{"id":122838,"date":"2025-04-02T12:55:26","date_gmt":"2025-04-02T10:55:26","guid":{"rendered":"https:\/\/prohoster.info\/?p=122838"},"modified":"2025-04-02T12:55:26","modified_gmt":"2025-04-02T10:55:26","slug":"novaya-statya-radeon-rx-9070-xt-protiv-geforce-rtx-5070-ti","status":"publish","type":"post","link":"https:\/\/prohoster.info\/en\/blog\/news\/novaya-statya-radeon-rx-9070-xt-protiv-geforce-rtx-5070-ti","title":{"rendered":"New Article: Radeon RX 9070 XT vs GeForce RTX 5070 Ti","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

GPU Specifications of the Blackwell Family and Tests GeForce RTX 5080<\/a><\/noindex> made it clear that all models in the 50 series, with the exception of the GeForce RTX 5090, are little more than minor upgrades over their previous generation counterparts. The situation has been worsened by the shortage of new graphics cards and, consequently, skyrocketing prices.<\/p>\n

Most buyers are only theoretically interested in high-end devices, which are monopolized by NVIDIA. However, just one step down, the technological stagnation of 'green' graphics has created a situation where AMD could potentially catch up in problematic areas for the RDNA architecture, such as ray tracing and upscaling. Yet, the prophecy that AMD will restore parity with its competitor (albeit without aspirations for the flagship sector) has been heard for a long time, each time sounding less convincing. Consequently, the announcement of the Radeon RX 9000 graphics cards did not attract much attention against NVIDIA's vibrant activity. However, AMD has carried out extensive work to address past mistakes, and while it may not have eliminated them entirely, it has at least compensated for the deficiencies of previous GPUs.<\/p>\n

We present the review of the Radeon RX 9070 XT in an unusual comparison format for 3DNews, as the 'red' newcomer and its main competitor\u2014the GeForce RTX 5070 Ti\u2014came into our hands simultaneously. One device cannot be viewed in isolation from the other, as they are sold in Russia for approximately the same price.<\/p>\n\n\n\n
\"New <\/td>\n <\/td>\n \"New <\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

<\/center> \t\t\t\t <\/p>\n

\u2193<\/a><\/noindex>#<\/a><\/noindex> Navi 48 GPU and RDNA 4 Architecture<\/strong><\/h2>\n

Only the lazy hasn't mentioned that NVIDIA has become a monopolist in the high-performance GPU market and behaves accordingly. Less attention is paid to the fact that all companies designing large processors are forced to rely on a single contractor for their production\u2014Taiwan's TSMC. The demand for TSMC's capacities has undoubtedly contributed to AMD halting the development of large consumer GPUs. Furthermore, the 'red' graphics processor line currently features just one chip from the new generation\u2014Navi 48.<\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Manufacturer<\/th>\nAMD<\/th>\n<\/tr>\n
Title<\/td>\nNavi 32<\/td>\nNavi 31<\/td>\nNavi 48<\/td>\n<\/tr>\n
Where Used<\/td>\n\n

Radeon RX 7700 XT;<\/p>\n

Radeon RX 7800 XT<\/p>\n<\/td>\n

\n

Radeon RX 7900 GRE;<\/p>\n

Radeon RX 7900 XT;<\/p>\n

Radeon RX 7900 XTX<\/p>\n<\/td>\n

\n

Radeon RX 9070;<\/p>\n

Radeon RX 9070 XTX<\/p>\n<\/td>\n<\/tr>\n

Architecture<\/td>\nRDNA 3<\/td>\nRDNA 3<\/td>\nRDNA 4<\/td>\n<\/tr>\n
Process Technology, nm<\/td>\nTSMC N5\/N6<\/td>\nTSMC N5\/N6<\/td>\nTSMC N4P<\/td>\n<\/tr>\n
Number of Transistors, billion<\/td>\n28,1<\/td>\n57,7<\/td>\n53,9<\/td>\n<\/tr>\n
Chip Area, mm\u00b2<\/td>\n346<\/td>\n522<\/td>\n357<\/td>\n<\/tr>\n
Number of CU\/WGP\/SE<\/td>\n<\/tr>\n
Compute Units (CU)<\/td>\n60<\/td>\n96<\/td>\n64<\/td>\n<\/tr>\n
Workgroup Processors (WGP)<\/td>\n30<\/td>\n48<\/td>\n32<\/td>\n<\/tr>\n
Shader Arrays (SA)<\/td>\nNo<\/td>\nNo<\/td>\nNo<\/td>\n<\/tr>\n
Shader Engines (SE)<\/td>\n4?<\/td>\n6<\/td>\n4<\/td>\n<\/tr>\n
Compute Unit Configuration<\/td>\n<\/tr>\n
Vector ALUs (FP32)<\/td>\n2 \u00d7 32<\/td>\n2 \u00d7 32<\/td>\n2 \u00d7 32<\/td>\n<\/tr>\n
Vector ALUs (FP32\/INT32)<\/td>\n2 \u00d7 32<\/td>\n2 \u00d7 32<\/td>\n2 \u00d7 32<\/td>\n<\/tr>\n
Scalar ALUs<\/td>\n2<\/td>\n2<\/td>\n2<\/td>\n<\/tr>\n
Specialized ALUs (SFU)<\/td>\n2 \u00d7 8<\/td>\n2 \u00d7 8<\/td>\n2 \u00d7 8<\/td>\n<\/tr>\n
Ray Tracing Blocks (Ray Accelerators)<\/td>\n1<\/td>\n1<\/td>\n1<\/td>\n<\/tr>\n
Texture Mapping Units (TMU)<\/td>\n4<\/td>\n4<\/td>\n4<\/td>\n<\/tr>\n
Vector\/Scalar Registers, Kbytes<\/td>\n384\/10<\/td>\n384\/10<\/td>\n384\/16<\/td>\n<\/tr>\n
L0 Cache Size, Kbytes<\/td>\n32<\/td>\n32<\/td>\n32<\/td>\n<\/tr>\n
Workgroup Processor (WGP) Configuration<\/td>\n<\/tr>\n
Shared Memory Size, Kbytes<\/td>\n128<\/td>\n128<\/td>\n128<\/td>\n<\/tr>\n
Instruction Cache Size, Kbytes<\/td>\n32<\/td>\n32<\/td>\n32<\/td>\n<\/tr>\n
Scalar Cache Size, Kbytes<\/td>\n16<\/td>\n16<\/td>\n16<\/td>\n<\/tr>\n
GPU Compute Units<\/td>\n<\/tr>\n
Vector ALUs (FP32)<\/td>\n3 840 (7 680)<\/td>\n6 144 (12 288)<\/td>\n4 096 (8 192)<\/td>\n<\/tr>\n
Texture Mapping Units (TMU)<\/td>\n240<\/td>\n384<\/td>\n256<\/td>\n<\/tr>\n
Raster Operations Blocks (ROP)<\/td>\n96<\/td>\n192<\/td>\n128<\/td>\n<\/tr>\n
Ray Tracing Blocks (Ray Accelerators)<\/td>\n60<\/td>\n96<\/td>\n64<\/td>\n<\/tr>\n
Memory Configuration<\/td>\n<\/tr>\n
L1 Cache Size, Kbytes<\/td>\n2 048<\/td>\n3 072<\/td>\n2 048<\/td>\n<\/tr>\n
L2 Cache Size, Mbytes<\/td>\n4<\/td>\n6<\/td>\n8<\/td>\n<\/tr>\n
Infinity Cache Size, Mbytes<\/td>\n64<\/td>\n96<\/td>\n64<\/td>\n<\/tr>\n
VRAM Bus Width, bits<\/td>\n256<\/td>\n384<\/td>\n256<\/td>\n<\/tr>\n
VRAM Chip Type<\/td>\nGDDR6<\/td>\nGDDR6<\/td>\nGDDR6<\/td>\n<\/tr>\n
<\/td>\n<\/tr>\n
PCI Express Interface<\/td>\n4.0 x16<\/td>\n4.0 x16<\/td>\n5.0 x16<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

