A series of rumors around NVIDIA's next-generation GeForce lineup, reportedly based on the Ampere GPU architecture, have appeared from various leakers on Twitter. The vast amounts of information reveal what to expect from the upcoming GPU architecture and also mentions the technical specifications of the various GPU SKUs that we would get in NVIDIA's next-gen Ampere GPU family.
NVIDIA Next-Gen GeForce 'Ampere' GPU Rumors - 10nm Process, Q4 2020 Launch, Ray-Tracing Across Entire Lineup, RTX 2080 Ti's Successor Up To 40% Faster
NVIDIA's Ampere GPU family has seen various rumors and so far, NVIDIA has been very tight-lipped about their upcoming family. We can't say for sure if Ampere would be the codename for its next-gen GPUs but the various SKUs that have slipped passed EEC seem to indicate that it might be the case after all.
Coming to the details, first are foremost, these leaks are not to be compared with the ones we had earlier which were only specific to the HPC parts. Those may feature a different GPU architecture & an entirely different fabrication process unlike the GeForce parts being mentioned here. With that cleared away, the two leakers from Twitter have reported that NVIDIA's next GPU architecture will be based on the Samsung 10nm process node and not go with TSMC as reported earlier. The new 10nm process node from Samsung will be utilized for the fabrication of NVIDIA's next-gen GeForce GPUs. The process is said to be closer to the 8LPP technology that Samsung has to offer but it isn't EUV-based.
"NO EUV. GA102 has taped out. It was based on SAMSUNG 10nm. (You can call it 8nm?) And the new tegra will use the same process node." - Kopite7kimi
Previously, Jensen Huang, CEO of NVIDIA, stated that TSMC will be handling the majority of its 7nm orders while Samsung would only get a small number of GPU orders. It makes sense that NVIDIA is just referring to 7nm parts and not 10nm parts like mentioned in this rumor. If TSMC were to get the larger order of 7nm GPUs, then that might be a different GPU altogether and it could be possible that the HPC oriented GA100 GPU is based on TSMC's 7nm process while the GeForce GPUs are to be fabricated at Samsungs (10nm / 8LPP) fab. The same process is also being used to produce the Orion SOC which utilizes NVIDIA's next-gen GPU architecture, reportedly Ampere.
The Rumored NVIDIA GeForce Ampere GPU Lineup - GA102, GA013, GA104, GA016 & GA107 For Gamers
Now with the process rumors out of the way, let's talk about the speculation that's been going on over the GeForce GPUs. According to KittyCorgi on Twitter, the GeForce lineup would be powered by a total of 5 Ampere GPUs and their respective SKUs featuring various core configurations. The topmost SKU in the lineup would be the GA102 which would be the successor to the TU102 GPU. The rest of the GPUs and their predecessors are mentioned below:
- NVIDIA GA102 - TU102 - GP102
- NVIDIA GA103 - No Predecessor
- NVIDIA GA104 - TU104 - GP104
- NVIDIA GA106 - TU106 - GP106
- NVIDIA GA107 - TU117 - GP107
Before we get into the details of these specific GPUs, it is mentioned that all GPUs will be made on the Samsung 10nm (8LPP) process node. SLI through a next-generation NVLINK interconnect will be available but only on the top-tier GA102 based graphics cards. Currently, this limitation only falls below the TU104 based GPUs but it looks like NVIDIA would impose further limitations with next-gen Ampere parts. NVIDIA would also enable RTX (real-time ray-tracing) across all GPUs, even the lower-end GA107 parts which means that there are some significant architectural enhancements made to the RT design on Ampere GPUs, allowing for more real-time raytracing performance on the new cards.
1 Surprisingly, the next gen chip will be based on Samsung 10nm node 2 Not very solid about what GA102 really got, may struggle to be 40% above 2080ti, but bound to be under that number if compared against full bore TU102 3 SLi will be available to GA102 ONLY 4 RTX for everyone pic.twitter.com/b5YAKmOmgf
— KittyCorgi (@CorgiKitty) March 11, 2020
Once again, it is mentioned that FP32 units on Ampere would be doubled but they don't translate into a 2x rasterization performance increase or twice the increase in the number of CUDA cores. The Ampere GPU wouldn't only improve the rasterization and shading performance but also deliver huge uplifts over Turing in ray tracing performance.
Since Turing was the 1st generation implementation of ray tracing on NVIDIA graphics cards, Ampere will further optimize it and is expected to offer more ray-tracing hardware onboard in the form of RT and Tensor cores. The Ampere GPUs are also mentioned to leverage performance from PCIe Gen 4.0 which is a given since this is a next-generation GPU architecture tuned for next-generation gaming platforms and the higher bandwidth from the Gen 4.0 protocol, plus NVLINK would be an essential feature. So let's talk about the GPU configurations now.
NVIDIA GA102 For Next-Gen Titan & GeForce RTX 3080 Ti Graphics Card- The Flagship Ampere GeForce GPU
The flagship GeForce GPU is rumored to be the GA102 and will be the successor to the Turing based TU102 GPU we got to see on the Titan RTX and RTX 2080 Ti graphics cards. The GPU will feature 84 SMs which equals 5376 CUDA cores. This is 16% more CUDA cores than the full-fat TU102 GPU featured on the Titan RTX. The GPU would be able to support up to 12 GB of VRAM across a 384-bit bus interface.
We can't say for sure if NVIDIA would be using the full configuration on the GeForce RTX 3080 Ti or whatever they call the RTX 2080 Ti's successor but if they follow the current plan which was used for Turing, then we would see a full fat 5376 core config on the next-generation Titan while the RTX 3080 Ti could use a more conservative core count of 5120 CUDA cores which may not be a huge difference which would be made up by higher clock speeds of AIB cards. The GA102 GPU is said to be up to 40% faster than the RTX 2080 Ti and once again, this is the full configuration that is being compared. It will be the same if not bigger leap as the GTX 1080 Ti was versus its non-Ti GTX 1080 brother.
