The Radeon RX 7900 XTX and RX 7900 XT graphics cards are powered by the world’s first chiplet gaming GPU, according to an alleged AMD RDNA 3 “Navi 31” GPU block diagram that has leaked online. In the gaming GPU market, the AMD RDNA 3 ushers in the chiplet era. With its Ryzen & EPYC CPUs, AMD has led the chiplet revolution.

The Red Team was also the first to release a chiplet GPU under the Instinct series name, but now it’s time for consumers to enjoy some decent chiplet gaming on their new GPUs. In contrast to a typical Monolithic design, AMD claims that the chiplet GPU design offers a revolutionary architectural approach.

A single GCD with 48 WGPs, 96 Compute Units with 6 Shader Engines, and 8 dual compute units per Shader Engine are all features of the AMD Navi 31 GPU with RDNA 3 architecture. There will be a total of 12,288 stream processors distributed as a result. This represents a 2.4x increase in cores above the 5120 SPs found in the Navi 21 GPU.

AMD uses a dual-issue SIMD unit design that doubles the rate at which instructions are sent, enabling it to perform FP, INT, and AI operations. Additionally, the GPU includes two AI accelerators per CU that increase performance by 2.7x. Additionally, BF16 is supported.

According to reports, the Navi 31 GCD, or GPU, will be packaged on TSMC’s 5nm production node and size 300mm2. The top die of AMD’s newest RDNA 3 GPU, which has 58 billion transistors overall and a maximum compute performance of 61 TFLOPs, can support. The 7nm RDNA 2 “Navi 21” GPU has 165% less transistors per mm2 than the Navi 31 GPU, according to AMD. According to the corporation, it improved performance per watt by 45% from generation to generation, exceeding its objective of 50% improvement.

Regarding the cache system, AMD’s RDNA 3 “Navi 31” GPU includes a total of four separate cache hierarchies.

The Infinity Cache LLC, which consists of 96 MB in 6 MCD pools and is interconnected by a 5.2 TB/s Infinity Fabric link, comes first. There are six megabytes (MB) of L2 cache, three megabytes (MB) of L1 cache, and two megabytes (MB) of L0 cache. The AD102 L3 cache is completely matched by the 96 MB Infinity Cache. According to AMD, the GPU’s effective bandwidth is 3.5 TB/s. Accordingly, the GDDR6 interface accounts for the remaining 960 GB/s of the Infinity Cache’s 2.5 TB/s ratings.

With a bigger RT accelerator that provides 50% more performance per CU, AMD’s second-generation ray tracing solution is more powerful. A few of the improvements include new ray box sorting and traversal, dedicated instructions, and 50% more rays in flight. AMD anticipates ray tracing speed improvements of up to 80% with the use of larger caches (helping with Complex RT workloads) (at 2.5 GHz).

AMD emphasises the utilisation of MDIA (Multi-Draw Indirect Accelerator), a 12 Primitive/CLK (50% boost), and double the hardware Prim/Vert Cull Rates while discussing the Geometry and Pixel pipelines. AMD boasts 192 Pixels/Clk, Random Order Opaque outputs, 6 Prims Rasterized/Clk (50% increase), and Pixel Wait Sync in Pixel. The rest of the information is largely what we already knew before the unveiling.

Also Read:

Source