Among older chips, it most closely resembles the second-tier model, Navi 32, in key specifications, but contains nearly as many transistors as Navi 31 (53.9 and 57.7 billion, respectively). Navi 48 is built on TSMC N4P technology, which belongs to the same 5nm photolithography node that was developed two years ago. However, Navi 48 occupies only 68% of the area of Navi 31 due to AMD abandoning the chiplet design, as the MCD chiplets, which contain the third-level cache and memory controllers, are manufactured using the coarser N6 process.<\/p>\n

\"New <\/p>\n

The architecture of the new graphics processor, RDNA 4, is structurally similar to previous iterations. Navi 48 contains 64 Compute Units, or 4,096 FP32-compatible shader ALUs, which is a typical formula for a mid-range GPU, as is the 256-bit memory bus. However, AMD contrasts Navi 48's performance with the flagship Navi 31, which has 6,144 full precision shader ALUs. The differences should be more than compensated by the numerous logic enhancements, and to understand them, it's advisable to refresh your memory on the RDNA 3 architecture, which we covered in detail in our review. Radeon RX 7900 XTX<\/a><\/noindex>.<\/p>\n

Compared to the 'green' architecture Blackwell<\/a><\/noindex> RDNA 4 represents a more extensive and profound upgrade of previous developments. However, the description of all innovations curiously ends up being a short list because NVIDIA traditionally pays increased attention to complex software or semi-software solutions that require elaborated commentary, while AMD's innovations tend to start and end with hardware.<\/p>\n

\"New <\/p>\n

Fundamental changes have occurred in the Compute Unit structure \u2014 the indivisible processing block that is analogous to NVIDIA's SM and CPU cores. For instance, the instruction scheduler now supports so-called split barriers, allowing instructions to be sent for execution during their active period. Instruction prefetching from the cache has also been optimized, and the scalar ALUs, which are used for efficient execution of conditional operations and branching, have inherited the ability to work with floating-point data from the mobile architecture RDNA 3.5. Additionally, the volume of scalar registers has been increased from 10 to 16 KB.<\/p>\n

\"New <\/p>\n

One of the most significant innovations of RDNA 4 is the doubled execution rate of matrix instructions WMMA, which are used in machine learning tasks and data processing through neural networks. Furthermore, the RDNA 4 instruction set contains commands for calculations on structurally sparse matrices and 8-bit floating-point data (FP8 and BF8).<\/p>\n

\"New <\/p>\n

Nevertheless, a throughput of 1,024 FLOPs per clock means that Navi chips have merely reached the level of the last 'green' architecture \u2014 Ampere. Moreover, while NVIDIA and Intel chips use separate ALU arrays for matrix operations, in the RDNA 4 architecture, this work still falls on the shader SIMD.<\/p>\n

However, we should not forget that in real applications, RDNA 3-based graphics processors never even come close to full utilization. The declared teraflops in FP32 general-purpose calculations are only achievable if the VOPD (Vector Operation Dual) format is used, which allows two independent vector instructions to be packed together. This format imposes numerous restrictions, so half of the 32-bit ALUs within the Compute Unit are often idle. In mixed workloads (such as gaming with upscaling), matrix operations will, of course, take away some computational resources from the shaders, but these losses will not be as significant as they seem at first glance.<\/p>\n\n\n\n\n\n\n
<\/th>\nCompute Unit (AMD RDNA 3)<\/th>\nCompute Unit (AMD RDNA 4)<\/th>\nStreaming Multiprocessor (NVIDIA Ada Lovelace)<\/th>\nStreaming Multiprocessor (NVIDIA Blackwell)<\/th>\nXe-core (Intel Xe-HPG)<\/th>\nXe-core (Intel Xe2)<\/th>\n<\/tr>\n
Execution units<\/td>\n\n