NVIDIA GeForce RTX 3080 Graphics Card - Up To 10 GB GDDR6, 320-bit Bus, GA103 GPU
As mentioned above, the NVIDIA GeForce RTX 3080 will be the successor to the GeForce RTX 2080. It will feature a total of 60 SMs with 64 CUDA cores each making 3840 CUDA cores. This would be a big leap over the RTX 2080's CUDA core count which was set at 3072 and we aren't even factoring in the clock speeds which NVIDIA will be able to churn up with the advanced process node.
The card is also said to feature a 320-bit bus interface which means that it could support either 10 or 20 GB of GDDR6 memory. If NVIDIA really wants to blow the audience away with Ampere, then 20 GB would be a perfect choice. However, we have to take into consideration the high costs that are associated with GDDR6 memory and the looming rumors of a memory shortage due to higher demand from the console market.
So let's be conservative and say 10 GB of GDDR6, which would still be 2 GB more than the current GeForce RTX 2080. If we are looking at 16 Gbps modules on the upcoming RTX 3080, it would give us 640 GB/s bandwidth which would be insane over the 496 GB/s of the RTX 2080 SUPER which features 15.5 Gbps memory alongside a 256-bit bus interface. The NVIDIA GeForce RTX 3080 is said to be 10% faster than the RTX 2080 Ti which is a better performance bump than both the RTX 2080 and the GTX 1080 which only matched and were slightly faster (under 5%) than the outgoing Ti parts.
NVIDIA GeForce RTX 3070 Graphics Card - Up To 8 GB GDDR6, 256-bit Bus, GA104 GPU
The third card is the GeForce RTX 3070 and as stated above, it will feature the GA104 GPU. The successor to the TU106 based GeForce RTX 2070 is suggested to feature the same CUDA core count as the RTX 2080, at 3072, all rounded up in its 48 streaming multiprocessor units. The card is expected to feature a 256-bit bus interface which once again would allow for 8 GB or up to 16 GB GDDR6 memory, but you have to take into consideration what I just said above about the GDDR6 supply and cost factor.
This GPU is said to deliver 95% performance of the RTX 2080 Ti which could be pushed further ahead with custom models and higher factory overclocks. This would mean that NVIDIA would deliver a card that offers the performance of an $1200 US card for under $500 US which would be pretty impressive.
Lastly, we have the GA106 and GA107 GPUs which would target the mainstream and budget segments. The GA106 GPU would feature 30 SMs or 1920 CUDA cores with up to 6 GB memory across a 192-bit bus interface while the GA107 GPU would feature 20 MS or 1280 CUDA cores with up to 4 GB of memory across a 128-bit bus interface.
NVIDIA Ampere GPU 'GeForce RTX 3080 Ti, RTX 3080, RTX 3070' Rumored Specs Comparison:
|Graphics Card Name||NVIDIA GeForce RTX 2070||NVIDIA GeForce RTX 2080||NVIDIA GeForce RTX 3070||NVIDIA GeForce RTX 3080||NVIDIA GeForce RTX 2080 Ti||NVIDIA GeForce RTX 3080 Ti|
|GPU Architecture||Turing TU106||Turing TU104||Ampere GA104||Ampere GA103||Turing TU102||Ampere GA102|
|Process||12nm FinFET NVIDIA (TSMC)||12nm FinFET NVIDIA (TSMC)||Samsung 10nm (8LPP)?||Samsung 10nm (8LPP)?||12nm FinFET NVIDIA (TSMC)||Samsung 10nm (8LPP)?|
|CUDA Cores||2304 CCs||3072 CCs||3072 CCs||3840 CCs||4608 CCs||5376 CCs|
|Memory||8 GB GDDR6||8 GB GDDR6||8/16 GB GDDR6||10/20 GB GDDR6||11 GB GDDR6||12 GB GDDR6|
|Memory Speed||14 Gbps||14 Gbps||16 Gbps?||16 Gbps?||14 Gbps||16 Gbps?|
|Memory Bandwidth||448 GB/s||448 GB/s||512 GB/s||640 GB/s||640 GB/s||768 GB/s|
What's interesting is that these two cards leave so much room open for a GeForce RTX 3080 Ti to just blow away the enthusiast market segment again. If the reports of NVIDIA's overall Ampere architecture gaining 50% higher performance than Turing while offering twice the efficiency are correct, than given those specs, an RTX 3070 would be able to deliver the performance or even outperform the RTX 2080 Ti at a much lower price point.
According to the Taipei Times, NVIDIA's Ampere GPU is going to deliver seriously impressive figures in both performance and efficiency. The overall performance gain is expected to be 50% higher than existing Turing GPUs while offering twice the power efficiency. It is clearly mentioned that Ampere would halve the power consumption compared to existing Turing GPUs. This means that a GPU with a lower TDP figure than the RTX 2060 (125-150W) would be able to outperform the 2080 Ti which is seriously impressive. One can only expect the kind of performance which a high-end GPU based on the Ampere architecture would offer as we can definitely expect true 4K 60 FPS ray-traced gaming from the updated architecture.
These are truly exciting details but we have to remember that all of these are just rumors at this point. NVIDIA had to recently push its annual GTC conference from a live webcast to just news releases in light of the coronavirus outbreak. That might change the timeline for major GPU and key announcements that were to be made so we have to wait for official confirmation before we can consider these specifications 100% accurate.