will there be a 4080 Ti at some point? Probably. Looking at the numbers of cores (CUDA, RT, Tensor) and pricing in the three Ada Lovelace cards announced, it appears that NVIDIA can slot in offerings around these three (above, in between, below).
NVIDIA did not release the entire Ampere lineup at the same time. People will have more choices if they choose to be patient whether it be additional NVIDIA 40 series models or whatever AMD releases. And like NVIDIA, AMD did not release the entire RNDA2 lineup at the same time.
They will be the last GPU to offer this port in...2024 on 5000 series, unless Nvidia consumers send them a clear message of disapproval and demand this upgrade on Ti cards.
Intel's A380, the lowest tier card, has DP 2.0 port at 40 Gbps.
Today you can grab a 6700XT for $350 with promo code +rebate off Newegg... lol. Both are scum.
but my point is: the 3080 had a core count and thus performance similar to the 3090, as the 2080ti & the titan rtx, et cetera.
In any case, it isn't worth comparing different generations of NVIDIA cards strictly by model number since they change how they think about them. It's really just a way to rank the various models within the product stack of a given generation.
Especially with Ada Lovelace generation, NVIDIA was motivated to change the numbering system yet again to make sure the Ada 4080 has performance above the Ampere 30 Series cards that they will continue selling alongside the new releases in order to draw down channel inventory.
So if i read this correct it affects the 3 series that AIB's need to have an extra conversion chip included. The 5 and 7 series per Ryan Shrout natively include this and thus should be 2.1 compliant. Just seems like an odd omission even on budget cards. HDMI 2.1 is not exactly new and AMD's 6400 card even supports it.
And Samsung 8N is a part of 10N series. It's 10N+++ speaking plain former Intel language.
Wow! I cannot believe that Intel waited for so long to tell us that only A750 and A770 Limited Edition cards have PCON converter chip on PCB for HDMI 2.1 FRL signal (unclear if 24Gbps?, 32 Gbps?, 40 Gbps? or 48 Gbps?), and all other cards are HDMI 2.0 with 18 Gbps speed.
This is what happens when HDMI Administrator publishes their decision to rebrand 2.0=2.1. And companies sell us a total mess...
Anyone can use either port and match it with whatever monitor or TV port they have at home.
Why not use HDMI 2.1 if someone has 4K/120 OLED TV? None of TVs has DisplayPort.
HDMI 2.1 port is currently superior in terms of bandwidth until DP 2.0 ports start working finally and become more mainstream.
Nvidia customers will unfortunately not enjoy this pleasure until 2024 and 5000 series, as the company has disappointed big time with omission of DP 2.0 on cards that cost up to $1600. Shambles. So much for promoting innovative connectivity technologies.
AD104 chip could cost 2x as much as GA102 chip that are on 3090Ti/3090/3080Ti/3080
That means Ampere will have the cost benefit for now while ADA will have performance and efficiency benefit.
Instead, they could have made more significant IPC/architectural improvements and backported AL to Samsung N8, so that the lineup would look something like this:
chip | die size | pricing
GAL104-2 750mm^2 1200
GAL104-2 750mm^2 1000
GAL104-2 750mm^2 800
GAL106-2 520mm^2 660
GAL106-2 520mm^2 500
GAL107-2 400mm^2 400