The rumor that Sony plans to launch its own upscaling technology with the upcoming PlayStation 5 Pro is followed by a wealth of tidbits about what the next-gen console's hardware will look like. The CPU will still be the same old Zen 2 core and the RAM will be slightly faster, but the GPU will be replaced with an entirely new design. And it seems to be a crazy mix of RDNA 3 and RDNA 4.
The leaks in question come from Inside Gaming And although these are unfounded claims, the details seem credible enough to me. For example, it is said that the PS5 Pro's custom AMD APU retains the same Zen 2 CPU as in the original PlayStation 5. So that's still eight cores and 16 threads, but that's fine for most games.
It won't run any faster, although there will apparently be an additional operating mode that will allow it to use more power to achieve a 10% higher clock speed. This comes at a cost to the GPU, which receives less power in this mode but only loses about 1% of its performance. Why shouldn't Sony go for a faster CPU? The answer to that is simple: backwards compatibility. All PS5 games and older games that also run on the platform expect a CPU that runs at 3.5GHz at best. Changing that too much can mess up a lot of things.
The same applies to system RAM. There's 16GB of average speed GDDR6 included in each PS5, and access to it is shared between the CPU and GPU. At 14 Gbps on a 256-bit aggregated memory bus, a paltry 448 GB/s of total bandwidth is available. The PS5 Pro will apparently offer a bandwidth of 576GB/s and assuming the bus width has changed, that equates to a speed of 18Gbps.
To put these numbers in perspective, the Radeon RX 6800 and RX 7800 XT both have 256-bit buses, but the former uses 16Gbps GDDR6 while the latter uses 19.5Gbps chips for a bandwidth of Have 624 GB/s. On paper, this would make the PS5 Pro look particularly good, as the RX 7800 XT is no slouch, but this GPU also has 64MB of Level 3 cache (aka Infinity Cache) to reduce the load on VRAM.
There's no sign that the PS5 Pro's APU will get Infinity Cache, as AMD and Sony need to keep the chip as small as possible to keep manufacturing costs down. But since the CPU And Since the GPU shares the same RAM pool, the 576 GB/s bandwidth is severely impacted. Especially when you read about the claimed changes to the graphics processor.
In the original PS5, the GPU is almost similar to an RDNA-2 processor, with 36 Compute Units (CUs) paired with 18 Workgroup Processors (WGPs). In each CU there are two banks of 32 ALUs that manage all shaders. According to the PS5 Pro rumors, the new GPU has 30 WGPs, just like the RX 7800 XT, for a maximum FP32 throughput of 33.5 TFLOPS.
If both of these numbers are correct, this could only happen if the CUs are based on the RDNA-3 architecture, meaning they have two banks of 64 ALUs. A quick calculation on a piece of paper puts the boost clock at 2,180MHz, which is lower than the PS5, but the additional shaders more than make up for it.
However, we were also told that the Ray Accelerators in the CUs support BVH8 traversal shaders. BVHs, or bounding volume hierarchies, are data structures designed to more quickly find out which object a light beam interacts with. In RDNA 3, the shaders work on four BVH children of each node (also known as BVH4).
Increasing it to eight theoretically means the PS5 Pro GPU can handle the throughput process much faster than before. At least potentially, because it's not as simple as “the number is doubled and so is the performance”. While there is no indication that AMD has moved these operations from the CUs to dedicated hardware units (like Nvidia does with its GPUs), the change in the traversal shaders shows us that this is the case with the new GPU not a pure RDNA 3, but an RDNA 3/4 hybrid.
It's unlikely to be a full RDNA 4 design as it doesn't have an Infinity Cache, but if all the numbers are to be believed, it's more RDNA 4 than 3. The key sign of this is the claim the AI ​​accelerators that first appeared in RDNA 3. In the PS5 Pro chip they should achieve 300 TOPS in INT8 mode and 67 TFLOPS in FP16 mode.
For example, the RDNA 3-powered Radeon RX 7800 XT can perform 512 INT8/FP16 operations per CU per clock cycle on its AI accelerators, meaning peaks of 75 TOPS and 75 TFLOPS, respectively. For comparison: The Tensor cores in Nvidia's GeForce RTX 3090 Ti have a maximum INT8 throughput of 320 TOPS.
What use would all of this be? AI, be it in the form of upscaling or any other task to which machine learning can be applied.
Everything is at this stage if so And But, and there is no guarantee that any of this is true. But machine learning and ray tracing performance has been RDNA's Achilles' heel so far, so given the importance of machine learning workloads these days, it makes perfect sense that AMD would improve them significantly for RDNA 4.
I'm unlikely to buy a PlayStation 5 Pro when it comes out, considering I already have a perfectly decent PS5 and a variety of gaming PCs in use at home, but from one perspective According to tech authors, things are looking quite promising.
However, from a GPU enthusiast's perspective, RDNA 4 looks extremely interesting: it may not land at the top of the performance charts, but it looks like it has all the features you could want.
