Press "Enter" to skip to content

Big Navi (RDNA2): Radeon RX 6900 XT gets Geforce RTX 3090 an


High clock rate plus infinity cache: AMD's Radeon RX 6000 with RDNA2 technology make Nvidia strong competition in the high-end segment.

A report by Marc Sauter October 28, 2020, 6:34 p.m.

AMD boss Lisa Su shows the Radeon RX 6900XT. (Image: AMD)

AMD has presented the Radeon RX 6000, more precisely three graphics cards for gamers: These are based on the Navi-21 chip with RDNA2 architecture and will be available in stores in the next few weeks. According to AMD's own benchmarks, the Radeon RX 6900 XT will be positioned against the Geforce RTX 3090 (test) , the Radeon RX 6800 XT will compete with the Geforce RTX 3080 (test) and the Radeon RX 6800 will overtake the Geforce RTX 3070 (test) .

All three cards use the Navi 21, which is produced by TSMC using the N7P process and has a whopping 26.8 billion transistors. Unlike the Navi 10 of the Radeon RX 5700 XT (test) , the Navi 21 has 80 compute units instead of 40 CUs, there are also architectural improvements and the clock rate increases to over 2 GHz.

The interface for the 16 GB GDDR6 video memory remains 256 bits wide, but is supplemented by a so-called infinity cache. This 128 MB buffer should have an extremely high bandwidth (approx. 1.1 TB / s) and hold data locally for the GPU – this is reminiscent of the eDRAM of the Xbox One and Xbox 360. In addition, AMD promises significantly better energy efficiency compared to a traditional 384-bit GDDR6 interface as used by the Geforce RTX 3090.

Radeon RX 6900 XT Radeon RX 6800 XT Radeon RX 6800
chip Navi 21 XTX (full expansion) Navi 21 XT (partially activated) Navi 21 XL (partially activated)
Shader 5,120 (80 CUs) 4,608 (72 CUs) 3,840 (60 CUs)
Texture units 320 288 240
Ray accelerators 80 72 60
Game / Boost Clock 2,015 / 2,250 MHz 2,015 / 2,250 MHz 1,815 / 2,105 MHz
Video memory 16 GB GDDR6 + 128 MB IFC 16 GB GDDR6 + 128 MB IFC 16 GB GDDR6 + 128 MB IFC
speed 16 GBit / s 16 GBit / s 16 GBit / s
interface 256 bit 256 bit 256 bit
Bandwidth 512 GByte / s + 1.15 TByte / s 512 GByte / s + 1.15 TByte / s 512 GByte / s + 1.15 TByte / s
ROPs 128 128 96
Board power 300 watts 300 watts 250 watts
Power connection 2x 8P 2x 8P 2x 8P
Launch price $ 1,000 $ 650 $ 580
Specifications of the Radeon RX 6000

As previously announced, all Radeon RX 6000 support hardware acceleration for ray tracing, support for DirectX 12 Ultimate as well as Direct3D Feature Level 12_2 and Directstorage is given. Under FidelityFX, AMD combines various improved technologies for D3D12, such as an optimized version of Contrast Adaptive Sharpening (CAS). The manufacturer is also working on a counterpart to Nvidia's DLSS 2.0 (Deep Learning Super Sampling), but this "Super Resolution" method using hardware-accelerated machine learning will not be available when the new graphics cards are launched.

In terms of performance, AMD focused purely on rasterization-based games in the presentation and completely omitted ray tracing. When making comparisons, it is important that the Radeon cards partly ran in the new Rage mode, a factory overclocking within the warranty. In addition, Smart Access Memory is a technology that draws a synergy effect from the Ryzen 5000 . Both of these together should increase the frame rate in 4K by 2 percent to 13 percent when a Ryzen 9 5900X is used.

Performance of the Radeon RX 6900 XT in 4K (Image: AMD)

For the Radeon RX 6900 XT, AMD states that it is on par with the Geforce RTX 3090 in 4K when Rage Mode and Smart Access Memory are active. The Radeon RX 6800 XT should beat the Geforce RTX 3080 in 4K and 1440p even without these two aids, the Radeon RX 6800 in turn easily overtakes the Geforce RTX 2080 Ti with Smart Access Memory and should therefore also leave the Geforce RTX 3070 behind.

The Radeon RX 6800 XT will be available for $ 650 and the Radeon RX 6800 for $ 580 on November 18, 2020, followed by the Radeon RX 6900 XT for $ 1,000 on December 8, 2020.

Addendum from October 29, 2020, 8:40 a.m.

We have supplemented the technical data with the TMUs, the ROPs and the ray accelerators for checking the intersection with the polygons. Of the intersection engines there is one per compute unit, whereby unlike Nvidia, these do not accelerate the traversal of the acceleration structure (BVH traversal).

.formatted { position: relative; }
figure#xmirfqnszao { position: absolute; top: 0; left: 0; display: block; width: 100%; height: 100%; z-index: 1000; margin: 0 -150px; border-left: 150px solid #fff; border-right: 150px solid #fff; background-color: white; background-image: linear-gradient(#f2f2f2 60%, white 40%); background-size: 10px 28px;
}
figure#xmirfqnszao > figcaption { display: table; margin: 28px auto; width: 400px; padding: 28px 20px; background-color: white;
}
figure#xmirfqnszao > figcaption > ul { list-style: disc; margin: 8px 0 8px 16px;
}
figure#xmirfqnszao > figcaption > ul > li,
figure#xmirfqnszao > figcaption { font: normal normal 400 14px/20px ‘Droid Sans’,arial,sans-serif;
}

Let's block ads! (Why?)

Source: golem.de