AMD RDNA 3 'Navi 31' MCM GPUs Allegedly Rock Up To 384 MB of Infinity Cache In 3D V-Cache Flavors

AMD RDNA 3 GPUs For Radeon RX 7000 Graphics Cards Detailed – Navi 31 “Plum Bonito”, Navi 32 “Wheat Nas”, Navi 33 “Hotpink Bonefish”

Tech outlet Angstronomics has published a very detailed rundown on the specifications of AMD’s RDNA 3 GPUs, such as the Navi 31, Navi 32, and Navi 33 chips that will power its next-generation Radeon RX 7000 series graphics cards.

GPU AMD Navi 31, Navi 32, Navi 33 “RDNA 3” for Radeon RX 7000 series graphics cards detailed

Previously leaked information within Linux patches has revealed that there will be at least four RDNA 3 IPs being released for the consumer segment. These include the Navi 31, Navi 32, Navi 33, and an APU design.

11.0.0 gfx1100 (Nav31) AMDGPU_FAMILY_GC_11_0_0 (FAMILY_GFX1100) dGPU
11.0.1 gfx1103 AMDGPU_FAMILY_GC_11_0_2 (FAMILY_GFX1103) APU
11.0.2 gfx1102 (Nav32) AMDGPU_FAMILY_GC_11_0_0 (FAMILY_GFX1100) dGPU
11.0.3? gfx1101 (Nav33)? AMDGPU_FAMILY_GC_11_0_0 (FAMILY_GFX1100)? dGPU?

Now, Angstronomics has reported information such as die sizes, GPU configurations, and features that may end up in final retail products. An interesting statement has been made by the tech outlet revealing that the configurations for the Navi 3X GPUs were first formulated in 2019 and then finalized sometime in 2020. These specifications have not been changed however various rumors have noted two different specifications for the GPUs. meaning either AMD has in fact made changes since their final designs or those rumors just weren’t aligned with the truth. With that said, let’s see what kind of configurations we can expect with the Navi 3x “RDNA 3” line of GPUs.

The AMD RDNA 3 Navi 3x GPU illustration shows possible chip configurations. (Image credits: Olrak_29)

AMD confirmed that its RDNA 3 GPUs will arrive later this year with a huge performance boost. The company’s SVP of Engineering, Radeon Technologies Group, David Wang, said that the next generation GPUs for the Radeon RX 7000 series will deliver more than 50% performance per watt compared to existing RDNA 2 GPUs. Some of the key features of the RDNA 3 GPUs highlighted by AMD will include:

  • 5nm process node
  • Advanced chip packaging
  • Redesigned compute unit
  • Optimized graphics pipeline
  • Next Generation AMD Infinity Cache
  • >50% yield/watt vs. RDNA 2

Gamers should expect smooth 4K gaming to be smooth as butter on these graphics cards and with the FidelityFX suite offering state-of-the-art FSR, SAS and SAM support, we could even see 60 FPS playable at 8K resolution.

GPU Configuration AMD Navi 31 “Pretty Plum”

The AMD Navi 31 GPU, the flagship RDNA 3 chip, would power next-gen enthusiast cards like the Radeon RX 7900 XT graphics card. We have heard that AMD will drop CUs (Compute Units) in favor of WGPs (Workgroup Processors) on their next generation RDNA 3 GPUs. Each WGP will house dual CUs (compute units), but with double the number of SIMD32 clusters instead of just 2 on each CU within RDNA 2. Rumor has it that AMD has the option to select between Samsung and TSMC for the 6nm die .

  • AMD Navi 31: 12288 cores, 384-bit bus, 192MB Infinity Cache, 308mm2 5nm GPU die
  • AMD Navi 21: 5120 cores, 384-bit bus, 128MB Infinity Cache, 520mm2 7nm GPU die

Based on the latest information, the AMD Navi 31 GPU with RDNA 3 architecture is expected to offer a single GCD with 48 WGP, 12 SA and 6 SE. This will give a total of 12,288 SPs or stream processors. This is a 2.4x increase in cores compared to the 5120 SPs that are included in the Navi 21 GPU. The Navi 31 GCD or GPU is said to be 308mmtwo and will come packaged on TSMC’s 5nm process node.

The Navi 31 GPU will also carry 6 MCDs that will feature 16MB of Infinity cache per die and will also likely carry 64-bit memory controllers (32-bit x 2) that will provide the chip with a 384-bit bus interface. While this equates to 96 MB of Infinity Cache, which is less than the 128 MB found on current Navi 21 GPUs, there is also a recently noted 3D-Stacked solution in the works that would double Infinity Cache to 32 MB ( 16MB). MB 0-hi + 16 MB 1-hi) capacities for a total of 192 MB of cache. This is a 50% increase compared to the current Navi 21 design and also makes Navi 31 the first GPU with 3D stacked and chiplet designs. These chiplets or MCDs will be manufactured in TSMC’s 6nm process node and will measure 37.5mm.two each.

The outlet reports that there was an even larger 288 MB MCD package (16 MB 0-Hi + 16 MB 1-Hi + 16 MB 2-Hi) planned, but it was scrapped due to cost-benefit reasons. There is also talk of a stripped down variant of the Navi 31 GPU which will feature 42 WGP or 10,752 cores and 5 MCDs for 80MB of Infinity Cache on a 320-bit bus interface.

Now, this will result in higher power consumption and AMD seems to have confirmed so much that their next-gen lineup of graphics cards will feature higher power consumption, but will still be a more efficient option than what NVIDIA has to offer. The AMD Radeon RX 6950 XT already has a TBP of 335 W for a >2x performance gain. The cards are expected to retain their dual 8-pin socket input for power and feature an updated triple-fan cooling design that’s a bit taller than what’s currently in use.

