The Caveat Wars: Gap Width Measurements *spin ad nauseam*

Supposedly but that nice theory is not supported by the facts we have. For instance in BF1, PS4 drops down to ~30fps where it's CPU heavy in the 64 players mode, XB1 version drops to ~30fps in scenes where there is lots of fog (so alphas).

In BF1 in the modes where it's less CPU heavy (but still with lots of alphas) both versions run very similarly (with a very slight advantage for PS4 actually using the average fps), while PS4 still having the usual ~40% resolution advantage.

In Titanfall 2 a game with tons of alphas, the XB1 game drops down to 480p when the PS4 game drops down to 720p in those cases (both running at solid 60fps).

My take on this is that while esram may give an advantage in those cases, PS4 has twice the number of ROPS the XB1 has.

 

Yeah, the one is a more balanced system used mostly for a naturally inbalanced task: gaming.
Of course it was designed with the assumption gaming was gonna be just a fraction of its use for most consumers, just before MS had the rude awakening that core gaming was still the soul of console early consumers.
So, being a balanced computer is a terrible design principle for consoles, and past success/failures cohoborate that.

 

But the netbook cpu inside the PS4 (and the One) is terrible for AI. Of course that was the best choice then because Zen wasn't available but still. DDR3 is a better choice for scattered read/writes.

http://techreport.com/news/27168/assassin-creed-unity-is-too-much-for-console-cpus

 

Balance
When you are writing an intensive CPU code you must think only about CPU cache and minimize cache misses. Lower DRAM latency does not help much (also GDDR5 does not have much more latency).
Fast scratch pad is good but it is relatively small.

 

This whole discussion is pretty pointless to be honest. The gap in power boils down to 1080p vs 900p in 3rd party titles like it always has. CPU intensive titles can have a slight advantage in performance on the Xbox one, while GPU intensive games have a slight advantage in performance on PS4.

The differences in resolution come down to less compute units and a more complicated memory setup.

All of these points are old news. Both systems have some beautiful 1st party titles which we all know are the true show ponies on consoles.

 

Of course it's pointless. I've got this theory though that to get a true noticeable leap in the types of games you play you need at least a consequent 5-6x leap in cpu performance. I think Dead Rising on the 360 was the first game to show something that could not be achieved on the previous gen. I think the CPU upgrades in this gen are a halfway step to something better and consequently games don't feel like a true evolution.

Imagine for instance walking through time square in new york in a GTA style game with thousands of people being controlled by their own AI routines.

 

Or, for extra realism, they all could be controlled by the media and the Illuminati new world order.

 

Jaguar is a descendant to Bobcat. Bobcat netbooks had 2 cores. Console Jaguars have 8 cores. Consoles are thus ~4x faster in multithreaded code than netbooks. Jaguar also has twice as wide vector units (64 bit vs 128 bit) and supports AVX, both nice additions regarding to games. It is modern OoO core. Not bad for branchy pointer heavy code like game AI (not to be confused with neural network based AI which runs best on GPU). Previous gen PPC based consoles were much worse in running general purpose code.

Intel's 20K$ Xeon Phi (Knights Landing) is also based on a netbook CPU. It has 72 modified Airmont Atom cores. It is the same CPU core used in netbooks, but the vector units are extended to be 4x wider (AVX-512).

Big amount of power efficient (slower) cores are better fit for large scale AI (GTA city for example) compared to a few high clocked ones. Lots of small cores = higher computational throughput for same TDP. I wouldn't expect a 4 core Zen (at 3.0 GHz) to be noticeably better than 8 core Jaguar (at 1.6 GHz) in game AI. 8 core Zen has too high TDP for console APU, unless severely downclocked. Most of the TDP of console APUs are allocated to the GPU.

 

I think jaguar probably hits that mark, at least for most game code. 2x the cores of the 360, plus, from what I understand they're something like 3-5x faster for general purpose code, maybe more in branch heavy code.

 

Well, one jaguar core is supposedly about a quarter as powerful as one Core i5 2500k core so if Zen is only as powerful as Sandy Bridge then 8 jaguar cores could be roughly equal to 4 Zen cores. However I'm hoping the next-gen consoles will feature at least 8 cores maybe on 7nm.

 

Actually just realised, the jaguar is only half the power of 2500K, so even worse.

 

Okay maybe that number increases on a logarithmic scale...so something like 10x?

 

IIRC, the general consensus was that the in-order PPC CPUs of last gen consoles reached around 0.2 IPC in generic code. Hand optimized inner loops of course ran faster. I remember reading from somewhere that Bobcat reaches around 0.9 IPC in generic code. Jaguar has slightly higher IPC than Bobcat -> let's estimate 1.0 -> Jaguar's IPC is 5x better than the previous gen PPC cores in general purpose code. PPC had double clocks -> Jaguar core is roughly 2.5x faster than old PPC core. 3 cores vs 8 cores lead to roughly 6.5x total performance increase in generic (unoptimized) code, assuming it is threaded perfectly. This is obviously a ballpark estimate, depending greatly on what kind of code you run.

PPC can get pretty close to Jaguar in some special cases like VMX128 vectorized inner loops (VMX128 had single cycle throughput multiply-add instruction). But getting that kind of throughput from the VMX128 unit requires acrobatics. 128 (x2) vector registers are there because you actually NEED them to reach any kind of decent throughput. Huge unrolled loops with pure VMX code is the only way to get close to the maximum flop rate. You need to separate your VMX code completely from other code, as the VMX pipeline latency is so long and there are huge stalls when you move data between VMX<->scalar registers. Cell SPUs had considerably shorter vector pipelines. But I would assume that generic game AI code doesn't use either. So Jaguar is a much better CPU choice than PPC/Cell.

 

I never thought I would see such thread in this webpage. Fanboys' war even in here? Serously? C´mon!

 

Where have you been since the console forums were born?

 

sandy bridge ? The tests AMD has released put it 5% faster than Skylake in Blender.

 

Its alright, the fanboy wars have mostly ended in the console forums. Unfortunately they remain extremely alive in the graphics card forums with Nvidia and AMD participants.

They existed before console forums and remain after them, namely the Graphics card sections with Nvidia and anyone else.

 

at least we can get some good info from sebbbi.

 

That should be worse case scenario. The benchmarks released by AMD could be contrived and massaged so to speak seeing as blender is open source. We'll see in about 10 days.

 

still though the newest spins of bulldozer coming out are getting ipc close to sandy bridge. Something would have had to go total wrong for AMD to still be 5 + years behind intel.