Igor over at Igor’s lab the date for the upcoming NVIDIA RTX 3080 Ti graphics card has just passed. This is the most anticipated card in NVIDIA’s RTX 3000 range, given that it has the full 20 GB vRAM buffer (unlike the 10 GB buffer on the 3080) and is expected to be the “true” flagship, since the 3090 is practically a replacement for the Titan. While Igor has been solid in the past, please note that NVIDIA’s launch programs are vulnerable to changing Jensen’s opinion and other supply and demand factors – so take the launch data with a grain of salt.
Igor: RTX 3060 will come in 6 GB and 12 GB flavors.
According to Igor, the RTX 3080 Ti will land in February 2021, while the RTX 3060 will land in early January 2021. This makes sense, as the RTX 3060 is scheduled to be the budget king for the NVIDIA range and should be one on which users will want to buy in large volumes. That being said, we’re not sure about NVIDIA’s supply lines because a few weeks ago, there was no visibility into when the company would be able to recover demand.
Even more interesting, Igor mentions that the RTX 3060 will have a 6 GB and 12 GB flavor. If this is true, then the value of a 10X RTX 3080 would seriously depreciate. I complained about the 10GB buffer right from the start – because it’s certainly not enough for a range of 2020 games that are based on on streaming assets. Who knows, NVIDIA might decide to offer RTX 3080 and the 20 GB treatment, once it has enough GDDR6 memory stock in stock.
Summary: RTX 3080 Ti 20GB
NVIDIA RTX 3080 Ti was originally supposed to be based on the PG132 SKU20 (although NVIDIA decided to rename it) and has been in the works for some time. Although the exact number of the CUDA core is not yet confirmed, it will have twice the amount of memory compared to vanilla RTX 3080. To be clear, even the nomenclature is not confirmed, as the SUPER suffix could appear anywhere in the name, if Jensen would consider it so.
We strongly recommend that users (who want to buy an NVIDIA card) choose RTX 3090 (if they can afford it) or wait for RTX 3080 Ti / SUPER 20 GB.
With this amount of shading power, you will want to have a free memory buffer. According to Videocardz, the card will land in January 2021. Given that killer apps such as Microsoft Flight Simulator 2020 are already removing all the vRAMs 2080 has offered you, 10 GB will feel small after a year. or two. State-of-the-art consoles already drive active streaming games in big worlds, so vRAM will start to become much more crucial than before. With AMD throwing out a flagship 16GB vRAM card, NVIDIA had to respond with their own high-end vRAM card.
According to Kopite, the RTX 3080 Ti will be based on the GA102 and will have specifications between an RTX 3080 and 3090. The exact nomenclature of the chip is GA102-250-A1 and will have a 384-bit bus with GDDR6X memory. The bus size means that NVIDIA will use either a 12 GB or a 24 GB buffer. Given that Microsoft Flight Simulator 2020 is already blocked by the 11GB branch on the RTX 2080 Ti, it would be disappointing to see NVIDIA deliver another powerful GPU with a small memory buffer. With RTX IO and asset streaming (the Unreal Engine demo for next-generation consoles) becoming a thing next year or so, every piece of buffer will help in this paradigm shift.
Jesus. A new GA102 specification, between 3080 and 3090.
– kopite7kimi (@ kopite7kimi) October 27, 2020
It is also unclear at this time how the unveiling of this new GPU fits into rumors about NVIDIA planning a move back to TSMC. However, it gives credence to the belief that the RTX 3000 series, at least for now, remains on the Samsung 8nm process. The company initially faced less-than-ideal returns and supply constraints at launch, but expects them to improve significantly as we enter the new year.
GA102-250-A1, 9984FP32, 384bits GD6X
– kopite7kimi (@ kopite7kimi) October 27, 2020
This will be an extremely powerful card with 34 TFLOP power. That being said, the current API, driver and application infrastructure cannot take full advantage of all this raw power. This is the real reason why NVIDIA’s extremely powerful cards do not escalate linearly with TLFOPs. The company has essentially made cards that are ahead of their time compared to the surrounding software ecosystem.
