A user on the Hardwareluxx forums has confirmed the existence of the NVIDIA Ampere GA102-250 GPU SKU, which was previously expected to appear on the GeForce RTX 3080 Ti graphics card. It was seen when the forum member disassembled his GeForce RTX 3090, only to reveal that the GPU had initially been labeled GA102-250 SKU instead of GA102-300, which is the official SKU used in the RTX 3090.
NVIDIA’s GeForce RTX 3080 Ti GPU, Ampere GA102-250, detected on a GeForce RTX 3090 graphics card
The NVIDIA GA102-250-KD-A1 GPU only existed at the rumor mill, but now we can confirm that it is a real chip. Initially planned for the GeForce RTX 3080 Ti graphics card, it seems that NVIDIA decided that they could use a smaller SKU to power the next flagship, while this SKU can be used for the GeForce RTX 3090 graphics card. You can clearly see that the label The GA102-250-KD-A1 GPU has been cut and instead reused as the GA102-300-A1 SKU. Interestingly, the GA102-250-KD-A1 was expected to feature 78 SM-enabled drives for a total of 9,984 CUDA cores, while the RTX 3090 with its GA102-300-A1 GPU includes 82 SM or 10,496 cores.
NVIDIA may have decided to reuse SKU GA102-250 on GA102-300 given the shortage of its high-end graphics cards. We can’t say for sure if the 250 variant offers the same performance / power limits as the 300 variant, but since it has a lower naming scheme, one might expect it to be a lower container than the standard GA102-300 SKU. The GPU was also manufactured during week 49 of 2020, which is between late November and early December. Again, we don’t know to what extent NVIDIA is repurposing its GA102-250 to GA102-300 chips, but since they exist and this is already available, we can expect more users to find out for themselves.
The NVIDIA GeForce RTX 3080 Ti, on the other hand, has undergone a specification change and will now use the Ampere GA102-225-A1 GPU SKU. Below are the expected specs of the upcoming RTX 30 series flagship.
NVIDIA GeForce RTX 3080 Ti 12GB ‘Rumored’ Graphics Card Specifications
NVIDIA’s GeForce RTX 3080 Ti 12GB FE (Founders Edition) graphics card is expected to feature the PG132-SKU18 PCB layout and GA102-225-KD-A1 graphics core. The GA102-225 GPU has also changed since we last saw it and will now feature 10,240 CUDA cores within a total of 80 SM drives. Since the main kernel configuration is completely new, the previously leaked TMU / ROP clocks and counts should now be discarded.
In terms of memory, the card will feature 12GB of GDDR6X memory. Unlike the 19.5 Gbps speeds of the RTX 3090, the GeForce RTX 3080 Ti is said to retain the same memory speeds as the RTX 3080 at 19 Gbps. Given that we get 12GB of memory, the NVIDIA GeForce RTX 3080 Ti will use a 384-bit bus interface which equates to a total bandwidth of 912GB / s. So even though we see an 8GB VRAM reduction over the previous spec, the larger bus interface should lead to higher memory bandwidth.
The TGP of the card is set to be the same as the RTX 3080 at 320 Watts. That’s definitely necessary to power the extra cores, so NVIDIA might have to tune the clocks a bit here. The NVIDIA GeForce RTX 3080 Ti is expected to launch in mid-April for an MSRP of around $ 999 US. The card is rumored to be released in May.
NVIDIA GeForce RTX 30 Series ‘Ampere’ Graphics Card Specifications:
|Graphics card name||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||GA107 Ampere||GA106 amp?||GA106-300 Amp||GA104-200 Amp||GA104-300 Amp||GA104-400 Amp||GA102-200 Amp||GA102-225 Ampere?||GA102-300 Amp|
|Process node||Samsung 8nm||Samsung 8nm||Samsung 8nm||Samsung 8nm||Samsung 8nm||Samsung 8nm||Samsung 8nm||Samsung 8nm||Samsung 8nm|
|Matrix size||TBA||TBA||TBA||395.2 mm2||395.2 mm2||395.2 mm2||628.4 mm2||628.4 mm2||628.4 mm2|
|Transistors||TBA||TBA||TBA||17.4 billion||17.4 billion||17.4 billion||28 billion||28 billion||28 billion|
|TMU / ROP||TBA||TBA||112/64||152/80||184/96||192/104?||272/96||320/112?||328/112|
|Tensor / RT Cores||TBA||TBA||112/28||152/38||184/46||192/48?||272/68||320/76?||328/82|
|Base clock||TBA||TBA||1320 MHz||1410 MHz||1500 MHz||TBA||1440 MHz||TBA||1400 MHz|
|Boost the clock||TBA||TBA||1780 MHz||1665 MHz||1730 MHz||TBA||1710 MHz||TBA||1700 MHz|
|FP32 Compute||TBA||TBA||12.7 TFLOP||16.2 TFLOP||20 TFLOP||TBA||30 TFLOP||TBA||36 TFLOP|
|RT TFLOP||TBA||TBA||25.4 TFLOP||32.4 TFLOP||40 TFLOP||TBA||58 TFLOP||TBA||69 TFLOP|
|Tensioner-TOPs||TBA||TBA||101 TOP||129.6 TOP||163 TOP||TBA||238 TOP||TBA||285 TOP|
|Memory capacity||4 GB GDDR6?||6GB GDDR6?||12GB GDDR6||8GB GDDR6||8GB GDDR6||8GB GDDR6X||10GB GDDR6X||12GB GDDR6X||24GB GDDR6X|
|Memory bus||128 bit||192 bits?||192 bits||256 bits||256 bits||256 bits||320 bits||384 bits||384 bits|
|Memory speed||TBA||TBA||15 Gbps||14 Gbps||14 Gbps||TBA||19 Gbps||19 Gbps||19.5 Gbps|
|Broadband||TBA||TBA||360 Gbps||448 Gbps||448 Gbps||TBA||760 Gbps||912 Gbps||936 Gbps|
|Price (MSRP / FE)||$ 149?||$ 199?||$ 329||399 US dollars||499 US dollars||599 US dollars?||699 US dollars||US $ 999?||$ 1499 dollars|
|Launch (availability)||2021?||2021?||February 2021||December 2020||October 29th||May 2021?||September, 17th||April 2021?||24th September|