NVIDIA Scrubs GeForce RTX 4080 12GB Launch; 16GB To Be Sole RTX 4080 Card
by Ryan Smith on October 14, 2022 1:20 PM EST- Posted in
- GPUs
- GeForce
- NVIDIA
- Ada Lovelace
- RTX 4080
In a short post published on NVIDIA’s website today, the company has announced that it is “unlaunching” their planned GeForce RTX 4080 12GB card. The lowest-end of the initially announce RTX 40 series cards, the RTX 4080 12GB had attracted significant criticism since it’s announcement for bifurcating the 4080 tier between two cards that didn’t even share a common GPU. Seemingly bowing to the pressure of those complaints, NVIDIA has removed the card from their RTX 40 series lineup, as well as cancelling its November launch.
NVIDIA’s brief message reads as follows:
The RTX 4080 12GB is a fantastic graphics card, but it’s not named right. Having two GPUs with the 4080 designation is confusing.
So, we’re pressing the “unlaunch” button on the 4080 12GB. The RTX 4080 16GB is amazing and on track to delight gamers everywhere on November 16th.
If the lines around the block and enthusiasm for the 4090 is any indication, the reception for the 4080 will be awesome.
NVIDIA is not providing any further details about their future plans for the AD104-based video card at this time. However given the circumstances, it’s a reasonable assumption right now that NVIDIA now intends to launch it at a later time, with a different part number.
NVIDIA GeForce Specification Comparison | |||||
RTX 4090 | RTX 4080 16GB | RTX 4080 12GB (Cancelled) |
|||
CUDA Cores | 16384 | 9728 | 7680 | ||
ROPs | 176 | 112 | 80 | ||
Boost Clock | 2520MHz | 2505MHz | 2610MHz | ||
Memory Clock | 21Gbps GDDR6X | 22.4Gbps GDDR6X | 21Gbps GDDR6X | ||
Memory Bus Width | 384-bit | 256-bit | 192-bit | ||
VRAM | 24GB | 16GB | 12GB | ||
Single Precision Perf. | 82.6 TFLOPS | 48.7 TFLOPS | 40.1 TFLOPS | ||
Tensor Perf. (FP16) | 330 TFLOPS | 195 TFLOPS | 160 TFLOPS | ||
Tensor Perf. (FP8) | 660 TFLOPS | 390 TFLOPS | 321 TFLOPS | ||
TDP | 450W | 320W | 285W | ||
L2 Cache | 72MB | 64MB | 48MB | ||
GPU | AD102 | AD103 | AD104 | ||
Transistor Count | 76.3B | 45.9B | 35.8B | ||
Architecture | Ada Lovelace | Ada Lovelace | Ada Lovelace | ||
Manufacturing Process | TSMC 4N | TSMC 4N | TSMC 4N | ||
Launch Date | 10/12/2022 | 11/16/2022 | Never | ||
Launch Price | MSRP: $1599 | MSRP: $1199 | Was: $899 |
Taking a look at the specifications of the cards, it’s easy to see why NVIDIA’s core base of enthusiast gamers were not amused. While both RTX 4080 parts shared a common architecture, they did not share a common GPU. Or, for that matter, common performance.
The RTX 4080 12GB, as it was, would have been based on the smaller AD104 GPU, rather than the AD103 GPU used for the 16GB model. In practice, this would have caused the 12GB model to deliver only about 82% of the former’s shader/tensor throughput, and just 70% of the memory bandwidth. A sizable performance gap that NVIDIA’s own figures ahead of the launch have all but confirmed.
NVIDIA, for its part, is no stranger to overloading a product line in this fashion, with similarly-named parts delivering unequal performance and the difference denoted solely by their VRAM capacity. This was a practice that started with the GTX 1060 series, and continued with the RTX 3080 series. However, the performance gap between the RTX 4080 parts was far larger than anything NVIDIA has previously done, bringing a good deal more attention to the problems that come from having such disparate parts sharing a common product name.