The hard proof of this lies in the fact that the RTX 3000 series experienced nonlinear positive scaling when it descended into the stack. An RTX 3070 that has slightly more cores than the RTX 2080 Ti beats the former flagship – exposing any rumors or accusations about the CUDA cores in the Ampere series that are not as strong as the Turing series or the deceptive fault in designing the missing architecture of INT performance.
AMD’s large Navi series drops later today and it remains to be seen how the GPU market will form as we enter the holiday season. NVIDIA’s prices are great, but they need to work with developers to fix the API and drivers to take full advantage of the gross performance offered by the Ampere series (fine wine of the highest quality) and in the meantime, AMD will come out cards budget friendly performance with what appears to be a wide range. They have also taken a lot of steps to ensure that the bot scaling that took place with NVIDIA does not happen at the launch of their Radeon RX 6000.
NVIDIA GeForce RTX 30 “Amp” graphics card specifications:
|The name of the graphics card||NVIDIA GeForce RTX 3050||NVIDIA GeForce RTX 3050 Ti||NVIDIA GeForce RTX 3060||NVIDIA GeForce RTX 3060 Ti||NVIDIA GeForce RTX 3070||NVIDIA GeForce RTX 3070 Ti?||NVIDIA GeForce RTX 3080||NVIDIA GeForce RTX 3080 Ti?||NVIDIA GeForce RTX 3090|
|GPU name||Ampere GA107||Ampere GA106?||Ampere GA106?||Ampere GA104-200||Ampere GA104-300||Ampere GA102-150||Ampere GA102-200||Ampere GA102-250||Ampere GA102-300|
|Process node||Samsung 8nm||Samsung 8nm||Samsung 8nm||Samsung 8nm||Samsung 8nm||Samsung 8nm||Samsung 8nm||Samsung 8nm||Samsung 8nm|
|The size of death||TBA||TBA||TBA||395.2 mm2||395.2 mm2||628,4mm2||628,4mm2||628,4mm2||628,4mm2|
|transistors||TBA||TBA||TBA||17.4 billion||17.4 billion||28 billion||28 billion||28 billion||28 billion|
|TMUs / PORs||TBA||TBA||TBA||152/80||184/96||232/80||272/96||328/112||328/112|
|Tensor / RT Colors||TBA||TBA||TBA||152/38||184/46||232/58||272/68||328/82||328/82|
|Basic watch||TBA||TBA||TBA||1410 MHz||1500 MHz||TBA||1440 MHz||TBA||1400 MHz|
|Boost Clock||TBA||TBA||TBA||1665 MHz||1730 MHz||TBA||1710 MHz||TBA||1700 MHz|
|Calculation FP32||TBA||TBA||TBA||16.2 TFLOPs||20 TFLOPs||TBA||30 TFLOPs||TBA||36 TFLOP|
|RT TFLOP||TBA||TBA||TBA||32.4 TFLOPs||40 TFLOP||TBA||58 TFLOPs||TBA||69 TFLOPs|
|Tensor-tops||TBA||TBA||TBA||TBA||163 TOPs||TBA||238 TOPs||TBA||285 TOPs|
|Memory capacity||4 GB GDDR6?||6 GB GDDR6?||6 GB GDDR6?||8 GB GDDR6||8 GB GDDR6||10 GB GDDR6X?||10 GB GDDR6X||20 GB GDDR6X||24 GB GDDR6X|
|Memory bus||128 bits||192 bits?||192 bits?||256 bits||256 bits||320 bits||320 bits||320 bits||384 bits|
|Memory speed||TBA||TBA||TBA||14 Gbps||14 Gbps||TBA||19 Gbps||19 Gbps||19.5 Gbps|
|Band width||TBA||TBA||TBA||448 Gbps||448 Gbps||TBA||760 Gbps||760 Gbps||936 Gbps|
|Price (MSRP / FE)||$ 149?||$ 199?||$ 299?||399 US dollars?||499 US dollars||$ 599?||699 US dollars||$ 899?||1499 US dollars|
|Launch (Availability)||2021?||2021?||2021?||November 2020?||October 29||Q4 2020?||September 17||January 2021?||September 24|