AMD Navi 32 “Wheat Nas” GPU Configuration

The AMD Navi 32 GPU codenamed “Wheat Nas” is also one of two MCM GPUs to be included in the RDNA 3 lineup. The GPU will feature a single GCD (Graphics Compute Dies) and four MCDs (Multi-Cache Die ). The GPU or GCD is said to be 200mmtwo while each MCD will retain the same size at 37.5mmtwo.

  • AMD Navi 32: 7680 cores, 256-bit bus, 64MB Infinity Cache, 200mm2 5nm GPU die
  • AMD Navi 22: 2560 cores, 192-bit bus, 96MB Infinity Cache, 335mm2 7nm GPU die

The GCD will be equipped with 3 Shader Engines and each Shader Engine has 6 Shader Arrays (18 total). The GPU will feature a total of 30 WGP or 60 compute units for a total of 7680 cores. This rounds up to >50% more cores than the flagship Navi 31.

Each GCD must also include 2 memory connection links (32 bits). That’s a total of 8 x 32-bit memory controllers for a 256-bit bus interface. Each MCD will retain the 16MB 0-Hi capacity, but there doesn’t appear to be any 1-Hi variants, meaning the chip will only have 64MB maximum configurations. The Navi 32 GPU is said to be released in 2023 and will be the highest chip for mobile devices and the mid-tier chip for desktop computers. The GPU would be implemented in the mobile segment first and could be announced at CES 2023, which makes a lot of sense since by then we will have the Phoenix Range and Dragon Range laptops with Zen 4 cores ready.

AMD Navi 32 “Hotpink Bonefish” GPU Configuration

The AMD Navi 33 GPU codenamed “Hotpink Bonefish” will start the monolithic segment within the RDNA 3 family. The GPU will feature a single die. The die is very similar to the flagship Navi 21 GPU and is expected to use a 6nm process node for manufacturing with a die size of 203mm.two.

  • AMD Navi 33: 4096 cores, 128-bit bus, 32MB Infinity cache, 335mm2 6nm GPU die
  • AMD Navi 23: 2048 cores, 128-bit bus, 32MB Infinity cache, 237mm2 7nm GPU die

Navi 33 GCD is expected to have 2 Shader Engines and each Shader Engine has 2 Shader Arrays (2 per SE / 4 total). This rounds up to 16 WGP or 32 compute units for a total of 4096 cores, which is higher than the Radeon RX 6800 “Navi 21 XL” GPU. The GPU will come packed with 32MB of Infinity Cache, the same amount as the Navi 23 GPU, and on a 128-bit wide bus. Like the Navi 32 GPUs, the Navi 33 chips will target the mobile lineup first with early graphics performance that will put them ahead of the Intel Arc Alchemist lineup at less than half the cost and with less power.

The outlet also talks about two new features coming to AMD RDNA 3 Navi 3x GPUs for Radeon RX 7000 series graphics cards, such as:

One of the features of the RDNA 3 graphics pipeline is OREO – Opaque Random Export Order, which is just one of many area saving techniques. With gfx10, pixel shaders are executed out of order, where the outputs go to a reorder buffer before being passed to the rest of the pipeline in order. With OREO, the next step (merge) can now receive and execute trades in any order and export to the next step in order. Therefore, the ROB can be replaced with a much smaller slide buffer, saving area.

Memory Attached Last Level (MALL) cache blocks are halved in size, doubling the number of banks for the same amount of cache. There are also changes and additions that increase graphics bandwidth to the MALL and reduce the exit penalty to VRAM.

via Angstronomics

The AMD Radeon RX 7000 “RDNA 3” GPU lineup based on the Nav 3x GPUs is expected to launch later this year with reports facing the flagship release of Navi 31, followed by the Navi 32 and Navi 33 GPUs.

AMD RDNA 3 Navi 3X GPU Configurations (Preview)

GPU name Navi 21 Navi 33 Navi 32 Navigation 31 Navi 3X
key name sienna cichlid hot pink bonefish Nas of wheat Plum Bonito Determined
GPU process 7nm 6nm 5nm/6nm 5nm/6nm 5nm/6nm
GPU package Monolithic Monolithic LCM (1 GCD + 4 GCD) LCM (1 GCD + 6 GCD) LCM (TBD)
GPU die size 520mm2 203 mm2 (GCD only) 200 mm2 (GCD only)
425 mm2 (with MCD)
308 mm2 (GCD only)
533 mm2 (with MCD)
shader engines 4 two 4 6 8
GPU WGP 40 sixteen 30 48 64
SP per WGP 128 256 256 256 256
Compute Units (per array) 80 32 60 96 128 (per GPU)
256 (Overall)
Cores (for dying) 5120 4096 7680 12288 8192
Nuclei (total) 5120 4096 7680 12288 16,384
memory bus 256 bit 128 bit 256 bit 384 bit 384 bit x2?
memory type GDDR6 GDDR6 GDDR6 GDDR6 GDDR6
memory capacity Up to 16GB Up to 8GB Up to 16GB Up to 24GB Up to 32GB
memory speed 16-18Gb/s Determined Determined 20Gb/s Determined
memory bandwidth 512-576GB/s Determined Determined 960GB/s Determined
infinite cache 128MB 32MB 64MB 96/192MB Determined
Flagship SKUs Radeon RX 6900 XTX Radeon RX 7600XT? Radeon RX 7800XT?
Radeon RX 7700XT?
Radeon RX 7900XT? Radeon Pro
TBP 330W ~150W ~250W ~350W Determined
Throw Q4 2020 Q4 2022? Q4 2022? Q4 2022? 2023?

Which latest generation AMD Radeon RX 7000 graphics card are you most interested in?

Products mentioned in this publication.

Leave a Comment

Your email address will not be published.