Baldur’s Gate 3: PS5 Pro Patch Technical Review
Share the Story:
Opening the floodgates
With every new CPU, GPU or console generation the customer base of each bay for blood. The launch of the PS5 Pro is extra special as it delivers all three in one, but the biggest question mark over the consoles long term boosts may stem from the first.
History repeats itself from the PS4 Pro which shared the same 8-core Jaguar CPU of the base PS4, albeit 33% faster clocks. Here the same Zen 2 8c/16t 3.5Ghz CPU is more modestly boosted to a potential 10% higher rate of 3.85Ghz.
Following my early look at another CPU bound title, on PC at least, in Dragon’s Dogma 2 in my full PS5 Pro review and earlier preview. We saw impressive results in that game pointing to an area I often discuss, memory or data bound. As the PS5 Pro is a whopping 50% faster at times than the base PS5 and even 20%+ faster than my 5600X 6c/12t CPU paired with an RX6800 and 32GB of 3600 DDR4 ram. Another collection of CPU limited titles are available and The PS5 Pro patch for Baldur’s Gate 3, that dropped on launch day of the 7th.The teams focus was on improved visuals for the PS5 Pro. I confirmed this with the same run on the PS5 Pro, before and after the patch, demonstrating performance is almost identical here, but visuals are not. The game offers dual modes, both on PS5 and PS5 Pro, with a native 1440p @ 30fps Quality mode and a 60fps upscaled from 1920x1080p base to 1440p mode entitled performance on the base console, which uses FSR2.
We now see a new 4K target in both modes, Quality is native 4K a huge 2.2x pixel hike over the PS5 and the performance mode now uses PSSR to again target 4K, a welcome visual improvement that we can use in a like for like DLSS vs FSR2 comparison a little later in the video, so stick around for that.
Inside Baldur’s Gate itself, CPU and memory become the bottleneck, on PC and console alike. The game supports dual APIs and testing here demonstrated the team did an excellent job getting great thread use our of the ageing DX11. It ended up being 8% on average faster than Vulkan using a selection of my PCs, such as my 5600x 6c/12t CPU and 32GB 3600Mhz DDR4 Ram. Although we see 50-66ms stutters in frame time spikes from the core cache or code issues.
My RTX2070 paired with a 2700 Zen+ CPU does struggle more than this PC, showing it can use threads very effectively helping the slower clocks and IPC. Those extra 2c/4t are not enough leaving it 42.6% faster on the newer Zen3 CPU, not helped by the slower 3200Mhz ram and older architecture. The PS5 and PS5 Pro here do well but not to the same level as we saw in Dragon’s Dogma 2, not something that is a surprise as I often say. Teams must be multi disciplined and if console or PC is your forte, then platforms can gain or lose based on those skill sets or experience.
In this section we are CPU and data bound and the 5600X is an excellent mid-range and high value CPU. Compared to the base PS5, we see no 66ms spikes, and no intrusive tearing which can be resolved on PC with double or triple buffering. Pointing to the fact they are more memory bound on consoles leaving the game running more hand to mouth on render time, causes constant tearing on PS5 and PS5 Pro which was something I hoped the team could resolve by now. This would be a focus for myself if the PS5 Pro has that extra 1GB+ of available ram to the game. The PS5 pitted against the 5600X machine we see a large gap open even with less cores. Highs and lows can be the same to 33% faster on the PC. Averages over this fixed run, using a variety of sections to stress data streaming, rendering and memory use. We are nearly 20% faster than the base PS5 CPU. Which is helped by the 31% clock boost of the 5600x @ 4.6Ghz countered by the 33% extra cores on the PS5 again highlighting paper specs are not linear and real world is never the same as theory.
The PS5 Pro does see an increase over that same comparison though and it can be 33% faster than the base console over the same run through town, ending up 10.5% faster but not enough to close the gap to the 5600x 6c/12t CPU, trailing it by 8.4% on average, again with peaks and troughs. Like PS5 to Pro though, this is so close to be practically identical to the player, but the PC takes the lead here.
