Digital Foundry Article Technical Discussion Archive [2013]

[jonabbey]

No it didn't.

It's still astonishing just how big The Duke was.

 

[Shifty Geezer]

Have you ever heard of ambilight tech? Anyway, health was just a random example.

Ambilight is behind the screen, not reflecting off it. Move could get pretty intrusive with its cyan and magenta screen presence. And regardless of the feedback being present, health or something else, putting that feedback outside of the player's line of sight and obscured behind a controller makes no sense. For that, you'd want at least a coloured LED on the top of the controller, such as the touch pad glowing. Which of course wouldn't help with camera tracking which is what I think the real purpose is. The coloured light is for motion tracking, on a camera Sony decided not to include as standard, and now the coloured light is a very obvious controller differentiator and not much else.

 

[Gubbi]

No it didn't.

Mea culpa. Guess it was down to ergonomics then. The S felt a lot better then the original controller

Cheers

 

[liolio]

So "at least" double the performance from an 8 core Jaguar CPU.. Where would that put it in terms of effective performance?

It is tough to estimate. The best I found is Anandtech A4-5000 review where Kabini (@1.5GHz) is compared to a dual core IVB @1.7GHz.
The result are a bit all over the place, the IVB is in turn 50%, 100% and in a corner case 200% faster (Kraken). Haswell is significantly better when you take advantage of the AVX2 and FMA instructions.

With a gross estimate of 2 cores Haswell is around twice as fast as a quad core kabini then a 8 jaguar set-up should be around the perfs of the 2 cores Haswell. Now if you do around twice as much in a console (which sounds as much as a gross estimate as the one I'm doing... so lets call it even... ) you also have desktop parts in the PC realm that run at around twice the clock.

It's a gross estimate, not true under any circumstances, clock speeds are not exactly the same (between the A4-5000, the jaguar in the next gen consoles supposedly 1.6GHz, the IVB runs a tad faster 1.7GHz), but like the x2 statement it should be good enough has a ballpark figures, not too mention that on consoles some core are reserved for the OS, in Windows environment well it depends but I usually don't do much while playing and pretty much the OS and services take really few resources, though some other users are different).

Imo I don't know if the CPU is going to be a bottleneck for next gen, I would think it is a nice improvement from say Xenon, but if PC ports (from consoles) benefit from the optimization done on console I think a lot of PC gamers won't have much intensive to upgrade their quad cores running @+3GHz. I would think that a Core i3 $xxx should be good enough if one doesn't aim at crazy FPS , I would think that any AMD quad core should also good enough. If AMD delivers I expect desktop parts based on streamrollers to run around next gen as far as CPU perfs are concerned. Those desktop CPUs are damned fast (some would say that they burn power accordingly).

 

[Aeoniss]

It is tough to estimate. The best I found is Anandtech A4-5000 review where Kabini (@1.5GHz) is compared to a dual core IVB @1.7GHz.
The result are a bit all over the place, the IVB is in turn 50%, 100% and in a corner case 200% faster (Kraken). Haswell is significantly better when you take advantage of the AVX2 and FMA instructions.

With a gross estimate of 2 cores Haswell is around twice as fast as a quad core kabini then a 8 jaguar set-up should be around the perfs of the 2 cores Haswell. Now if you do around twice as much in a console (which sounds as much as a gross estimate as the one I'm doing... so lets call it even... ) you also have desktop parts in the PC realm that run at around twice the clock.

It's a gross estimate, not true under any circumstances, clock speeds are not exactly the same (between the A4-5000, the jaguar in the next gen consoles supposedly 1.6GHz, the IVB runs a tad faster 1.7GHz), but like the x2 statement it should be good enough has a ballpark figures, not too mention that on consoles some core are reserved for the OS, in Windows environment well it depends but I usually don't do much while playing and pretty much the OS and services take really few resources, though some other users are different).

Imo I don't know if the CPU is going to be a bottleneck for next gen, I would think it is a nice improvement from say Xenon, but if PC ports (from consoles) benefit from the optimization done on console I think a lot of PC gamers won't have much intensive to upgrade their quad cores running @+3GHz. I would think that a Core i3 $xxx should be good enough if one doesn't aim at crazy FPS , I would think that any AMD quad core should also good enough. If AMD delivers I expect desktop parts based on streamrollers to run around next gen as far as CPU perfs are concerned. Those desktop CPUs are damned fast (some would say that they burn power accordingly).

