How Powerful is the PS5 Pro? (Part 2)

Overnight, a Twitter user, "brunno_fast", has uploaded a photo from the Playstation Pro Safety Guide revealing some hitherto unknown information about the PS5 Pro. Additionally, we have a teardown from "TAG" over on YouTube! The dam has broken and Sony no longer has control over the flow of information...

Let's dig into the details!

Safety Specs...

Let's start with the details:

Some interesting differences, there...

From my cursory glance, I spot the following differences between the specs listed for the Pro and base digital console:

• 16.7 TFLOPs of GPU compute vs 10 TFLOPs
• DDR5 2 GB listed in the memory section
• 2 TB SSD vs 825 GB in storage
• Replacement of a Hi-speed USB-A port with a Hi-speed USB-C port (480 Mbps)
• Wireless support for 802.11 be (WiFi 7)
• Higher current draw: 1.7 A vs 1.46 A
• Higher Wattage of the PSU: 390 W vs 340 W
• Lighter weight: 3.1 kg vs 3.4 kg

This is a confusing mish-mash of updates for the Pro console over the base model!

On the one hand, we have the slow USB connector updated to a type-C, instead of one or both of the USB-A Superspeed (5 Gbps) connectors updated - which makes zero sense to me! We have a reduction in weight, which would imply a smaller (or at least less dense) heatsink, despite the larger chassis size. We also have the TFLOP number of 16.7, which is not the previously leaked value of 33.4 TFLOPs - mind you, the PS5 base model's 10.23 TFLOP capability is also similarly misreported as 10 TFLOPs... but that's a smaller discrepancy.

On the other hand, we have the increased power and current requirements - which tally with each other: Wattage = Voltage x Current; Base = 240 x 1.46 = 350.4 W, Pro = 240 x 1.7 = 408 W. However, we know they're not pulling 100% and without 100% efficiency (plus some overhead for safety) so divide the two to get a ratio of 1.16 and multiply the 340 W by that ratio and you get ~394 W - close enough to the new value of 390 W to be accurate. 

However, the point still stands - power use has gone up...

Previously, I had looked at the power efficiency of RDNA2 compared with RDNA3 and come away with the conclusion that a move to RDNA3 would allow Sony to keep similar power levels to the base unit - which is now confirmed to be untrue.

Then, there's that 2 GB of DDR5 memory... which is a bit of a mystery to me. It seems like this is the primary reason for freeing up memory for games. However, does it have its own memory bus, like the PS4 Pro did? Or will it eat up some of that precious 256 bit main bus bandwidth? The other side of the coin is that a typical 2 TB SSD with DRAM cache would have a 2 GB DRAM chip as buffer. Could we have an increase in that, as well?

Larger L0, L1 and L2 caches are a feature of RDNA3 over RDNA2...

Flip-flopping...

Coming back to Digital Foundry's coverage from several months ago, the GPU specs detailed by them were as follows:

• 30 WGPs (60 CU)
• GPU L0 cache 32 kb
• GPU L1 cache 256 kb
• GPU L2 cache 4 MB
• Up to 2.35 GHz GPU clock frequency but confirming the averaging of around 33.4 TFLOPs

Those cache sizes match well with what we observe in the RDNA3 Shader Engine architecture. However, the teraflop value listed in the official specs of the PS5 Pro is matching only that of an RDNA2 design - with single a ALU pipeline per Compute Unit (CU). Similarly, the Pro is actually matching the increase in power usage of an RDNA2 design - despite being on the same* or smaller manufacturing process node (either TSMC N6 or N4). We're looking at a 1.42x power increase from RX 6700 to RX 6800 on N7 but, taking into account the process node optimisations and a slightly increased clock frequency, 1.16x could be a feasible number for N4...

I say slightly increased clock frequency because, like Digital Foundry state in that segment linked above, the calculations put the core clock frequency at 2174 MHz (the stock desktop equivalent RDNA2 part has a boost frequency of 2105 MHz) but to increase further to potential 2350 MHz, would likely require an associated reduction in CPU power draw to achieve...

*I say "the same" but there are some optimisations on N6 versus the N7 node used to manufacture RDNA2. So, we would actually expect some power-savings if all things remained the same...

