NVIDIA's GeForce RTX 40 GPUs based on the Ada Lovelace graphics architecture have their SM configurations allegedly leak out, pointing out to over 18,000 cores for the flagship AD102 chip.
NVIDIA GeForce RTX 40 'Ada Lovelace' GPU Configurations Leak Out: AD102 With 144 SMs, AD103 With 84 SMs, AD104 With 60 SMs, AD106 With 36 SMs, AD107 With 24 SMs
Recently, NVIDIA got hacked and hackers were able to steal over 1 TB of confidential information which has started leaking out. Some information that has leaked out in the public already includes a bypass for the LHR technology, source code for DLSS technology, & codenames of next-gen GPU architectures. We have seen information regarding Hopper's successor, Blackwell, leak out that will feature at least two Data Center chips but this latest leak is specific to the consumer GPU lineup on the Ada Lovelace GPU architecture.
NVIDIA GeForce RTX 40 'Ada Lovelace' GPU Configurations
According to the leak, NVIDIA will have at least six GPUs within its Ada Lovelace lineup. These will include AD102, AD103, AD104, AD106, AD107, and AD10B. The first five SKUs will be designed for the desktop and mobility segments and featured in both GeForce RTX 40 and RTX Workstation solutions. The last part is reported by Kopite7kimi to be specific to the next-gen Tegra SOC while the Ampere-based GA10F could go on to power the next-gen Switch handheld console and Tegra Drive solutions.
So coming to the leaked SKUs, the top AD102 GPU which is likely going to power the next-gen GeForce RTX 4090, RTX 4080 Ti graphics cards will make use of 144 SMs, a 71% increase over the existing GA102 GPU and house a massive 18,432 CUDA core count. The interesting thing here is that the AD102 GPU is the only SKU that is getting over a 50% increase in SM count & considering what we have heard about the flagship chip, in terms of performance and power consumption, it looks very likely that NVIDIA is going all out with its top chip in the Ada Lovelace family.
The AD103 GPU will replace the GA103 GPU which was recently introduced on mobile and feature the same SM count as the GA102 GPU at 84. The AD104, AD106, and AD107 GPUs will feature 60, 36, and 24 SM units, respectively. Besides the AD103 GPU which is a 40% SM increase over GA103, every other GPU gets a 25-20% SM count increase over its predecessor. It's not as significant as the AD102 GPU but considering this is the mainstream segment, we are likely going to get RTX 3080 or similar performance out of an RTX 4060 Ti & RTX 3070 or higher performance out of the standard RTX 4060. The RTX 4050 should be close or on par with an RTX 3060 given the addition of IPC and clock improvements aside from architectural upgrades.
In addition to the SM counts, the Ada Lovelace GPUs will also feature increased L2 cache sizes. Starting with the AD102 GPU, the flagship would be outfitted with up to 96 MB of L2 cache, an insane 16x increase over the 6 MB L2 cache featured on GA102. The AD103 GPU will feature 64 MB, AD104 will feature 48 MB while both AD106/AD107 GPUs will feature 32 MB of L2 cache. As for the memory bus, the flagship AD102 GPU will feature a 384-bit bus interface, the AD103 GPU will get a 256-bit bus interface, AD104 will feature a 192-bit bus interface, while the AD106/AD107 GPUs will get a 128-bit bus interface.
