GPU Architecture and specs

GPU architecture and specs

You may expect several graphics cards within the new 6000 lineups. AMD strickly is focussing on high-end to enthusiast-class with this release. But we have no doubt that RDNA2 based architecture of course will find its way to the lower spectrum of the product line over time as well. For the initial Big Navi release, we'll see the two Radeon RX 6800 (XT) graphics cards and a Radeon RX 6900 XT; they all are based on the same GPU holding a whopping 26.8 billion transistors. Inevitably, of course, there will be a Radeon RX 6700 series released as well. AMD claims a 50% perf per watt improvement in performance, double that of last-gen performance. Let's first have a look and overview of the specifications per product.

As you can see AMD has been able to nearly double up the shader/stream processor count for a full enabled GPU. However take a look at the difference in transistor count, yes that's a big chip alright, reasonably explained due to the new infinity cache, as well as additional RT cores. As you can observe, a fully enabled Big Navi GPU holds 80 CUs (Compute units); each CU holds 64 shading/stream cores. So when multiplied you'll end up at 80x64= 5120 shader cores. Each CU has 1 RT (Raytracing) core thus 80 cores for the 6900XT, 72 for the 6800 XT, and 60 for the 6800. Historically AMD has had 4 texture units per CU, so that's 320 units for the 6900 XT, 288 for 6800 XT, and 240 units for the 6800. The ROP count is 128 units for 6800 XT and 6900 XT, 96 for the 6800. 

Infinity Cache

One of the biggest changes from previous GPU architecture is Infinity Cache (IC). Why IC? well, the choice of GDDR6 memory is a far cheaper approach than what NVIDIA is doing with GDDRX. This will help in the bill of materials for a graphics card. However, gDDR6 tied to 16GB AMD faces the challenge that the memory bus is a bit limited at 256-bits. 512-bit is complicated signal and wires wise so they figured that adding a level of cache memory will take the load off the memory bus. and that helps tremendously in performance per watt really, but also greatly helps raytracing.  So very simply put, IC is cache memory, and that cache memory (128MB) is placed directly into the chip itself (on-die). This also is one of the reasons that the Navi 21 GPU is considerably larger than Navi 10. Normally a GPU has a few megabytes of cache memory (L1 and L2). Then there's a huge gap in between the many gigabytes of VRAM that the frame buffer has. This gap is bridged with Infinity Cache. Arbitrarily speaking you could look at IC as an L3 cache that is more capable to provide the GPU with sufficient and faster data in a faster manner and reduces frame buffer utilization.

AMD injected 128MB Infinity Cache used in combination with a 256-bit memory bus provides more than twice the bandwidth of a 384-bit memory bus. So a lot of the transistor budget was used for this feature. The cache does not just help in performance, as the GPU now takes advantage of lower energy consumption because there is less utilization on the memory controllers. 128MB Infinity Cache is implemented and used on all the RX 6900 XT and 6800 XT and RX 6800 graphics cards. In our findings, this big 'L3' cache helps out greatly.  

Smart Access Memory 

You will have noticed that AMD introduced SAM, smart Access memory. AMD's Smart Memory Access gives the new Zen 3-based Ryzen 5000 series processors full access to GPU memory if you have a Zen 3 chip on an X570 chipset with the new RNDA 2-powered Radeon RX 6000 series graphics card that is. Normally you are limited to a block of 256MB max. Smart Access Memory thus efficient use of the combined memory of the CPU and GPU, reducing buffering and latency - great news for gamers and it utilizes the PCI-express bus. It is suggested this can brings in another 5 to10 % additional bonus in your framerate. SAM requires that CSM Support is turned off in the BIOS in order to enable the above 4G decode function, which will allow Resizable Bar Support (SAM) to be enabled. One problem here is that if your Windows installation is configured as non-UEFI, Windows will be unable to boot. In order to test it, you would need to reinstall Windows with CSM support disabled. Probably 9 out of 10 Windows 10 PCs these days have CSM support enabled. Also and please understand this is mandatory, is a Series 500 chipset based motherboard, a Ryzen Seris 5000 processor, and RX 6000 Series processor.

Raytracing

So what is raytracing really? Well, with raytracing, you basically are mimicking the behavior, looks, and feel of a real-life environment in a computer-generated 3D scene. Wood looks like wood, however, the leaking resin will shine and refract its environment and lighting accurately. Glass and waves of water get refracted as glass based on the surroundings and lights/rays. Can true 100% raytracing be applied in games? Short term answer, no, partially. As you have just read and hopefully remembered, Microsoft has released an extension to DirectX; DirectX Raytracing (DXR). AMD now has dedicated hardware built into their GPUs to accelerate certain raytracing features. You have seen these in current games mostly as Shadow optimization, but most commonly used are reflections (water, puddles, windows, tiles, and so on). Rasterization has been the default renderer for a long time and you can add to that a layer of raytracing. Combining rasterization and raytracing we like to call Hybrid raytracing. and offers the best of both worlds. But make no mistake, the RT cores inside Big Navi can do full scene raytracing, however, would never be fast enough for real-time rendering.  

A DLSS alternative? ML Super Sampling - DirectML 

DLSS is a proprietary NVIDIA solution tied to Tensor cores in an NVIDIA RTX GPU.  AMD however is working on more open support for a somewhat similar technology. DLSS is an acronym for "deep learning super sampling." An AI based algorithm. That feature reduces the workload of the GPU without having significant losses in visual quality. Why is this important? Well, it boosts your framerate and thus compensates the losses of Hybrid Ray Tracing performance. Combined you can game in 4K resolution and preferably above 60 frames per second. AMD will answer this challenge, as simply put they'll need to do that, that answer is super-resolution DirectML. Like NVIDIA, AMD plans to use machine learning to improve the visual quality of games, but AMD's solution will be supported by Microsoft.Microsoft confirmed that its two new consoles already support DirectML Super Resolution. AMD has the opportunity to replace DLSS with an alternative that will apply to gamers on both PC and Xbox, giving developers little reason to support DLSS over DirectML Super Resolution. DirectML is not an exclusive Radeon technology and its applications extend far beyond super-resolution capabilities.  AMD's Super Resolution technology is currently in development and should be released as part of AMD's cross-platform GPU Open / FidelityFX technology suite. It has not stated when this will be released and how big the performance hit will be, as AMD does not offer any dedicated hardware for this to accelerate this functionality.  When released though, it'll likely be located under AMD's FidelityFX suite dubbed Super Resolution. FidelityFX technologies currently include graphical tools such as Contrast Adaptive Sharpening, Ambient Occlusion, Variable Shading, Screen Space Reflections, and new, a Denoiser that should help out in Raytracing.