Of equal criticism has been NVIDIA’s decision to sell an AD104 part as an RTX 4080 card to begin with. Traditionally in NVIDIA’s product stack, the next card below the xx80 card is some form of xx70 card. And while video card names and GPU identifiers are essentially arbitrary, NVIDIA’s early performance figures painted a picture of a card that would have performed a lot like the kind of card most people would expect from the RTX 4070 – delivering performance upwards of 20% (or more) behind the better RTX 4080, and on-par with the last-generation flagship, the RTX 3090 Ti. In other words, there has been a great deal of suspicion within the enthusiast community that NVIDIA was attempting to sell what otherwise would have been the RTX 4070 as an RTX 4080, while carrying a higher price to match.
In any case, those plans are now officially scuttled. Whatever NVIDIA has planned for their AD104-based RTX 40 series card is something only the company knows at this time. Meanwhile come November 16th when the RTX 4080 series launches, the 16GB AD103-based cards will be the only offerings available, with prices starting at $1199.
100 Comments
View All Comments
Kangal - Monday, October 17, 2022 - link
See below:(_% of Max Shaders - MSRP Price : Chipset Name ~ Shader Cores = Name of GPU Card)
100% - $1000 : GM-200-400 _3072 = 9-GTX Titan X
100% - $1200 : GP-102-450 _3840 == 10-Titan Xp
100% - $2000 : GA-102-350 _10752 ======= RTX 3090 Ti
100% - $2500 : TU-102-400 _4608 ==== 20-Titan RTX
~98% - $1500 : GA-102-300 _10496 ======= RTX 3090
~97% - $1600 : AD-102-300 _16384 =========== RTX 4090
~95% - $1200 : GA-102-250 _10240 ======= RTX 3080 Ti
~94% - $1000 : TU-102-300 _4352 ==== RTX 2080 Ti
~93% - $700 : GP-102-350 _3584 == GTX 1080Ti
~92% - $650 : GM-200-200 _2816 = GTX980Ti
~81% - $700 : GA-102-200 _8704 ======= RTX 3080
~80% - $1200 : AD-103-300 _9728 ========= RTX 4080
~67% - $550 : GM-204-300 _2048 = GTX980
~67% - $600 : GP-104-410 _ 2560 == GTX 1080
~67% - $800 : TU-104-450 _ 3072 ==== RTX 2080-S
~64% - $700 : TU-104-410 _ 2944 ==== RTX 2080
~63% - $450 : GP-104-300 _ 2432 == GTX 1070Ti
~58% - $900 : AD-104-400 _ 7680 =========== RTX 4080-12GB
~57% - $600 : GA-104-400 _ 6144 ======= RTX 3070 Ti
~56% - $600 : TU-104-400 _ 2560 ==== RTX 2070-S
~55% - $500 : GA-104-300 _ 5888 ======= RTX 3070
~54% - $330 : GM-204-200 _1664 = GTX970
~50% - $380 : GP-104-200 _ 1920 == GTX 1070
~50% - $500 : TU-106-410 _ 2304 ==== RTX 2070
~47% - $400 : TU-106-400 _ 2176 ==== RTX 2060-S
~45% - $400 : GA-104-200 _ 4864 ======= RTX 3060 Ti
~42% - $350 : TU-106-200 _ 1920 ==== RTX 2060
~33% - $200 : GM-206-400 _1024 = GTX960
~33% - $250 : GP-106-410 _ 1280 == GTX 1060Ti*
~33% - $280 : TU-116-400 _ 1536 === GTX 1660Ti
~33% - $330 : GA-106-300 _ 3584 ======= RTX 3060
~31% - $230 : TU-116-300 _ 1440 === GTX 1660-S
~30% - $200 : GP-106-300 _ 1152 == GTX 1060
~30% - $220 : TU-116-200 _ 1408 === GTX 1660
~28% - $160 : TU-116-300 _ 1280 === GTX 1650-S
~25% - $160 : GM-206-250 _ 768 = GTX950
~24% - $250 : GA-106-150 _ 2560 ======= RTX 3050
~21% - $130 : GM-206-200 _ 640 = GTX750Ti
~20% - $140 : GP-107-400 _ 768 == GTX 1050Ti
~17% - $100 : GM-206-100 _ 512 = GTX750
~17% - $110 : GP-107-300 _ 640 == GTX 1050
~11% - $150 : TU-117-300 _ 896 === GTX 1650
* GTX 1060 Ti, is just a nickname for the enhanced 6GB model with 9Gbps memory, as opposed to the cut-down, slower model, with only 3GB memory.
