Brand new information regarding AMD's RDNA 3 flagship GPU, the Navi 31, has been revealed by @Kepler_L2 according to which the chip can rock up to 384 MB of Infinity Cache in standard & 3D packaging.
AMD RDNA 3 'Navi 31' MCM GPUs Allegedly Rock Up To 384 MB of Infinity Cache In 3D V-Cache Flavors
The information is specifically for the AMD RDNA 3 flagship, the Navi 31 chip. So far, we don't have any information regarding the top chip which is only known as Navi 3x so Navi 31 should retain the flagship position when it launches this year and might get replaced by a faster variant next year.
With that said, the AMD Navi 31 GPU is confirmed to rock an advanced chiplet packaging design with next-generation Infinity Cache. According to @Kepler_L2 who digs around various patches & logs within AMD's drivers, it looks like the Navi 31 GPU will be rocking up to 6 MCDs (Multi-Cache Dies).
384MB I guess.
— Kepler (@Kepler_L2) July 13, 2022
32MB + 64-bit MC normally, +32MB 3D V-Cache on some models.
— Kepler (@Kepler_L2) July 13, 2022
Each MCD allegedly features up to 32 MB of standard cache and will also come in 3D V-Cache flavors. This rounds up to 192 MB of standard cache but the insider mentions that there will also be a few variants with 3D V-Cache and those will go up to 384 MB Infinity Cache. The Infinity Cache also comes in 32 MB size, half of what is featured on the Zen 3 3D V-Cache parts (64 MB) but since there are 6 MCDs to work around, the chip can theoretically offer an additional 384 MB of Infinity Cache through 3D packaging. So that's 192 MB of standard and 192 MB of 3D V-Cache for a total of 384 MB Infinity Cache.
That's a 3x increase over the AMD RDNA 2 flagship, the Navi 21, which only features 128 MB of total Infinity Cache. The Infinity Cache architecture helped AMD pump more bandwidth to the GPU, which allows it to remove memory bottlenecks and offer a fast cache close to the GPU rather than going through the VRAM.
AMD Radeon RX 7900 XT Graphics Card Specifications
The AMD Navi 31 GPU, the flagship RDNA 3 chip, would power the next-gen enthusiast cards such as the Radeon RX 7900 XT graphics card. We have heard that AMD will drop CU (Compute Units) in favor of WGP (Work Group Processors) on its next-gen RDNA 3 GPUs. Each WGP will house dual CU (Compute Units) but with twice the SIMD32 clusters as opposed to just 2 on each CU within RDNA 2. Rumors are that AMD has the option to select between Samsung and TSMC for the 6nm die.
According to the latest information, the AMD Navi 31 GPU with RDNA 3 architecture is expected to offer a single GCD with 48 WGPs, 12 SAs, and 6 SEs. This will give out a total of 12,288 stream processors which is lower than the previous count. This will also drop the overall compute performance unless AMD goes crazy with over 3.0 GHz clock frequencies on its flagship part. The Navi 31 GPU will also carry 6 MCD's which will feature 32 MB Infinity Cache per die and are also likely to carry the 64-bit (32-bit x 2) memory controllers that will provide the chip with a 384-bit bus interface.
As for clock speeds, the AMD Navi 31 GPU is said to offer clock speeds that can hit or even exceed 3 GHz. NVIDIA's flagship GPUs are also said to offer close to 2.8 GHz clock speeds but AMD has had a clear advantage in clock speeds over NVIDIA during the past generation so it's expected to continue. A 3 GHz clock speed means that we can expect over 75 TFLOPs of FP32 performance on the newest flagship which will be a 2.3x increase over the current RDNA 2 flagship, the RX 6950 XT.
Now, this is going to result in higher power draw and AMD seems to have confirmed this much that their next-generation graphics card lineup will feature higher power consumption but they will still be a more efficient option than what NVIDIA has to offer. The AMD Radeon RX 6950 XT already has a TBP of 335W so for a >2x performance gain, we can expect a final TBP for the Radeon RX 7900 XT to end up close to 400-450W. The cards are expected to launch later this year as confirmed by AMD themselves.
AMD RDNA GPU (Generational Comparison) Preliminary:
| GPU Name | Navi 10 | Navi 21 | Navi 31 |
|---|---|---|---|
| GPU Process | 7nm | 7nm | 5nm/6nm |
| GPU Package | Monolithic | Monolithic | MCD (Multi-Chiplet Die) |
| Shader Engines | 2 | 4 | 6 |
| GPU WGPs | 20 | 40 | 48 |
| SPs Per WGP | 128 | 128 | 256 |
| Compute Units (Per Die) | 40 | 80 | 192 |
| Cores (Per Die) | 2560 | 5120 | 12288 |
| Cores (Total) | 2560 | 5120 | 12288 |
| Peak Clock | 1905 MHz | 2250 MHz | ~3000 MHz |
| FP32 Compute | 9.7 | 23 | ~75 |
| Memory Bus | 256-bit | 256-bit | 384-bit |
| Memory Type | GDDR6 | GDDR6 | GDDR6 |
| Memory Capacity | 8 GB | 16 GB | 24 GB |
| Infinity Cache | N/A | 128 MB | 96-192 MB |
| Flagship SKU | Radeon RX 5700 XT | Radeon RX 6900 XTX | Radeon RX 7900 XT |
| TBP | 225W | 330W | 350W |
| Launch | Q3 2019 | Q4 2020 | Q4 2022 |
News Sources: FreeDesktop #1, FreeDesktop #2
