Capabilities and Specs for Future Radeon GPUs Suggested with Xbox Series X Details and Leak

[Guest_Jim_*]

When AMD released its Navi-based Radeon RX 5000 series GPUs it succeeded in providing an amount of disruption it had less success at previously. The one area these graphics cards have not entered yet, however, is the high end where NVIDIA has been secure with the likes of the GTX 1080 Ti, RTX 2080 Super, and RTX 2080 Ti, but AMD has stated it will release a GPU for that segment. In fact one of AMD's employees stated it intended to disrupt performance at 4K similar to Ryzen disrupted performance for CPUs. Now thanks to Microsoft sharing some details about the GPU in the upcoming Xbox Series X as well as a leak with some specifications for a discrete graphics card, we may have an appreciably better idea of what to expect from AMD this year.

The current Navi GPUs utilize AMD's new RDNA architecture, which has replaced the GCN design AMD has been using for many years. A number of tweaks and improvements came with RDNA, but AMD also announced RDNA 2 at the same time, stating this architecture would feature the company's solution for hardware accelerated ray tracing. This is something the Xbox Series X GPU will feature and now Microsoft has specifically stated it will use RDNA 2, as opposed to a custom design between RDNA and RDNA 2.

Another feature RDNA 2 will have is variable rate shading, VRS, which you can already find in NVIDIA's Turing-based GPUs and some Intel CPUs. Shading is the process for determining the final colors of pixels and is actually somewhat intense work. Normally the shading rate is the same across all pixels in a frame, but with VRS it can be decreased for better performance and would normally be applied to the background, so the resulting loss in image definition is not easily noticed. It is also possible to leverage VRS to increase the shading rate for pixels, reducing performance but improving image quality, though I have not seen any indications it is being used this way. For a console, supporting VRS may prove quite valuable but it does also have significant performance for desktop graphics, especially for those who demand higher frame rates.

According to Microsoft the RDNA 2 GPU in the Xbox Series X will offer 12 TFLOPS of compute performance, which is pretty significant as that surpasses the RX 5700XT. The RX 5700XT is capable of 9.754 TFLOPS for FP32 computations (math using 32-bit floating point numbers) and if we assume Microsoft is also referring to FP32 performance, that is more than 20% faster. Of course TFLOPS do not directly relate to graphics performance, but this is definitely an impressive improvement. It is worth noting, however, that TFLOPS performance should not blindly be compared between GPU architectures as there can be differences. The RX Vega 64 offers 12.66 TFLOPS at FP32 but the RX 5700XT is able to surpass it in graphics and in some workloads, thanks to its redesigned compute units. Still, Microsoft states this GPU will be twice as powerful as that in the Xbox One X and eight times that of the original Xbox One. The Xbox Series X will also support HDMI 2.1, which sports features such as variable refresh rate and auto low latency mode, so we should be able to expect that coming from AMD with its RDNA 2 GPUs.

Also appearing recently is an image supposedly from an SK Hynix document with specifications for a new GPU, possibly the RX 5950XT, which WCCFTech has examined. Being a leak though, this information should be treated as a rumor and thus with some amount of skepticism. According to the document, this GPU will feature 80 compute units, which translates to a total of 5120 shading units, as well as 96 ROPs and 320 TMUs. The RX 5700XT has just 40 CUs and just 64 ROPs, which translates to 2560 shading units and 160 TMUs so we are looking at double the rendering hardware. With the RX 5700XT already being a very effective GPU for 1440p gaming, this speculated RX 5950XT would be quite powerful.

Paired with these components will be 12 MB of L2 cache, which is three times that of the RX 5700XT, and 24 GB of HBM2e on a 4096-bit bus with a bandwidth of 2048 GB/s. Though the current Navi GPUs are all using GDDR6, it appears AMD will indeed return to some form of HBM for the products at the top of its stack. The Vega-based GPUs all used HBM2 and HBM2e is a step up, allowing for greater bandwidth per stack and for 12-Hi stacks but AMD might not go with such high stacks. The more dies stacked together to form HBM, the more capacitive and resistance limits you run into. In any case, this will potentially enable as much as 2 TB/s of memory bandwidth.

What is not indicated by this leak is when the GPU may release, or a price. With these specifications though, this potential RX 5950XT would definitely be a top-of-the-stack product and call for such a price as well. That naming does allow for RX 5950 non-XT, RX 5900, and RX 5800 GPUs with other, less expensive configurations though. That is of course assuming this is all true.

What we do we can expect is the RDNA 2 GPUs will be made using TSMC's 7 nm+ node, which may prove less expensive over time as it requires fewer photolithography masks and steps than the current 7 nm process of the foundry's. We can also expect some improvement to efficiency as it was stated at CES 2020 that methods learned from working on the Ryzen 4000-series APUs could be applied to other GPUs. A top-end discrete graphics card will likely still pull more power than mid-range products, but it may be able to do more with that power. Hopefully it will not be long before these products are unveiled and we can see what AMD has achieved and if it will finally bring competition back at the top-end.

Source: Microsoft and WCCFTech

Back to original news post

[Guest_Jim_*]

Have updated the news item as SK Hynix has released a stated that the document is not from them and that specifications given on it for HBM2e are incorrect. On the bright side, it did not take long to learn the accuracy of the rumor, but now we do need to get back to waiting on information for what will be correct.

SKY Hynix statement