Based on this we can see a trends happening, especially when it comes to price:
9-series* -> 10-series = average +12% price difference (-30% min, max +50%)
10-series -> 16-series = average +16% price difference (+11% min, max +21%)
10-series -> 20-series = average +54% price difference (+32% min, max +109%)
20-series -> 30-series = average +5% price difference (-20% min, max +22%)
30-series -> 40-series = average +43% price difference (+8% min, max +69%)
Averages of averages: -1% min, +26% ave, +55% max = Overall price trend per generation
Bruzzone - Tuesday, October 18, 2022 - link
Kangai (san) very complete table, thank you. mbKangal - Saturday, October 22, 2022 - link
Here's the same list!But shorter, placed from Best Value to Least Value, according to USD $ price and the amount of hardware/silicon you're purchasing.
($5.71 per %) = (0.1750 % per $) = GTX 1650-S
($5.88 per %) = (0.1700 % per $) = GTX750
($6.06 per %) = (0.1650 % per $) = GTX960
($6.11 per %) = (0.1421 % per $) = GTX970
($6.19 per %) = (0.1615 % per $) = GTX750Ti
($6.40 per %) = (0.1563 % per $) = GTX950
($6.47 per %) = (0.1545 % per $) = GTX 1050
($6.67 per %) = (0.1500 % per $) = GTX 1060
($7.00 per %) = (0.1429 % per $) = GTX 1050Ti
($7.07 per %) = (0.1415 % per $) = GTX980Ti
($7.33 per %) = (0.1364 % per $) = GTX 1660
($7.42 per %) = (0.1348 % per $) = GTX 1660-S
($7.58 per %)= (0.1320 % per $) = GTX 1060Ti*
($7.60 per %) = (0.1316 % per $) = GTX 1070
($7.14 per %) = (0.1400 % per $) = GTX 1070Ti
($7.53 per %) = (0.1329 % per $) = GTX 1080Ti
($8.21 per %) = (0.1218 % per $) = GTX980
($8.33 per %) = (0.1200 % per $) = RTX 2060
($8.48 per %) = (0.1179 % per $) = GTX 1660Ti
($8.51 per %) = (0.1175 % per $) = RTX 2060-S
($8.64 per %) = (0.1157 % per $) = RTX 3080
($8.89 per %) = (0.1125 % per $) = RTX 3060 Ti
($8.96 per %) = (0.1117 % per $) = GTX 1080
($9.09 per %) = (0.1100 % per $) = RTX 3070
($10.00 per %) = (0.1000 % per $) = RTX 3060
($10.00 per %) = (0.1000 % per $) = RTX 2070
($10.00 per %) = (0.1000 % per $) = 9-GTX Titan X
($10.42 per %) = (0.0960 % per $) = RTX 3050
($10.53 per %) = (0.0950 % per $) = RTX 3070 Ti
($10.64 per %) = (0.0940 % per $) = RTX 2080 Ti
($10.71 per %) = (0.0933 % per $) = RTX 2070-S
($10.94 per %) = (0.0914 % per $) = RTX 2080
($11.94 per %) = (0.0838 % per $) = RTX 2080-S
($12.00 per %) = (0.0833 % per $) = 10-Titan Xp
($12.63 per %) = (0.0791 % per $) = RTX 3080 Ti
($13.64 per %) = (0.0733 % per $) = GTX 1650
($15.00 per %) = (0.0667 % per $) = RTX 4080
($15.31 per %) = (0.0653 % per $) = RTX 3090
($15.52 per %) = (0.0644 % per $) = RTX 4080-12GB
($16.49 per %) = (0.0606 % per $) = RTX 4090
($20.00 per %) = (0.0500 % per $) = RTX 3090 Ti
($25.00 per %) = (0.0400 % per $) = 20-Titan RTX
So here is the upgrade path for graphics cards. This is over the span of 8-Years, using logical thinking and in the most robotic way, to maximise the most value possible. With a limitation to never go downwards in performance (given), while hitting the maximum value (obvious), and never going backwards in generation (severe limitation), while making each upgrade step.