It's probably also worth mentioning all of the dedicated silicon to offload processes from the CPU as well on the PS4 APU..

I think it'd be helpful to know how much CPU is really and truly 'available' for use on PC's. At least on Windows PC's. How much game developers can reasonably expect to use.

From a GPU perspective I think the PS4 is fairly potent out the gate compared to other generations at launch vis a vis their PC counter-parts. Excluding balls-out absurdity like TITAN and HD 7990's at any rate.

CPU, though, I just don't know really what to make of the Jaguar. I mean Cell was so radically different and Xenon was a PowerPC architecture (same with cell) 'traditional' design with 3 cores. Hard to make real comparisons to desktop parts there.

 

[sebbbi]

It is tough to estimate. The best I found is Anandtech A4-5000 review where Kabini (@1.5GHz) is compared to a dual core IVB @1.7GHz.
The result are a bit all over the place, the IVB is in turn 50%, 100% and in a corner case 200% faster (Kraken). Haswell is significantly better when you take advantage of the AVX2 and FMA instructions.

With a gross estimate of 2 cores Haswell is around twice as fast as a quad core kabini then a 8 jaguar set-up should be around the perfs of the 2 cores Haswell. Now if you do around twice as much in a console (which sounds as much as a gross estimate as the one I'm doing... so lets call it even... ) you also have desktop parts in the PC realm that run at around twice the clock.

That 1.7 GHz IVB is rarely running at it's nominal 1.7 GHz frequency, especially in pure CPU benchmarks. It turbos up to 2.6 GHz. Because the CPU has shared TDP with the integrated GPU, you always get a considerable turbo boost in all synthetic CPU benchmarks (as GPU is idling). Similarly a 1.9 GHz IVB turbos up to 3.0 GHz, and that's already a 2x clock rate advantage over the 1.5 GHz Kabini for pure CPU benchmarks. Hard to compare the real IPC in this scenario.

Console games are specially optimized for a single set of hardware. All the available CPU cores are used (all games use all cores, there's no single threaded games), and all the GPU cycles are always used (no TDP left unused by the GPU). On PC you can frequently turbo clock up to +50% advantage. On a console, you could only turbo clock if the game was poorly programmed (didn't use all cores or left some GPU cycles unused). So the PC based pure CPU benchmark comparison is not valid.

If you compare IVB and Jaguar architectures, you will notice that IVB is around 2x capable per clock in many ways. It's SIMD units process 2x8 floats per cycle instead of 2x4 floats. It can sustain (decode and retire) four uops per cycle vs two (both scalar and vector uops). It can take instructions from two threads instead of one (Hyperthreading). A single IVB core can sustain exactly the same amount of scalar instructions (or SIMD flops) per cycle than two Jaguar cores. So basically a 1.7 GHz dual core IVB should generally match a 1.7 GHz quad core Jaguar (when all cores are 100% fed and turbo cannot be used). And a double clocked (3.4 GHz) dual core IVB should again match a hypothetical 1.7 GHz 8 core Jaguar.

If we go down to gritty details, we will notice that a single IVB core is actually better than two equally clocked Jaguar cores in some cases, and worse in others.

First, each Jaguar core has their own 32KB + 32KB L1 instruction and data caches, while IVB core has just one set of 32KB + 32KB caches shared between two threads. Jaguar has 1 MB of L2 cache shared for pair of cores (2 MB for four cores), while IVB core (running two threads) has just 256 KB of L2 cache. Clearly IVB doesn't double the resources here, and thus doesn't reach the 2x performance goal in (common) memory bound scenarios. However IVBs caches have lower latency, and it has a large shared L3 cache to back up the smaller caches (getting back some lost ground).