2 \u00d7 SIMD32 (FP32\/INT32);<\/p>\n

2 \u00d7 SIMD32 (FP32);<\/p>\n

2 \u00d7 SIMD2 (FP64);<\/p>\n

2 \u00d7 SIMD8 (SFU);<\/p>\n

2 \u00d7 scalar ALUs<\/p>\n<\/td>\n

\n

2 \u00d7 SIMD32 (FP32\/INT32);<\/p>\n

2 \u00d7 SIMD32 (FP32);<\/p>\n

2 \u00d7 SIMD2 (FP64);<\/p>\n

2 \u00d7 SIMD8 (SFU);<\/p>\n

2 \u00d7 scalar ALUs<\/p>\n<\/td>\n

\n

4 \u00d7 SIMD16 (FP32\/INT32);<\/p>\n

4 \u00d7 SIMD16 (FP32);<\/p>\n

2 \u00d7 SISD? (FP64);<\/p>\n

4 \u00d7 SIMD4 (SFU);<\/p>\n

4 \u00d7 scalar ALUs;<\/p>\n

4 \u00d7 tensor cores<\/p>\n<\/td>\n

\n

8 \u00d7 SIMD16 (FP32\/INT32);<\/p>\n

2 \u00d7 SISD? (FP64);<\/p>\n

4 \u00d7 SIMD4 (SFU);<\/p>\n

4 \u00d7 scalar ALUs;<\/p>\n

4 \u00d7 tensor cores<\/p>\n<\/td>\n

\n

16 \u00d7 SIMD8 (FP32);<\/p>\n

16 \u00d7 SIMD8 (INT32);<\/p>\n

16 \u00d7 SISD (FP64);<\/p>\n

16 \u00d7 SIMD2 (SFU);<\/p>\n

16 \u00d7 XMX<\/p>\n<\/td>\n

\n

8 \u00d7 SIMD16 (FP32);<\/p>\n

8 \u00d7 SIMD16 (INT32);<\/p>\n

8 \u00d7 SIMD2 (FP64);<\/p>\n

8 \u00d7 SIMD4 (SFU);<\/p>\n

8 \u00d7 XMX<\/p>\n<\/td>\n<\/tr>\n

SIMD operations per cycle<\/td>\n\n

128 \u00d7 FP32;<\/p>\n

64 \u00d7 INT32;<\/p>\n

256 \u00d7 FP16;<\/p>\n

4 \u00d7 FP64;<\/p>\n

16 \u00d7 transcendental functions<\/p>\n<\/td>\n

\n

128 \u00d7 FP32;<\/p>\n

64 \u00d7 INT32;<\/p>\n

256 \u00d7 FP16;<\/p>\n

4 \u00d7 FP64;<\/p>\n

16 \u00d7 transcendental functions<\/p>\n<\/td>\n

\n

128 \u00d7 FP32;<\/p>\n

64 \u00d7 INT32;<\/p>\n

128 \u00d7 FP16;<\/p>\n

2 \u00d7 FP64;<\/p>\n

16 \u00d7 transcendental functions<\/p>\n<\/td>\n

\n

128 \u00d7 FP32;<\/p>\n

128 \u00d7 INT32;<\/p>\n

128 \u00d7 FP16;<\/p>\n

2 \u00d7 FP64;<\/p>\n

16 \u00d7 transcendental functions<\/p>\n<\/td>\n

\n

128 \u00d7 FP32;<\/p>\n

128 \u00d7 INT32;<\/p>\n

256 \u00d7 FP16;<\/p>\n

16 \u00d7 FP64;<\/p>\n

32 \u00d7 transcendental functions<\/p>\n<\/td>\n

\n

128 \u00d7 FP32;<\/p>\n

128 \u00d7 INT32;<\/p>\n

256 \u00d7 FP16;<\/p>\n

16 \u00d7 FP64;<\/p>\n

32 \u00d7 transcendental functions<\/p>\n<\/td>\n<\/tr>\n

Matrix operations, FLOP per cycle (FP16)<\/td>\n512<\/td>\n1 024<\/td>\n2 048<\/td>\n2 048<\/td>\n2 048<\/td>\n2 048<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

Frame scaling using neural networks is no longer a theoretical situation for Radeon accelerators. The new, fourth version of FSR finally relies on machine learning, including for frame generation, with the models used representing a hybrid of convolutional networks and transformers. Unfortunately, unlike previous versions of the upscaler that worked on virtually any hardware, FSR 4 supports only RDNA 4 architecture graphics processors (due to the widening of the WMMA instruction set for FP8 data). A universal mode with simplified computational cores, as seen in Intel's XeSS, is not available.<\/p>\n

\"New <\/p>\n

Another direction of the RDNA architecture reform was ray tracing. The speed of ray intersection testing with a BVH box and a triangle has doubled per Ray Accelerator (and therefore per Compute Unit): from 4 and 1 per cycle to 8 and 2, respectively. However, the RT block of Intel's Battlemage chips operates at a pace of 18 and 2, while Blackwell GPUs can find four ray-triangle intersections per cycle and an undisclosed number of ray-box intersections. Furthermore, competitors perform the two types of tests (against the BVH box and the triangle) simultaneously, while AMD still does not. To some extent, parallelism is possible, as each Ray Accelerator now contains two independent \"testers,\" but not at full speed.<\/p>\n

\"New <\/p>\n

Additionally, the Ray Accelerator has gained the ability to process rays out of order. Rays that do not require access to slow memory stack levels are processed first, and the testing results are saved so that the shader program can receive them in the expected order.<\/p>\n

Ray tracing logic now uses a denser BVH8 structure instead of BVH4 (eight instead of four nodes per each volume hierarchy branch), which increases traversal speed while simultaneously reducing pressure on video memory. The size of the BVH in memory has also decreased due to the new primitive compression algorithm. Finally, AMD now allows the orientation of BVH boxes to be decoupled from the absolute scene coordinates. Rotating the boxes according to the geometry of objects reduces the number of steps needed for search.<\/p>\n

\"New <\/p>\n

It's unfortunate that, despite all the mentioned changes, RDNA 4 still executes BVH traversal on shader ALUs rather than using specialized logic. According to the classification of hardware ray tracing solutions proposed by Imagination Technologies, AMD's solutions are stuck at the second level, Intel's GPUs are at the third, while NVIDIA's hardware is already moving from level 3 to level 4 thanks to the SER feature\u2014dynamic grouping of instruction threads aimed at improving memory access coherence and, consequently, reducing latency.<\/p>\n

\"New <\/p>\n

There is no equivalent of SER in the RDNA 4 architecture, but a function called Out of Order Memory has been introduced, which helps address the same problem in a different way. Previous versions of RDNA, working with multiple instruction groups (wavefronts in AMD's terminology), returned data from memory strictly in the same order that requests were made. When data requested by a shader is in the fast cache, the return can still be delayed if another shader made its request first and its data is in relatively slow VRAM. Frankly, this limitation seems completely senseless in the context of multithreaded computing, but it surely appeared to simplify GPU organization and remained a reasonable compromise for some time (read: in rasterized games). RDNA 4 allows instruction groups to return data independently of the order of requests.<\/p>\n