So, this has left me with a bit of a question in my mind. The leaked specifications from the developer portal all point to RDNA3 being utilised in the Pro. However, the official specs point to the ALU numbers per CU of RDNA2 being present.

Logically speaking, Sony would NOT "sandbag" about the real TFLOP capabilities of the PS5 Pro - they have no incentive to do so. AMD's RDNA3 and Nvidia's Ampere do not pretend to be capable of doing less work than they can theoretically manage. Which professionally-acting, long-established company would do such a thing?

No, the only conclusion is that the TFLOP value in the official consumer documentation is accurate. Taking these conclusions and data at face value, it appears that the developer portal documentation that was leaked could not have been accurate or was purposefully falsified in some manner in order to throw-off leakers and damage their credibility.

Programmer Matias Goldberg gave his thoughts on architectural changes...

My own PS5 Pro simulations pointed to this outcome, throughout this whole time period. Especially when I showed that, even when CPU-limited, the RDNA3 part would have an uplift of 1.62 - 1.70x the peformance of over the PS5 base at a core clock frequency of 2230 MHz. Whereas the uplift in performance on the equivalent RDNA2 part over the base of 1.42 - 1.43x, again at 2230 MHz! 

During my prior calculations, I also worked out that 2x the raytracing performance would leave the theoretical PS5 Pro at just above RX 6800-levels of RT performance, with 2.5x equating to approximately RX 7700 XT-levels of performance.

I have also pointed out that updating to RDNA3 might cause software compatibility issues. There are ways around that - by being able to "gate-off" various features but things like changes cache size are transparent to sofware in terms of programme knowledge but opaque in terms of performance optimisations. If you have written your programme to expect certain data movement behaviour and optimised for it, changing the cache size will mess with the expected output of the programme and potentially cause problems.

For the time being, until ironclad evidence is presented, I have to return to my prior standpoint that the PS5 Pro is using 60 CU of RDNA2 design. There is just too much conflicting data to confidently state that RDNA3 is in use in the cache and CU design!

Video by TAG, comparison of the original PS5 base from iFixit...

Increasing Memory...

The new circuitboard design is interesting. Sony have dispensed with the expensive custom-cut layout, instead opting for a cheaper rectangular cut - that will save some money and really did seem like a silly extravagance in the original design. The SSD controller and DRAM cache are still shifted to the rear of the board, as was the case with the PS5 Slim design. We still have four flash memory modules as per the Slim (though probably now 512 MB each). However, the DRAM module doesn't appear different - though I can't read the model specs from the video - so, maybe no increase in buffer compared to the increased size of the SSD, which would be disappointing.

Judging from the layout, the extra chip in the bottom left-hand corner is the 2 GB DDR5 and appears to be linked directly to the APU - so it seems it is connected on a different bus than the GDDR6 memory modules, just like the PS4 Pro!

Ironically, the 16.7 TFLOPs of compute has been online in Asia for at least some time, this information has only just broken into the Western tech sphere...

Conclusions...

This was just a quick breakdown of the new (and final) information we're getting about the Playstation 5 Pro. From this information, I'm no longer confident in the prior claims that the console is RDNA3-based. All data and even my own calculations and hardware-based simulations have pointed to 60 CU RDNA2 and I had been unable to explain the discrepancy in performance based on that work compared to the claims of 1.45x performance uplift and 2-3x the raytracing calculation performance. RDNA3 just provides too much of a performance uplift compared to those definitive claims from Sony.

In this context, the extra power requirements are understandable and the extra DRAM to help with game memory allocation is apparently needed and, given that it will likely sit on its own bus, will potentially leave more bandwidth for games to utilise!

There are also some nice cost-optimisations on the Pro model internal design, including the heatsink and heatpipes - which I didn't get into in this article, but which look better at addressing the heat transfer from the APU, as well as the separate system for both VRMs, memory modules, and SSD controller and DRAM.

Finally, the seemingly random switch from the Type-A to Type-C adapter for the slowest USB port makes little sense to me and thus maybe a cost-reduction exercise...

Anyway, that's all for today. Let me know any thoughts in the comments or on Twitter!