Second, IVB core can load balance between two threads. Two Jaguar cores are fully separate (except for the L2 cache). If one IVB thread stalls (full cache miss = wait 150+ cycles for memory request), the other thread gets the whole (four uop sustained) CPU core for itself (for the next 150+ cycles). In general, even the highly optimized (game) programs have some cache misses, and thus gain extra performance from the additional memory latency hiding provided by Hyperthreading. Some parts of game code, such as script execution (LUA, Flash, Python, etc) and complex search algorithms (from sparse data structures) have significant gains. However the flip side of the coin is extra cache contention. A single IVB core doesn't have as large caches as two Jaguar cores combined. It must execute two threads using half the cache storage for each, so the caches trash easily, and in absolutely worst case cache trashing can hurt performance more that Hyperthreading helps.

While IVB core is capable of twice as many floating point SIMD operations per cycle compared to a Jaguar core, it is only capable of equal amount of (8, 16 and 32 bit per lane) integer SIMD operations per cycle. So in this metric, two Jaguar cores beat single IVB core (2x sustained rate per clock). Haswell is the first chip to support 256 bit wide integer operations. And Haswell also doubles the peak theoretical flops by adding FMA (however no real code is pure 100% FMA instructions, so real gains are closer to +50%, which is still very nice). Ivy/Haswell also beat Jaguar in 64 bit integer and floating point math by a wide margin (Jaguar 64 bit per lane SIMD operations are not full rate). However this doesn't affect game performance much (scientific software however would run considerably better on Ivy/Haswell).

 

[TheChosenOne]

That 1.7 GHz IVB is rarely running at it's nominal 1.7 GHz frequency, especially in pure CPU benchmarks. It turbos up to 2.6 GHz. Because the CPU has shared TDP with the integrated GPU, you always get a considerable turbo boost in all synthetic CPU benchmarks (as GPU is idling). Similarly a 1.9 GHz IVB turbos up to 3.0 GHz, and that's already a 2x clock rate advantage over the 1.5 GHz Kabini for pure CPU benchmarks. Hard to compare the real IPC in this scenario.

Console games are specially optimized for a single set of hardware. All the available CPU cores are used (all games use all cores, there's no single threaded games), and all the GPU cycles are always used (no TDP left unused by the GPU). On PC you can frequently turbo clock up to +50% advantage. On a console, you could only turbo clock if the game was poorly programmed (didn't use all cores or left some GPU cycles unused). So the PC based pure CPU benchmark comparison is not valid.

If you compare IVB and Jaguar architectures, you will notice that IVB is around 2x capable per clock in many ways. It's SIMD units process 2x8 floats per cycle instead of 2x4 floats. It can sustain (decode and retire) four uops per cycle vs two (both scalar and vector uops). It can take instructions from two threads instead of one (Hyperthreading). A single IVB core can sustain exactly the same amount of scalar instructions (or SIMD flops) per cycle than two Jaguar cores. So basically a 1.7 GHz dual core IVB should generally match a 1.7 GHz quad core Jaguar (when all cores are 100% fed and turbo cannot be used). And a double clocked (3.4 GHz) dual core IVB should again match a hypothetical 1.7 GHz 8 core Jaguar.

If we go down to gritty details, we will notice that a single IVB core is actually better than two equally clocked Jaguar cores in some cases, and worse in others.

First, each Jaguar core has their own 32KB + 32KB L1 instruction and data caches, while IVB core has just one set of 32KB + 32KB caches shared between two threads. Jaguar has 1 MB of L2 cache shared for pair of cores (2 MB for four cores), while IVB core (running two threads) has just 256 KB of L2 cache. Clearly IVB doesn't double the resources here, and thus doesn't reach the 2x performance goal in (common) memory bound scenarios. However IVBs caches have lower latency, and it has a large shared L3 cache to back up the smaller caches (getting back some lost ground).

Second, IVB core can load balance between two threads. Two Jaguar cores are fully separate (except for the L2 cache). If one IVB thread stalls (full cache miss = wait 150+ cycles for memory request), the other thread gets the whole (four uop sustained) CPU core for itself (for the next 150+ cycles). In general, even the highly optimized (game) programs have some cache misses, and thus gain extra performance from the additional memory latency hiding provided by Hyperthreading. Some parts of game code, such as script execution (LUA, Flash, Python, etc) and complex search algorithms (from sparse data structures) have significant gains. However the flip side of the coin is extra cache contention. A single IVB core doesn't have as large caches as two Jaguar cores combined. It must execute two threads using half the cache storage for each, so the caches trash easily, and in absolutely worst case cache trashing can hurt performance more that Hyperthreading helps.