\"New <\/p>\n

Out of Order Memory \u2014 is just one of the changes in the memory stack in the RDNA 4 architecture. Previously, the allocation of registers to the Compute Unit's shader programs happened statically, based on maximum consumption during a shader's lifespan, meaning that a portion of the register file often remained unused. RDNA 4 allows shaders to request and free registers dynamically, making the use of the register file more efficient.<\/p>\n

Interestingly, the Shader Engine \u2014 a significant part of the GPU that combines the CU array, rasterizer, primitive block, and rasterization operation blocks \u2014 in RDNA 4 has lost its own level 1 cache. However, each segment of the L2 cache now contains twice as many banks. Consequently, both its volume and throughput increased, notably in both directions \u2014 to the Shader Engine and to the Infinity Cache (in other words, L3). The size of the Infinity Cache itself remains the same as in Navi 32 \u2014 64 MB.<\/p>\n

Navi 48 features 16 lines of PCI Express 5.0, but obviously cannot fully utilize such a fast interface in gaming tasks.<\/p>\n

\"New <\/p>\n

Finally, AMD reports that the hardware video codec in the next-generation GPU provides enhanced encoding quality at low bitrates popular for streaming, but nothing more. This means that support for YUV 4:2:2 color sub-sampling, which is present in NVIDIA and Intel GPUs, has not been introduced.<\/p>\n

As for display output capabilities, the Navi 48 display controller is compatible with the latest versions of DisplayPort and HDMI interfaces: 2.1a and 2.1b respectively.<\/p>\n

\u2193<\/a><\/noindex>#<\/a><\/noindex>Specifications, Pricing<\/strong><\/h2>\n

The Radeon RX 9070 XT comes equipped with a fully functional Navi 48 chip featuring 64 Compute Units, which translates to 4,096 standard precision floating-point ALUs. Finding a formal predecessor for this new model among the previous generation 'red' accelerators is challenging, as both the lineup and the GPU numbering principle have changed (Navi chips never reached the eighth iteration before). In terms of computing block counts and memory stack characteristics, the Radeon RX 9070 XT resembles the Radeon RX 7800 XT, but AMD positions it as a replacement for the Radeon RX 7900 XTX.<\/p>\n

The Game Clock and Boost Clock (target clock frequency in games and clock frequency limit) of the graphics processor have significantly increased compared to the flagship model of the previous generation, reaching 2.4 and 2.97 GHz, respectively. The theoretical performance of the Radeon RX 7900 XTX based on the maximum clock frequency is 26% higher, but AMD plans to compensate for this difference through improvements in the RDNA 4 architecture.<\/p>\n

The Radeon RX 9070 XT has 16 GB of GDDR6 memory with a bandwidth of 20 Gbps on a 256-bit bus. Reports suggest that AMD considered increasing the VRAM to 32 GB, but ultimately deemed that configuration excessive for consumer solutions. It's a pity, as game demands continue to grow, and the amount of video memory has been a key competitive advantage for many previous AMD devices.<\/p>\n

