I. Preface: the new RDAN3 architecture floating point performance soared 2.7 times
After 2 full years of the RTX 30 series and RX 6000 series graphics cards being available, the market for the Solo Display has finally seen an upgrade.
First debuted NVIDIA's RTX 4090/4080, and now the AMD RX 7900 XT/XTX with the new RDNA3 architecture is here.
The following is an explanation of the new technologies in the RDNA3 architecture.
AMD's new graphics card takes its cues from the Riptide processor Chiplets with a new MCM design. the Navi 31 has 6 MCDs and 1 GCD.
MCD means Memory Cache Die, which is where the Infinity Cache cache and graphics memory controller are located, using a relatively mature 6nm process.
A single MCD measures 37mm2 and contains 16MB Infinity Cache and a 64Bit GDDR6 memory controller. 6 MCDs cover a total area of 220mm2, making up a 384Bit memory bit width and 96MB Infinity Cache.
The GCD is a Graphics Compute Die, which includes a stream processor compute unit, VGPR media unit, AI gas pedal, and RT optical chase gas pedal, etc. The GDC uses an advanced 5nm process with an area of approximately 306 mm2.
The most direct benefit of such a design is that it can significantly reduce the cost of the chip!
Generally speaking, the larger the chip area, the lower the yield, MCM can not only significantly improve the yield, relatively do not eat the performance of the part of the mature 6nm process can also effectively reduce costs.
2, dual-emission design of the flow processor unit
Many people may be confused by the fact that the RX 7900 XTX transistor count has multiplied while the stream processor count has only increased by 20%!
The RDNA3 uses a new stream processor design scheme with 64 FP32 units and 64 INT32 units in each group of CU cells. Each FP32 and INT32 unit can perform integer or floating point operations as required.
At the limit, with all FP/INT32 units performing floating-point operations, a group of CU units is equivalent to having 12 stream processors. With 96 CU groups, the RX 7900 XTX can theoretically equate to a maximum of 12,488 stream processors and a floating point capability of 61 TFLOPS, compared to 23 TFLOPS for the previous generation RX 6950 XT, a direct jump of 2.7x.
This design concept is similar to NVIDIA's Ada Lovelace architecture, except that NV counts INT32 units as stream processors, so the RTX 4090's stream processor count looks horrible.
In additionRDAN3 also adds a new AI computing unit (AI Accelerator), similar to the role of the N card's Tensor Core.With 2 AI Accelerators in each CU group, it can improve the efficiency of related operations by 2.7 times. If there are operations involving deep learning, the operation efficiency of GPU will be greatly improved.
In RDNA3 GPU, AMD introduced DP2.1 interface for the first time, the transmission bandwidth from DP1.4 of 32Gbps to 54Gbps, can output 8K165Hz or 4K480Hz, and support 12 bit color depth, can display 68 billion colors.
Since the RX 7900 XTX has a 384Bit bit width with 20GHz GDDR6 memory, it is not as tight on memory bandwidth as its predecessor.
Therefore, RDNA3 reduces the Infinity Cache capacity, but at the same time increases the cache operating frequency to 2.3GHz, giving Infinity Cache a terrifying bandwidth of 5.3TB/s, equivalent to 2.7 times that of its predecessor.
In terms of FSR, FSR 2.0 is currently the main push, with 85 mainstream games currently supported, followed by FSR 2.2 to further improve the image quality.
For NVIDIA's focus on the single DLSS 3, AMD also has a response plan, and is expected to launch FSR 3.0 in the first quarter of next year, supporting the new AMD Fluid Motion Frame complementary frame technology, which can bring up to 2 by frame rate improvement compared to FSR 2.0.
The RX 7900 XTX debuts at $7,999, the same price as its predecessor, the RX 6900 XT, which is $1,500 cheaper than the competing RTX 4080. The RX 7900 XT, on the other hand, is priced at $7,499, which looks like a better price/performance ratio than the $500 more expensive RX 7900 XTX.
We will be using the RX 7900 XTXX Red Devil for this test.