While IVB core is capable of twice as many floating point SIMD operations per cycle compared to a Jaguar core, it is only capable of equal amount of (8, 16 and 32 bit per lane) integer SIMD operations per cycle. So in this metric, two Jaguar cores beat single IVB core (2x sustained rate per clock). Haswell is the first chip to support 256 bit wide integer operations. And Haswell also doubles the peak theoretical flops by adding FMA (however no real code is pure 100% FMA instructions, so real gains are closer to +50%, which is still very nice). Ivy/Haswell also beat Jaguar in 64 bit integer and floating point math by a wide margin (Jaguar 64 bit per lane SIMD operations are not full rate). However this doesn't affect game performance much (scientific software however would run considerably better on Ivy/Haswell).

Great info! So the 1.6Ghz 8 core jaguars in the PS4 and XBO are, roughly, equal to a 3.2 Ghz core i3 (ivy bridge)? Jaguar wins in some cases and i3 in others...

 

[sebbbi]

Great info! So the 1.6Ghz 8 core jaguars in the PS4 and XBO are, roughly, equal to a 3.2 Ghz core i3 (ivy bridge)? Jaguar wins in some cases and i3 in others...

There's no official information about Xbox One CPU yet (it is an unnamed eight core CPU), and no specifics about PS4 cores either. I am purely talking about the Jaguar cores found in PC (laptop/tablet) Kabini/Temash APUs. AMD has released full spec sheets about the Jaguar CPU cores used in these APUs and full instruction rate/latency sheets. Intel has similar spec sheets available for Ivy bridge, so the comparison is easy to do.

You could say that a dual core 1.6 GHz Ivy Bridge (turbo disabled) equals quad core Jaguar (Kabini) at same clocks, and a quad core 1.6 GHz Ivy would likely equal a eight core (double Jaguar CU) Kabini model at the same clocks. Dual core 3.2 GHz Ivy would roughly mach a quad core 1.6 GHz Ivy when theoretical throughput is considered. However doubled core clocks increase the memory latency from CPUs point of view. The CPU must wait twice the clocks cycles to get the data from the memory, assuming memory speed stays intact. Because of this, the 3.2 GHz model wouldn't provide linear gains over the 1.6 GHz model. But neither does a quad core provide linear gains over a dual core (multithreaded scaling is never 100% efficient).

 

[TheChosenOne]

Thanks for the clarification, sebbbi. Yes, I'm aware you were going by publicly released/known info about Kabini and Ivy Bridge. I was just extending that to what is "generally" speculated as what will/may be in the consoles.

 

[Aeoniss]

Is this assessment including or not including the "atleast 2x" performance advantage due to low level access and closed box advantages?

 

[TheChosenOne]

Well, it's all based on theoretical performance. Not real world performance, so no "at least 2x" performance there. No way to quantify that at this point, if ever.

Unless someone on the inside spills the beans.

 

[BRiT]

But why? Why is it better to have the left thumb more forwards than the right thumb?

Because in typical games you want your thumbs on the controls you're using the most. Typically this means using left thumb for movement and the right thumb for buttons. In this design, the thumbs are exactly the same distance forward, they're just used for different activities. Left thumb for directional control and right thumb for button presses.

 

[(((interference)))]

Yeah, I was playing the Last of Us and did notice it was a bit awkward moving forward or sprinting with the DualShock since the left movement thumbstick is so inboard that you have to more pull it forward rather than push (which is much easier to do when the thumbstick is above your thumb like on the 360 controller).

 

[djskribbles]

People still argue about the stick placement? It's personal preference... I bet there are just as many people that prefer a symmetrical design than an asymmetrical design, and there are many that don't care either way. I do see how someone coming from a 360 controller might find the DS3 awkward, but it's the same for me when I hold a 360 controller. The way I hold the DS3, it's not uncomfortable at all. The DS4 looks to be a bit better for people that prefer an asymmetrical setup because the sticks are in a higher position, and maybe slightly closer to the outside.

edit: http://www.ps4site.com/news/dualshock-4-vs-dualshock-3/

 

[RedVi]

I've had every DS since the first and I agree that asymmetrical sticks are more comfortable. When I tried out a dreamcast I couldn't really tell since I didn't use it for long enough, but when my friend got a Xbox S controller, it was definitely better than the DS2. The Duke controller made my hands sweat and I hated the skewed face buttons, but when the S controller came around, I could see that it was a better layout.