The reference power consumption of the Radeon RX 9070 XT is 304 watts, and the recommended price is $599 (which is $100 more than the Radeon RX 7800 XT).<\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Manufacturer<\/th>\nAMD<\/th>\nNVIDIA<\/th>\n<\/tr>\n
Model<\/td>\nRadeon RX 7900 XT<\/td>\nRadeon RX 7900 XTX<\/td>\nRadeon RX 9070<\/td>\nRadeon RX 9070 XT<\/td>\nGeForce RTX 4070 Ti SUPER<\/td>\nGeForce RTX 4080 SUPER<\/td>\nGeForce RTX 5070 Ti<\/td>\nGeForce RTX 5080<\/td>\n<\/tr>\n
Graphics Processor<\/td>\n<\/tr>\n
Title<\/td>\nNavi 31 XT<\/td>\nNavi 31 XTX<\/td>\nNavi 48 XT<\/td>\nNavi 48 XTX<\/td>\nAD103<\/td>\nAD103<\/td>\nGB203<\/td>\nGB203<\/td>\n<\/tr>\n
Architecture<\/td>\nRDNA 3<\/td>\nRDNA 3<\/td>\nRDNA 4<\/td>\nRDNA 4<\/td>\nAda Lovelace<\/td>\nAda Lovelace<\/td>\nBlackwell<\/td>\nBlackwell<\/td>\n<\/tr>\n
Process Technology<\/td>\nTSMC N5\/N6<\/td>\nTSMC N5\/N6<\/td>\nTSMC N4C<\/td>\nTSMC N4C<\/td>\nTSMC 4N<\/td>\nTSMC 4N<\/td>\nTSMC 4NP<\/td>\nTSMC 4NP<\/td>\n<\/tr>\n
Number of Transistors, billion<\/td>\n57,7<\/td>\n57,7<\/td>\n53,9<\/td>\n53,9<\/td>\n45,9<\/td>\n45,9<\/td>\n45,6<\/td>\n45,6<\/td>\n<\/tr>\n
Clock Frequency (Base Clock\/Game Clock\/Boost Clock), MHz<\/td>\n1 500 \/2 025\/2 394<\/td>\n1 855 \/2 269\/2 499<\/td>\n1 330 \/2 070\/2 520<\/td>\n1 660 \/2 400\/2 970<\/td>\n2 310\/2 610<\/td>\n2 205\/2 550<\/td>\n2 295\/2 452<\/td>\n2 295\/2 617<\/td>\n<\/tr>\n
Shader ALUs (FP32)<\/td>\n5 376 (10 752)<\/td>\n6 144 (12 228)<\/td>\n3 584 (7 168)<\/td>\n4 096 (8 192)<\/td>\n8 448<\/td>\n10 240<\/td>\n8 960<\/td>\n10 752<\/td>\n<\/tr>\n
Texture Mapping Units (TMU)<\/td>\n336<\/td>\n384<\/td>\n224<\/td>\n256<\/td>\n264<\/td>\n320<\/td>\n280<\/td>\n336<\/td>\n<\/tr>\n
Raster Operations Blocks (ROP)<\/td>\n192<\/td>\n192<\/td>\n128<\/td>\n128<\/td>\n112<\/td>\n112<\/td>\n96<\/td>\n112<\/td>\n<\/tr>\n
Tensor Cores<\/td>\nNo<\/td>\nNo<\/td>\nNo<\/td>\nNo<\/td>\n264<\/td>\n320<\/td>\n280<\/td>\n336<\/td>\n<\/tr>\n
RT Cores<\/td>\n84<\/td>\n96<\/td>\n112<\/td>\n128<\/td>\n66<\/td>\n80<\/td>\n70<\/td>\n84<\/td>\n<\/tr>\n
Last Level Cache Size, MB<\/td>\n80<\/td>\n96<\/td>\n64<\/td>\n64<\/td>\n48<\/td>\n64<\/td>\n64<\/td>\n64<\/td>\n<\/tr>\n
Memory<\/td>\n<\/tr>\n
Bus Width, bits<\/td>\n320<\/td>\n384<\/td>\n256<\/td>\n256<\/td>\n256<\/td>\n256<\/td>\n256<\/td>\n256<\/td>\n<\/tr>\n
Chip Type<\/td>\nGDDR6 SGRAM<\/td>\nGDDR6 SGRAM<\/td>\nGDDR6 SGRAM<\/td>\nGDDR6 SGRAM<\/td>\nGDDR6X SGRAM<\/td>\nGDDR6X SGRAM<\/td>\nGDDR7 SGRAM<\/td>\nGDDR7 SGRAM<\/td>\n<\/tr>\n
Bandwidth per Pin, Gbps<\/td>\n20<\/td>\n20<\/td>\n20<\/td>\n20<\/td>\n21<\/td>\n23<\/td>\n28<\/td>\n30<\/td>\n<\/tr>\n
Total Bandwidth, GB\/s<\/td>\n800<\/td>\n960<\/td>\n640<\/td>\n640<\/td>\n672<\/td>\n736<\/td>\n896<\/td>\n960<\/td>\n<\/tr>\n
Capacity, GB<\/td>\n20<\/td>\n24<\/td>\n16<\/td>\n16<\/td>\n16<\/td>\n16<\/td>\n16<\/td>\n16<\/td>\n<\/tr>\n
Performance<\/td>\n<\/tr>\n
Peak Performance FP32, TFLOPS<\/td>\n25,7 (51,5)<\/td>\n30,7 (61,1)<\/td>\n18,1 (36,1)<\/td>\n24,3 (48,7)<\/td>\n44,1<\/td>\n52,2<\/td>\n43,9<\/td>\n56,3<\/td>\n<\/tr>\n
FP64\/FP32 Performance<\/td>\n1\/32 (1\/64)<\/td>\n1\/32 (1\/64)<\/td>\n1\/32 (1\/64)<\/td>\n1\/32 (1\/64)<\/td>\n1\/64<\/td>\n1\/64<\/td>\n1\/64<\/td>\n1\/64<\/td>\n<\/tr>\n
FP16\/FP32 Performance<\/td>\n2\/1<\/td>\n2\/1<\/td>\n2\/1<\/td>\n2\/1<\/td>\n1\/1<\/td>\n1\/1<\/td>\n1\/1<\/td>\n1\/1<\/td>\n<\/tr>\n
Other<\/td>\n<\/tr>\n
PCI Express Bus<\/td>\nPCI Express 4.0 x16<\/td>\nPCI Express 4.0 x16<\/td>\nPCI Express 5.0 x16<\/td>\nPCI Express 5.0 x16<\/td>\nPCI Express 4.0 x16<\/td>\nPCI Express 4.0 x16<\/td>\n5.0 x16<\/td>\n5.0 x16<\/td>\n<\/tr>\n
Display Interfaces<\/td>\nDisplayPort 2.1, HDMI 2.1a<\/td>\nDisplayPort 2.1, HDMI 2.1a<\/td>\nDisplayPort 2.1a, HDMI 2.1b<\/td>\nDisplayPort 2.1a, HDMI 2.1b<\/td>\nDisplayPort 1.4a, HDMI 2.1<\/td>\nDisplayPort 1.4a, HDMI 2.1<\/td>\nDisplayPort 2.1b, HDMI 2.1b<\/td>\nDisplayPort 2.1b, HDMI 2.1b<\/td>\n<\/tr>\n
TDP\/TBP, W<\/td>\n315<\/td>\n335<\/td>\n220<\/td>\n304<\/td>\n285<\/td>\n320<\/td>\n300<\/td>\n360<\/td>\n<\/tr>\n
Retail Price (USA), $<\/td>\n899 (recommended at launch)<\/td>\n999 (recommended at launch)<\/td>\n549 (recommended at launch)<\/td>\n599 (recommended at launch)<\/td>\n799 (recommended at launch)<\/td>\n999 (recommended at launch)<\/td>\n749 (recommended at launch)<\/td>\n999 (recommended at launch)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

In turn, the GeForce RTX 5070 Ti is based on the same GB203 GPU used in the GeForce RTX 5080, but the lower model has lost 14 streaming multiprocessors, meaning 1,792 FP32-compatible shader ALUs, and its gaming clock frequency has been reduced. As a result, the GeForce RTX 5080 has 27% more theoretical performance than the RTX 5070 Ti, and the RTX 5070 Ti itself does not differ from the RTX 4070 Ti SUPER in this regard. Similar to the Radeon RX 9070 XT, the lack of raw computing power is compensated by the new GPU architecture, primarily through the ability to generate multiple frames with DLSS (up to three generated frames between two 'real' ones). <\/p>\n

The GeForce RTX 5070 Ti is rated for a power consumption of 300 watts and, with an MSRP of $749, it is $50 cheaper than the GeForce RTX 4070 Ti SUPER.<\/p>\n

The difference of $150 in the recommended currency value between the Radeon RX 9070 XT and the GeForce RTX 5070 Ti is not noticeable given the current Russian prices: the most affordable partner models we found in DNS are priced at 99,999 and 104,999 rubles, respectively. These are the models we will be testing in benchmarks. The Radeon RX 9070 XT features the GIGABYTE GAMING OC graphics card with a factory overclocked GPU up to 2.52\/3.06 GHz, while the GeForce RTX 5070 Ti is the Palit GamingPro accelerator with reference specifications.<\/p>\n