In terms of like for like performance the PS5’s CPU, both base and Pro, competes well with the 5600x and 5700x range here, backing up what I said on many prior PS5 reviews and PS5 Pro predictions. The console CPU, memory sub system, dedicated I/O co-processor and API/SDK enable the 3.5Ghz zen 2 based component to compete and even exceed a Zen3 6c/12t 4.6Ghz CPU and memory system in some titles. This is still very good and close but lower than DD2 and other 1st party titles such as Horizon Zero dawn, Spider man, and Ragnarök amongst others. However, the fact is this game is a PC focused game and just as with 1st party or console focused games play to those strengths and can hinder PC Here that PC focus favours that platform and hinders both PS5 machines. Loading is a good way for me to demonstrate this, the game likely does not use the I/O decompression blocks well, or maybe at all, instead using all CPU load and decompression or maybe other overheads here on PS5. The 5600X with a fast SSD but longer load chain is a whopping 58.8% faster than the Pro, which itself highlights the CPU work here being only 6.4% faster than the PS5, leaving the PS5 6% slower than the PS5 pro and 69% slower than the 5600X. This is how Delta’s work.
PSSR vs FSR2 vs DLSS2.4.2
Moving to image quality and we see a mixed blessing of improvements, reinforcing why pixel counts are not all they are cracked up to be. The bump from 1440 to 4K sounds great on paper but in practise the benefit is smaller than many would expect.
FSR2 vs PSSR first and we can see that image stability and sub pixel sampling is far superior with PSSR. Notice the high contrast of the leaves against the sky fizzle and flicker with FSR2 while PSSR maintains a stable and flicker free presentation. The thin rail tracks and moire pattern they create can be a challenge for TAA and upscalers as a collective, and FSR2 struggles here, often a weakness, due to an over sharpened image, dithered sampling and a lack of convergence of pixel weighting. This causes the tracks to jitter and almost ripple constantly, which gets slightly worse when occluded and then disocluded. PSSR handles this brilliantly, with almost no flicker and moire patterns preset, only when the Occlusion test crops up do we see it occur but is quickly resolved due to a better convergence on the pixel data samples over time settling the area back to stability. Shadows also struggle under FSR with pixel trails and break up which we have seen happen often with the AMD solution, not so under PSSR. Finally alpha and sparse shading are worse with FSR2, such as hair cards which, along with the high. Sharpening, cause more ringing and fizzle on camera cuts or motion, see the worst case here against PSSR after a camera cut. Every object is AA free, raw, untouched, naked pixels like they day they were born. Leaving a rough and sharp image that softens and smooths over multiple frames. PSSR solves all this with a much cleaner, less aliased and more pleasing resolve but highlights the main weakness of PSSR in this iteration and implementation at least. Lack of fine detail, notice the tree bark, shades on the leaves and clothes all lack the high frequency information. Again, the hair is more solid, blended, but softer, with less detail present. We see this in all surfaces and even rocks, grass, shadows all remain more consistent of a higher resolution base but can smudge all those fine elements together, which can leave a softer, blurred, albeit more consistent, image that is less prone to noise and flicker.
DLSS v2.4.2 here by default, is surprisingly close, with the gap being much closer than FSR2, in some areas at least but not all. Both handle camera cuts, hair, shadows and the train track pattern test very well. DLSS is better on the balance of jittering noise versus clarity, with hair being the perfect target of the sharpness of FSR2 with the stability of PSSR. Again, texture details, alpha, shadows are all as stable but retain more of the high frequency info which is far lower on PS5 and Pro due to those memory constraints again as textures are lower quality as is sampling. Both keep fizzle and noise to a minimum, with PSSR being slightly better here, such as the train track test causing more jittering noise when disocluded and pixel fizzle around characters as the pass over it, PSSR manages to minimise the breakup and pixel fade and flicker of DLSS. However, the far softer image and reduced clarity mean it comes at a cost that could be worth a trade. As PSSR is really running around the balanced mode, that is closer than quality as you have seen so far, but even comparing the native 4K quality mode to DLSS2 quality confirms that the assets are a large reason for the much lower quality, with it managing improve on the Sony solution and full 4K image within that temporal consistency versus image clarity equilibrium. However, for a launch day First time PSSR solution. Likely from a small team and time, this is a strong showing and certainly exceeds FSR2 as the baseline TAA and Reconstruction solution for the Pro and I am confident it will improve drastically in the coming months ahead.