NVIDIA Ada Lovelace 'GeForce RTX 40' GPU Configurations
| GPU Name | GPCs / TPCs | SMs Per TPC / Total | CUDA Cores | L2 Cache | Memory Bus | Power Limit Desktop (Peak) | Power Limit Mobile (Expected) |
|---|---|---|---|---|---|---|---|
| AD102 | 12 / 6 | 2 / 144 | 18432 | 96 MB | 384-bit | 800W | N/A |
| AD103 | 7 / 6 | 2 / 84 | 10752 | 64 MB | 256-bit | 450W | 175W |
| AD104 | 5 / 6 | 2 / 60 | 7680 | 48 MB | 192-bit | 400W | 175W |
| AD106 | 3 / 6 | 2 / 36 | 4608 | 32 MB | 128-bit | 260W | 140W |
| AD107 | 3 / 4 | 2 / 24 | 3072 | 32 MB | 128-bit | 180W | 80W |
NVIDIA Ada Lovelace & Ampere GPU Comparison
| Ada Lovelace GPU | SMs | CUDA Cores | Graphics Card Series | Memory Bus | Ampere GPU | SMs | CUDA Cores | Top SKU | Memory Bus | SM Increase (% Over Ampere) |
|---|---|---|---|---|---|---|---|---|---|---|
| AD102 | 144 | 18432 | RTX 4090? | 384-bit | GA102 | 84 | 10752 | RTX 3090 Ti | 384-bit | +71% |
| AD103 | 84 | 10752 | RTX 4080? | 256-bit | GA103S | 60 | 7680 | RTX 3080 Ti | 256-bit | +40% |
| AD104 | 60 | 7680 | RTX 4070? | 192-bit | GA104 | 48 | 6144 | RTX 3070 Ti | 256-bit | +25% |
| AD106 | 36 | 4608 | RTX 4060? | 128-bit | GA106 | 30 | 3840 | RTX 3060 | 192-bit | +20% |
| AD107 | 24 | 3072 | RTX 4050? | 128-bit | GA107 | 20 | 2560 | RTX 3050 | 128-bit | +20% |
Previously rumored specs have shown us a huge update to the core specs. The NVIDIA AD102 "ADA GPU" appears to have 18432 CUDA Cores. This is almost twice the cores present in Ampere which was already a massive step up from Turing. A 2.2 GHz clock speed would give us 81 TFLOPs of compute performance (FP32). This is more than twice the performance of the existing RTX 3090 which packs 36 TFLOPs of FP32 compute power.
We have also heard that to support such extreme specifications and the massive increase in SM / Core Count on the AD102 GPU, the top NVIDIA GeForce RTX 40 GPUs such as the RTX 4090 or RTX 4090 Ti could feature a TDP of up to 850W.
NVIDIA is already investing development around the new PCIe Gen 5 connector that offers up to 600W power input per connector. The delayed GeForce RTX 3090 Ti is one example where the card is expected to rock at a TGP of 450W and will be the first desktop graphics card to utilize such a connector interface. The next-gen cards are also expected to utilize the same PCIe standard but it looks like the top variant could end up with two Gen 5 connectors to supplement the ~800W power requirement.
Several PSU makers have already started releasing their brand new Gen 5 power supplies which would include the necessary connectors to support the next-gen GPUs but they only feature one primary Gen 5 connector which means that if NVIDIA was to use a second 16-pin port, users will have to use a 2x 8-pin to 1x 16-pin adapter which will ship with most of these PSUs.
NVIDIA / AMD Next-Gen Rumored GPU Performance Estimates
Kopite7kimi also hinted at some specification details of the NVIDIA Ada Lovelace chips a while back which you can read more about here and check out the specs in the table provided below:
NVIDIA CUDA GPU (RUMORED) Preliminary:
| GPU | TU102 | GA102 | AD102 |
|---|---|---|---|
| Flagship SKU | RTX 2080 Ti | RTX 3090 Ti | RTX 4090? |
| Architecture | Turing | Ampere | Ada Lovelace |
| Process | TSMC 12nm NFF | Samsung 8nm | TSMC 4N? |
| Die Size | 754mm2 | 628mm2 | ~600mm2 |
| Graphics Processing Clusters (GPC) | 6 | 7 | 12 |
| Texture Processing Clusters (TPC) | 36 | 42 | 72 |
| Streaming Multiprocessors (SM) | 72 | 84 | 144 |
| CUDA Cores | 4608 | 10752 | 18432 |
| L2 Cache | 6 MB | 6 MB | 96 MB |
| Theoretical TFLOPs | 16 TFLOPs | 40 TFLOPs | ~90 TFLOPs? |
| Memory Type | GDDR6 | GDDR6X | GDDR6X |
| Memory Capacity | 11 GB (2080 Ti) | 24 GB (3090 Ti) | 24 GB (4090?) |
| Memory Speed | 14 Gbps | 21 Gbps | 24 Gbps? |
| Memory Bandwidth | 616 GB/s | 1.008 GB/s | 1152 GB/s? |
| Memory Bus | 384-bit | 384-bit | 384-bit |
| PCIe Interface | PCIe Gen 3.0 | PCIe Gen 4.0 | PCIe Gen 4.0 |
| TGP | 250W | 350W | 600W? |
| Release | Sep. 2018 | Sept. 20 | 2H 2022 (TBC) |
The NVIDIA Ada Lovelace GPU family is expected to bring a generational jump similar to Maxwell to Pascal. It is expected to launch in the second half of 2022 but expect supply and pricing to be similar to current cards despite NVIDIA spending billions of dollars to accquire those good good TSMC 5nm wafers.
News Source: @davideneco25320