\u2193<\/a><\/noindex>#<\/a><\/noindex>GIGABYTE Radeon RX 9070 XT GAMING OC: Design<\/strong><\/h2>\n

The GIGABYTE version of the Radeon RX 9070 XT has fairly modest dimensions for a graphics card with a power consumption of at least (and, as tests will show, much more than) 300W: 288 mm in length and three slots in width. The only element of the LED backlighting is a short strip on the side. The manufacturer's logo badge is movable: if shifted onto the LED strip, the logo begins to illuminate.<\/p>\n

\"New <\/p>\n

The cooling system is serviced by three fans with a blade diameter of 87 mm. Air passes straight through a small section of the radiator thanks to a cutout in the backplate.<\/p>\n

\"New <\/p>\n

The graphics processor and the memory chips transfer heat to the vapor chamber, but the former does so directly, while the latter does it through a U-shaped metal spacer. Individual plates are attached to the radiator fins, which serve as heat sinks for the VRM components. There are a total of seven heat pipes with a diameter of 6 mm. It should be noted that GIGABYTE opted for liquid thermal pads instead of conventional ones. This may explain how the VRAM chips heat up under load (see the empirical section of the review).<\/p>\n

\"New <\/p>\n

The backplate aids in dissipating heat from the printed circuit board via several thermal pads (which are conventional, not liquid).<\/p>\n

\"New <\/p>\n

\u2193<\/a><\/noindex>#<\/a><\/noindex>GIGABYTE Radeon RX 9070 XT GAMING OC: Printed Circuit Board<\/strong><\/h2>\n

Some variations of the Radeon RX 9070 XT receive power through a 12V-2\u00d76 connector, but GIGABYTE opted for three universal eight-pin connectors. Given the high power consumption of the graphics card and ongoing complaints about the new connector, this is a more reliable, though less convenient and aesthetically pleasing solution. Of the 17 phases in the power system, 14, managed by the Monolithic Power Systems MP2868A PWM controller, are dedicated to the graphics processor, while three are allocated for the memory. Both VRMs are equipped with power stages MP87993 with an assumed current rating of 90A.<\/p>\n

The VRAM is comprised of SK hynix chips, where the marking H56G42AS8DX-014 encodes a bandwidth of 20 Gbps.<\/p>\n

In addition to the BIOS copy, which is active by default, the graphics card has a backup 'silent' firmware.<\/p>\n

\"New \"New <\/p>\n

\u2193<\/a><\/noindex>#<\/a><\/noindex>Palit GeForce RTX 5070 Ti GamingPro: Design<\/strong><\/h2>\n

The Palit GeForce RTX 5070 Ti GamingPro measures 331.9 mm in length but occupies only three expansion slots. RGB lighting LEDs are located under the GamingPro logo and adjacent section of the front panel. The operation of the LEDs can be synchronized with the motherboard through the standard ARGB connector located near the 12V-2\u00d76 power input.<\/p>\n

\"New <\/p>\n

The side surfaces of the casing are formed by a die-cast aluminum frame with ventilation slots on the long sides. The protective plate on the reverse side of the PCB features a familiar grid for airflow through the heatsink. The fan blades have a diameter of 94 mm.<\/p>\n

\"New <\/p>\n

At the base of the cooler lies an evaporation chamber of sufficient size to cover the GPU die and its surrounding memory chips, while separate plate-type heat sinks are provided for the VRM components. The radiator fins are threaded onto eight heat pipes with a diameter of 6 mm.<\/p>\n

\"New <\/p>\n

The metal backplate serves only a protective and decorative function, as there are no thermal pads between it and the PCB.<\/p>\n

\"New <\/p>\n

The graphics card comes with an adapter from two eight-pin power connectors to 12V-2\u00d76 and a small fabric mouse pad.<\/p>\n

\"New <\/p>\n

\u2193<\/a><\/noindex>#<\/a><\/noindex>Palit GeForce RTX 5070 Ti GamingPro: printed circuit board<\/strong><\/h2>\n

Unlike the graphics card Palit GameRock<\/a><\/noindex>, which we covered in a recent review of the GeForce RTX 5080, the next model in the series is built on a spacious PCB, but its power system is simpler. The GPU has a 15-phase VRM managed by the PWM controller Monolithic Power Systems MP29816. The power for the GDDR7 memory chips is three-phase, based on the MP2988 controller. Both regulators are equipped with MP87993 transistor assemblies with a presumed current rating of 90 A.<\/p>\n

The marking of the Samsung K4VAF325ZC-SC28 memory chips indicates a bandwidth of 28 Gbps.<\/p>\n

Like the GIGABYTE Radeon RX 9070 XT, the Palit accelerator has two BIOS options - 'performance' and 'quiet'.<\/p>\n

