in from Tyrannosaurus 55. The leaker reports that, based on new information, AMD has created a more powerful design than previously expected.
AMD Radeon RX 7900 XT, the flagship RDNA 3 ‘Navi 31’ GPU-driven graphics card that delivers over 3 GHz clock speeds and nearly 100 TFLOPs of performance
Over the past few days, multiple sightings of AMD RDNA 3 (GFX11) GPUs in the LLVM project in Linux OS Drivers and Development Tools have begun to appear. It looks like AMD is currently working on at least four new GPUs, which have been confirmed via their ID leaks as shown below (Credit: @Kepler_L2):
- GFX (1100) – possible Navi 31 GPU
- GFX (1101) – possible Navi 32 GPU
- GFX (1102) – possible Navi 33 GPU
- (GFX 1103) – possible Phoenix APU
Now, while these IDs don’t reveal much, such as specs or configurations, they do tell us that preliminary support is being added, and soon, these GPUs will begin testing within their respective platforms to optimize performance delivery. But the topic is not about the leaked ID, but about the latest rumors about the flagship RDNA 3 GPU Navi 31, which will power the Radeon RX 7900 series graphics cards.
According to the latest rumors from Greymon55, it seems that the clock speed has increased substantially from 2.5 GHz to now hovering even above the 3 GHz frequency mark. This means that a flagship chip with 15,360 stream processors will be able to deliver nearly 100.0 TFLOPs of FP32 compute performance (92 TFLOPs to be exact). That’s without increasing the OC limit, which will definitely help it get closer to the 100 TFLOPs barrier it’s already approaching.
≈92T FP32
— Greymon55 (@greymon55) April 30, 2022
3GHz+ 7900xt
— Greymon55 (@greymon55) April 30, 2022
But it seems that recent reports of NVIDIA making significant changes to the architecture of its Ada Lovelace GPUs for its GeForce RTX 40 series have also been factored in, with leakers expecting the Ada Lovelace flagship to not only be the first to break the computing barrier of 100 TFLOPs GPUs, but also Offers clock rates similar to AMD’s RDNA 3, likely thanks to the switch from Samsung’s 8nm to TSMC’s leading-edge 4N node, an optimized variant based on TSMC’s N5 process (5nm).
AMD RDNA 3 Navi 31 GPU for Radeon RX 7900 Series
The flagship RDNA 3-chip AMD Navi 31 GPU will power the next generation of enthusiast graphics cards like the Radeon RX 7900 XT graphics card. We’ve heard that AMD will be ditching CUs (Compute Units) in favor of WGPs (Workgroup Processors) on its next-gen RDNA 3 GPUs. Each WGP will hold dual CUs (compute units), but twice the number of SIMD32 clusters compared to 2 on each CU in RDNA 2.
According to the Linux driver, the compute unit still seems to be there, and the WGP is still the DualCU. The biggest change is going back to 4x SIMD per CU like GCN, but now using RDNA’s 32 ALU design. pic.twitter.com/hAzGs9M9E6
— Kepler (@Kepler_L2) April 30, 2022
The Navi 31 GPU configuration shown here has two GCDs (graphics core chips) and one MCD (multi-cache chip). 3 shader engines per GCD (6 total) and 2 shader arrays per shader engine (2 per SE/6 per GCD/12 total). Each shader array consists of 5 WGPs (10 per SE / 30 per GCD / 60 total), each WGP has 8 SIMD32 units and 32 ALUs (40 SIMD32 per SA / each SE 80 / 240 per GCD / 480 total). Combined, these SIMD32 cells make up 7,680 cores per GCD and 15,360 cores in total.
A preliminary block diagram of AMD’s next-generation RDNA 3-based Navi 31 GPU, which will power the flagship Radeon RX 7900 XT graphics card. (Image credit: Orlak)
The Navi 31 (RDNA 3) MCD will be connected to dual GCDs via next-generation Infinity Fabric interconnects and feature 256-512 MB of Infinity Cache. Each GPU should also have 4 memory connection links (32-bit). A 256-bit bus interface has a total of eight 32-bit memory controllers.
There are also rumors of 3D Stacking on the flagship Navi 31 GPU, which is currently expected to launch by the end of the fourth quarter of 2022, but we will likely start seeing the actual available paper product in the first quarter of 2023.
Upcoming flagship AMD, Intel, NVIDIA GPU specs (preliminary)
GPU name | AD102 | Navigation 31 | Xe2-HPG |
---|---|---|---|
code name | Ada Lovelace | rDNA 3 | Battle Mage |
Flagship SKUs | GeForce RTX 4090 Series | Radeon RX 7900 series | Arc B900 Series |
GPU process | TSMC 4N | TSMC 5nm+ TSMC 6nm | TSMC 5nm? |
GPU package | Monolithic | MCD (Multi-Chip Chip) | MCM (Multi-Chiplet Module) |
GPU freezes | Mono x 1 | 2 x GCD + 4 x MCD + 1 x IOD | Quad-Tile (tGPU) |
GPU Large Cluster | 12 GPCs (Graphics Processing Clusters) | 6 shader engines | 10 render slices |
GPU super cluster | 72 TPC (Texture Processing Cluster) | 30 WGPs (per MCD) 60 WGPs (total) |
40 Xe cores (per block) 160 Xe cores (total) |
GPU cluster | 144 Streaming Multiprocessor (SM) | 120 Compute Units (CUs) 240 computational units (total) |
1280 Xe VE (per block) 5120 Xe VE (total) |
core (per chip) | 18432 CUDA cores | 7680 SPs (per GCD) 15360 SPs (total) |
20480 ALUs (total) |
peak clock | ~2.85 GHz | ~3.0 GHz | pending |
FP32 computing | ~105 TFLOPs | ~92 TFLOPs | pending |
memory type | GDDR6X | GDDR6 | GDDR6? |
Memory Capacity | 24 GB | 32 GB | pending |
memory bus | 384 bit | 256 bits | pending |
memory speed | ~21Gbps | ~18Gbps | pending |
cache subsystem | 96 MB L2 cache | 512 MB (unlimited cache) | pending |
pending | ~600W | ~500W | pending |
emission | Q4 2022 | Q4 2022 | 2023 |
Which next-gen GPU are you most looking forward to?
News source: Videocardz