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
Also, look at how AMD manages to solve the problem with the high wafers cost. It's called multi-chip module first, and second - AMD uses the less expensive N5:
AMD's main Navi 31 GPU reportedly has 96MB of Infinity Cache, Navi 32 is up to 7680 cores - VideoCardz.com
GCD 308mm^2 on N5 + 6 x MCD 37.5mm^2 on N6.
:D :D
Plus, no consumer group has ever got organised to mount a legal challenge against HDMI Administrator for changing naming scheme without educating consumers properly and mandating vendors to be more transparent with features and signalling in advertisements.
Perhaps you could be the first one to organise consumer group and a legal team?Exactly all those things make it cheaper and ensure better margins for AIB, so there is more room for price tuning, unlike Ada.
MORE bad news for NVIDIA's 40 Series...
The video is by a youtube tech commentator JayzTwoCents so maybe his info is good or maybe it isn't, but it does raise questions.
If Samsung 8LPP wafer cost was $4000 like some posters said, this means $50 or less per die on a 12" wafer. Even if TSMC 4N has 3X the wafer cost we are talking $150 or less.
In AD103 and AD104 cases the wafer cost alone is even less a deciding factor in relation with GA104/GA106 if you factor the insane price increase.
That's the moment when any vendor could publish on their website that their old 2.0 port is 2.1 port. This created an explosion of deceptive marketing, from motherboards to CPUs and GPUs. Vendors do not seem obliged to mention HDMI 2.1 TMDS (2.0 at 18 Gbps) or HDMI 2.1 FRL (real 2.1 up to 48 Gbps). Some of course do and thjey are honest.
TFT Central wrote an expensive article about it in December 2021. Have a look when you get a spare moment.
On Intel's official website, the situation is still confusing and they should be called out for this by tech community. Here is the screen shot of the official spec.
As you can see, both HDMI 2.0b and HDMI 2.1 have a star "*" attached to it, but at the bottom of the page this star small print is nowhere to be found. Intel, WTF?! DP has two stars "**" and there is a clear explanation. This is what always makes me suspicious that something is being hidden from the public.
They should simply say the following, for the sake of simplicity and transparency:
*HDMI 2.0b TMDS 18 Gbps - A380 and A580 cards
*HDMI 2.1 FRL 40/48 Gbps - installed on founders A750 and A770 Limited Edition only; AIBs have an option to add FRL support via PCON
There are two ways to get FRL signal on GPU:
1. install native HDMI 2.1 chip on PCB (Nvidia and AMD GPUs)
2. install PCON chip on PCB that is fed by DP1.4 signal and then converted into HDMI FRL protocol (A750 and A770)
Why do you care about how the product name stack against previous gen as long it is fits your needs? How the die size effecting the usege of the product??
I can understand your point if you need to decide whether to buy now of a fetures in 6-12 monthIf you but gpu by the die size than, well, good luck.
I advise to buy according to pragmatic factors- price to preformance in a given budget factoring spacific needs (say you must have cuda cors or need free sync).
You factor your felling s about the market status and the company general pricing. That might get you a bad deal, when involving your fellings in it.
For sure, bigger dies will generate more heat which requires bigger and more expensive thermal solutions. Even if you're not taking die physical dimensions directly into account, most likely you are looking indirectly while observing power specifications.
The same way you don't shop by memory-bus and, in most situations, not by memory size.
This is why Big chip's have a limited future, the yield rate's don't go up, only down with node swaps.
For my primary gaming build, I went for something with high performance, performance-per-watt metric be damned. If it's not as efficient as another system, that's fine while I'm gaming.
And I don't run my gaming computer 24x7. A lot of my desktop computing (like answering your post) is done on a Mac mini 2018 -- not on the build described in my System Specs. If I'm just surfing the 'net (couch, bed, etc.) I often just pick up my iPad.
Besides, instead of blindly accepting what I see on price tags, I like at least trying to understand why things are the way they are. Is that bad?But new nodes aren't new forever, so maybe it's time for the industry to sit back a bit and wait for these nodes to come down in costs? They surely can't double the price of every new generation of GPUs forever.
Realising an item's real worth vs. what the price tag says is part of what makes you an informed consumer. Just because something costs $900 it doesn't mean it's actually worth $900, even if you can afford it.
In other words, ~80% more than the Ampere cores (CUDA, RT, Tensor) and the added performance of a new generation. They can increase now because the disadvantage of the manufacturing node is no longer there.
The RTX 4090 promises to be a beast.