\"New \"New <\/p>\n

\u2193<\/a><\/noindex>#<\/a><\/noindex>Test bench, testing methodology<\/strong><\/h2>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Test Stand<\/th>\n<\/tr>\n
CPU<\/td>\nAMD Ryzen 9 7950X3D (PBO +150 MHz, CU -20)<\/td>\n<\/tr>\n
Motherboard<\/td>\nASUS ROG Crosshair X670E Hero<\/td>\n<\/tr>\n
Memory<\/td>\nG.Skill Trident Z5 Neo RGB (F5-6000J3040G32GX2-TZ5NR), 2 \u00d7 32 GB (6200 MT\/s, CL30)<\/td>\n<\/tr>\n
SSD<\/td>\nSolidigm P44 Pro, 2 TB<\/td>\n<\/tr>\n
Power Supply<\/td>\nCorsair AX1600i, 1600 W<\/td>\n<\/tr>\n
CPU Cooling System<\/td>\nCustom liquid cooling (EK-Quantum Velocity\u00b2 DDC 4.2 PWM D-RGB + EK-Quantum Surface X280M)<\/td>\n<\/tr>\n
Case<\/td>\nOpen stand<\/td>\n<\/tr>\n
Operating system<\/td>\nWindows 11 Pro<\/td>\n<\/tr>\n
GPU Software for AMD<\/td>\n<\/tr>\n
All graphics cards<\/td>\nAMD Software Adrenalin Edition 25.2.1\/25.3.1<\/td>\n<\/tr>\n
GPU Software for NVIDIA<\/td>\n<\/tr>\n
All graphics cards<\/td>\nNVIDIA GeForce Game Ready Driver 532.60\/532.70<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Games without ray tracing<\/th>\n<\/tr>\n
Game<\/td>\nAPI<\/td>\nTesting Method<\/td>\nGraphics Settings<\/td>\n<\/tr>\n
Alan Wake 2<\/td>\nDirectX 12<\/td>\nOCAT, Bright Falls location<\/td>\nMax. graphics quality<\/td>\n<\/tr>\n
Black Myth: Wukong<\/td>\nDirectX 12<\/td>\nBuilt-in benchmark<\/td>\nMax. graphics quality<\/td>\n<\/tr>\n
Cyberpunk 2077<\/td>\nDirectX 12<\/td>\nBuilt-in benchmark<\/td>\nMax. graphics quality<\/td>\n<\/tr>\n
F1 24<\/td>\nDirectX 12<\/td>\nBuilt-in benchmark, Monaco track (rain)<\/td>\nMax. graphics quality<\/td>\n<\/tr>\n
Hogwarts Legacy<\/td>\nDirectX 12<\/td>\nOCAT, cart ride in Path to Hogwarts<\/td>\nMax. graphics quality<\/td>\n<\/tr>\n
Horizon Zero Dawn Remastered<\/td>\nDirectX 12<\/td>\nBuilt-in benchmark<\/td>\nMax. graphics quality<\/td>\n<\/tr>\n
Metro Exodus<\/td>\nDirectX 12<\/td>\nBuilt-in benchmark<\/td>\nMax. graphics quality; Shading Rate: 100%<\/td>\n<\/tr>\n
Red Dead Redemption 2<\/td>\nVulkan<\/td>\nBuilt-in benchmark<\/td>\nMax. graphics quality<\/td>\n<\/tr>\n
Returnal<\/td>\nDirectX 12<\/td>\nBuilt-in benchmark<\/td>\nMax. graphics quality<\/td>\n<\/tr>\n
Total War: WARHAMMER III<\/td>\nDirectX 11<\/td>\nBuilt-in benchmark (Mirrors of Madness Benchmark)<\/td>\nMax. graphics quality<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Games with ray tracing<\/th>\n<\/tr>\n
Game<\/td>\nAPI<\/td>\nTesting Method<\/td>\nGraphics Settings<\/td>\nFrame scaling<\/td>\n<\/tr>\n
AMD<\/td>\nIntel<\/td>\nNVIDIA<\/td>\n<\/tr>\n
Alan Wake 2<\/td>\nDirectX 12<\/td>\nOCAT, Bright Falls location<\/td>\nMax. graphics quality, high ray tracing quality<\/td>\nFSR Balanced<\/td>\nFSR Balanced<\/td>\nDLSS Balanced + Ray Reconstruction (+ Frame Generation)<\/td>\n<\/tr>\n
Black Myth: Wukong<\/td>\nBuilt-in benchmark<\/td>\nMax. graphics and ray tracing quality (Path Tracing)<\/td>\nFSR Balanced (+ Frame Generation)<\/td>\nXeSS Balanced\/FSR Balanced + Frame Generation<\/td>\nDLSS Balanced (+ Frame Generation)<\/td>\n<\/tr>\n
Cyberpunk 2077<\/td>\nBuilt-in benchmark (OCAT for frame generation)<\/td>\nMax. graphics and ray tracing quality (Path Tracing)<\/td>\nFSR Balanced (+ Frame Generation)<\/td>\nXeSS Balanced\/FSR Balanced + Frame Generation<\/td>\nDLSS Balanced (Transformer Model) + Ray Reconstruction (+ Frame Generation)<\/td>\n<\/tr>\n
F1 24<\/td>\nBuilt-in benchmark, Monaco track (rain)<\/td>\nMax. graphics quality and ray tracing<\/td>\nFSR Balanced (+ Frame Generation)<\/td>\nXeSS Balanced\/FSR Balanced + Frame Generation<\/td>\nDLSS Balanced (+ Frame Generation)<\/td>\n<\/tr>\n
Hogwarts Legacy<\/td>\nOCAT, cart ride in Path to Hogwarts<\/td>\nMax. graphics quality and ray tracing<\/td>\nFSR Balanced<\/td>\nXeSS Balanced<\/td>\nDLSS Balanced + Ray Reconstruction (+ Frame Generation)<\/td>\n<\/tr>\n
Indiana Jones and the Great Circle<\/td>\nOCAT, location Sukhothai<\/td>\nMax. graphics quality, high quality ray tracing (Path Tracing); Ray Traced Shadows: Only Sunlight<\/td>\nFSR Balanced (+ Frame Generation)<\/td>\nXeSS Balanced\/FSR Balanced + Frame Generation<\/td>\nDLSS Balanced (Transformer Model) + Ray Reconstruction (+ Frame Generation)<\/td>\n<\/tr>\n
Metro Exodus Enhanced Edition<\/td>\nBuilt-in benchmark<\/td>\nMax. graphics quality and ray tracing<\/td>\nN\/A<\/td>\nN\/A<\/td>\nDLSS Balanced<\/td>\n<\/tr>\n
Returnal<\/td>\nBuilt-in benchmark (OCAT for frame generation)<\/td>\nMax. graphics quality and ray tracing<\/td>\nFSR Balanced (+ Frame Generation)<\/td>\nXeSS Balanced\/FSR Balanced + Frame Generation<\/td>\nDLSS Balanced (+ Frame Generation)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