GTX750, GTX960, GTX 1050Ti, GTX 1660, RTX 2060, RTX 3080, RTX 4080.
$100 -> $200 -> $140 -> $220 -> $350 -> $700 -> $1200.
($5.88 per %), ($6.06), ($7.00), ($7.33), ($8.33), ($8.64), ($15.00).
Kangal - Saturday, October 22, 2022 - link
I also find it interesting how weird the Turing period was!The higher priced and early options were mostly terrible value. Whilst the latter "Super" options and the lower priced (16-series) ranged from decent to great value. It really was a weird era, see below:
"GTX 1650-Super" it has the best value ($5.71 per %) in history.
"RTX 2080" then offers exactly half the value ($10.94 per %), making it a bad choice.
"RTX Titan 20" then offers exactly half of that value ($25.00 per %), sits as the worst value in history.
lilkwarrior - Tuesday, November 1, 2022 - link
There are considerable flaws in evaluating the value, especially for creative and machine-learning professionals. It's terrible math with the proliferation of dedicated hardware for ray-tracing and machine-learning (tensor cores) in RTX GPUs.Especially with the AV1 and real-time path-tracing capabilities of the 4090 + invaluable capabilities like saturating what's possible on HDMI 2.1 panels.
Kangal - Friday, November 4, 2022 - link
Not really.Ray Tracing still isn't mainstream, it's pretty niche. Also, it doesn't really make drastic differences to the experience. In many cases, people would rather have a more fluid image (+120fps), or sharper image (+1800p resolution), or even allocate those funds into a better monitor (size, brightness, contrast, colour, GtG, etc etc). I think you have a point, when RT becomes dominant, but that will only mean we can't compare the newer cards against the older one, but the comparison itself would still be sound.
This is the fairest way to compare them. By comparing one card to its own family. And basing the product based on the amount of hardware/transistors you're buying. Then giving those cards a value rating based on its MSRP price versus the hardware. You can discern patterns happening in the market, as you compare how the price and hardware is trending between the different generations.
Another way of doing comparisons, would be to run each GPU at Stock Settings and test them individually on a Test System. Then compare their performance on a suite of games. Run an aggregate (eg 50 games), to derive a FPS per Dollar figure. But this is tricky and also imperfect, since some new games will be incompatible with older hardware. Some new hardware will not scale well on older games as they hit a ceiling / point of diminishing returns. And some titles will not have specific optimised per-game-ready drivers which also skews the results (drivers may get updated later, or a later game update may break the optimisations/bugs).
Again, the focus of my post was primarily about gaming, the hardware, and the money. As soon as you start mixing in other factors like Machine Learning, Encoding, Professional Applications that's when you have a problem. Because the target goal could be achieved by technological improvements in the software side too.
thestryker - Friday, October 14, 2022 - link
After seeing the RTX 4090 *not* top all of the charts in 1080p/1440p resolution I couldn't help but wonder what the performance would have been like for these cards. That's ignoring the fact that this card was also using a chip which historically would have been for a high end 60 or 70 series card. The gloves are clearly off over at nvidia and they don't care about what they're doing because nobody's leaving them as a customer. I've known for a while that they weren't a great company, but between the price reset on the 20 series and everything that went on during the 30 series I hope people start looking elsewhere. The GPU market is in a terrible place and needs a hard reset.WaltC - Friday, October 14, 2022 - link
I read that the 12GB GPU will simply be renamed--possibly 4070, etc.thestryker - Friday, October 14, 2022 - link
Sure, but will nvidia really drop what was going to be a $900 card down to $500-600? I don't think so.haukionkannel - Friday, October 14, 2022 - link
Of course not!This will be 4070ti at $899 and they will release cut down version as 4070 at $699 later with 10Gb or 6Gb of memory!