In most games, average and minimum frame rate (we indicate the 1st percentile of the distribution) is derived from an array of individual frame render times or the instant frame rate obtained with the built-in benchmark. The exception is games without a built-in benchmark and tests using frame generation: in these cases, to capture the inter-frame intervals, we use the OCAT program.<\/p>\n\n\n\n\n\n\n\n\n
Productivity applications<\/th>\n<\/tr>\n
Application<\/td>\nBenchmark<\/td>\nSettings<\/td>\n<\/tr>\n
Adobe Premiere Pro 25.x<\/td>\nPugetBench for Premiere Pro 1.1.1 (test composition in the link)<\/a><\/noindex><\/td>\nStandard (4K)<\/td>\n<\/tr>\n
Blender 4.x<\/td>\nDemo Agent 327 Barbershop from Blender's website<\/td>\nCycles Renderer<\/td>\n<\/tr>\n
Blackmagic Design DaVinci Resolve Studio 19.x<\/td>\nPugetBench for DaVinci Resolve 1.1.1 (test composition in the link)<\/a><\/noindex><\/td>\nStandard (4K); H.264\/HEVC Encoding Mode: Auto<\/td>\n<\/tr>\n
CAD applications<\/td>\nSPECviewperf 2020 v3.1<\/td>\nScreen resolution: 3840 \u00d7 2160<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Video decoding (ffmpeg 5.x)<\/th>\n<\/tr>\n
Format<\/td>\nResolution<\/td>\nEncoding settings<\/td>\nAPI<\/td>\n<\/tr>\n
H.264 (YUV 4:2:0, 8 bits\/channel)<\/td>\n1920 \u00d7 1080<\/td>\nHigh Profile, L4.1<\/td>\nD3D11VA<\/td>\n<\/tr>\n
3840 \u00d7 2160<\/td>\nHigh Profile, L5.1<\/td>\n<\/tr>\n
HEVC (YUV 4:2:0, 8 bits\/channel)<\/td>\n1920 \u00d7 1080<\/td>\nMain Profile, L4.0<\/td>\n<\/tr>\n
3840 \u00d7 2160<\/td>\nMain Profile, L5.0<\/td>\n<\/tr>\n
7680 \u00d7 4320<\/td>\nMain Profile, L6.0<\/td>\n<\/tr>\n
VP9 (YUV 4:2:0, 8 bits\/channel)<\/td>\n1920 \u00d7 1080<\/td>\nN\/A<\/td>\n<\/tr>\n
3840 \u00d7 2160<\/td>\n<\/tr>\n
7680 \u00d7 4320<\/td>\n<\/tr>\n
AV1 (YUV 4:2:0, 8 bits\/channel)<\/td>\n1920 \u00d7 1080<\/td>\nMain Profile, L4.0<\/td>\n<\/tr>\n
3840 \u00d7 2160<\/td>\nMain Profile, L5.0<\/td>\n<\/tr>\n
7680 \u00d7 4320<\/td>\nMain Profile, L6.0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Video encoding (ffmpeg 5.x)<\/th>\n<\/tr>\n
Format<\/td>\nResolution<\/td>\nEncoding settings<\/td>\nAPI<\/td>\n<\/tr>\n
AMD<\/td>\nIntel<\/td>\nNVIDIA<\/td>\nAMD<\/td>\nIntel<\/td>\nNVIDIA<\/td>\n<\/tr>\n
H.264 (YUV 4:2:0, 8 bits\/channel)<\/td>\n1920 \u00d7 1080<\/td>\n-c:v h264_amf -quality speed -coder cabac -refs 1 -b:v 3M<\/td>\n-c:v h264_qsv -preset veryfast -profile:v main -level 4.1 -b:v 3M<\/td>\n-c:v h264_nvenc -preset fast -coder cabac -refs 1 -b:v 3M<\/td>\nAMF<\/td>\noneVPL<\/td>\nNVENC<\/td>\n<\/tr>\n
3840 \u00d7 2160<\/td>\n-c:v h264_amf -quality speed -coder cabac -refs 1 -b:v 7.5M<\/td>\n-c:v h264_qsv -preset veryfast -profile:v main -level 5.1 -b:v 7.5M<\/td>\n-c:v h264_nvenc -preset fast -coder cabac -refs 1 -b:v 7.5M<\/td>\n<\/tr>\n
HEVC (YUV 4:2:0, 8 bits\/channel)<\/td>\n1920 \u00d7 1080<\/td>\n-c:v hevc_amf -quality speed -b:v 3M<\/td>\n-c:v hevc_qsv -preset veryfast -tier main -b:v 3M<\/td>\n-c:v hevc_nvenc -preset fast -b:v 3M<\/td>\n<\/tr>\n
3840 \u00d7 2160<\/td>\n-c:v hevc_amf -quality speed -b:v 7.5M<\/td>\n-c:v hevc_qsv -preset veryfast -tier main -b:v 7.5M<\/td>\n-c:v hevc_nvenc -preset fast -b:v 7.5M<\/td>\n<\/tr>\n
7680 \u00d7 4320<\/td>\n-c:v hevc_amf -quality speed -b:v 20M<\/td>\n-c:v hevc_qsv -preset veryfast -tier main -b:v 20M<\/td>\n-c:v hevc_nvenc -preset fast -b:v 20M<\/td>\n<\/tr>\n
AV1 (YUV 4:2:0, 8 bits\/channel)<\/td>\n1920 \u00d7 1080<\/td>\n-c:v hevc_amf -quality speed -b:v 3M<\/td>\n-c:v av1_qsv -preset veryfast -profile main -b:v 3M<\/td>\n-c:v hevc_nvenc -preset fast -b:v 3M<\/td>\n<\/tr>\n
3840 \u00d7 2160<\/td>\n-c:v hevc_amf -quality speed -b:v 7.5M<\/td>\n-c:v av1_qsv -preset veryfast -profile main -b:v 7.5M<\/td>\n-c:v hevc_nvenc -preset fast -b:v 7.5M<\/td>\n<\/tr>\n
7680 \u00d7 4320<\/td>\n-c:v hevc_amf -quality speed -b:v 20M<\/td>\n-c:v av1_qsv -preset veryfast -profile main -b:v 20M<\/td>\n-c:v hevc_nvenc -preset fast -b:v 20M<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

The power of the graphics cards is recorded separately from the CPU and other PC components using the NVIDIA PCAT device. For the power and noise level tests, the game Cyberpunk 2077 is used at a resolution of 3840 \u00d7 2160 with maximum graphics quality settings (no ray tracing), as well as the FurMark stress test with the most aggressive settings (resolution 3840 \u00d7 2160, MSAA 8x). All parameters are measured after warming up the graphics card when the GPU temperature and clock frequencies stabilize.<\/p>\n

\u2193<\/a><\/noindex>#<\/a><\/noindex>Testing participants<\/strong><\/h2>\n

The following graphics cards participated in the performance testing:<\/p>\n