I've only ever owned a Playstation (and a master system, if that counts). So I personally don't think it's down to what you're used to at all. Asymmetrical sticks are better ergonomically to me. However, if they have changed the grip shape and size which it looks like they have it should be a huge improvement and aid in drastically reducing the awkwardness of the left sticks position.

 

[(((interference)))]

Ergonomics is not strictly down to personal preference.

The DS3 is an older design, Sony just slapped on two sticks to the original PS controller back in 97 and didn't give it much thought.

Both MS and Nintendo chose an assymetric stick layout when designing the GC and original Xbox controllers.

While the PS3's stick layout is better for games that use the dpad extensively (like fighters) the 360's layout is better for everything else. Just try moving forward in an FPS on PS3 vs 360 and you'll see how it's easier to tilt the stick forward when the stick is positioned above your thumb rather than to the right of it.

This is less important for the right stick since you're mainly moving that left and right rather than tilting it completely forward or back (unless you want to stare at the sky or the ground).

This hands on comparison of the two controllers also mentions the XB1's asymmetric layout being more comfortable:
http://stuffmideast.com/2013/06/23/...t-the-xbox-one-controller-or-ps4-dualshock-4/

 

[patsu]

Sony abandoned the boomerang controller for a DS2-like design in DS3.

For the newer controllers, I think we will need to hold them in ours hands to appreciate. I doubt asymmetric placement alone will make or break ergonomics. All the elements complement each other as a whole. Neither will be awkward to use. Both will be comfortable.

Personally, I think the game's control scheme makes more significant differences than mere stick placement (e.g,, lag, aim acceleration, button combo, inverted/default etc.). I have never had any issue with DS3 or the Xbox controllers. I got used to them very quickly.

 

[SedentaryJourney]

Ergonomics is not strictly down to personal preference.

[snip]

This hands on comparison of the two controllers also mentions the XB1's asymmetric layout being more comfortable:
http://stuffmideast.com/2013/06/23/...t-the-xbox-one-controller-or-ps4-dualshock-4/

It seems the only thing backing your statement IS personal preference. I've gamed on both and have my own preferences, but I can perform just as well in any shooter on either controller with ease. I've yet to see an ergonomic comparison between the controllers that would qualify as an objective comparison based on performance metrics and anatomical study, so the ergonomic superiority of one stick placement vs. the other is moot.

 

[djskribbles]

Ergonomics is not strictly down to personal preference.

The DS3 is an older design, Sony just slapped on two sticks to the original PS controller back in 97 and didn't give it much thought.

Both MS and Nintendo chose an assymetric stick layout when designing the GC and original Xbox controllers.

While the PS3's stick layout is better for games that use the dpad extensively (like fighters) the 360's layout is better for everything else. Just try moving forward in an FPS on PS3 vs 360 and you'll see how it's easier to tilt the stick forward when the stick is positioned above your thumb rather than to the right of it.

This is less important for the right stick since you're mainly moving that left and right rather than tilting it completely forward or back (unless you want to stare at the sky or the ground).

This hands on comparison of the two controllers also mentions the XB1's asymmetric layout being more comfortable:
http://stuffmideast.com/2013/06/23/...t-the-xbox-one-controller-or-ps4-dualshock-4/

If it's not personal preference, how do you explain the fact that some people prefer the DS3 and its symmetrical sticks? The only problems I have with the DS3 are the triggers and the rounded sticks, both issues were addressed in the DS4 so I'm happy.

The Gamecube had asymmetrical sticks, but the Wii controller has symmetrical sticks.
http://gdgt.com/nintendo/wii/classic-controller/pro/

The link you posted is just one person's opinion. How does that prove anything?

Honestly I think the whole stick placement thing is overblown and the bigger issue is the shape of the DS3, which Sony seems to have improved significantly. If Sony would've moved the sticks, I wouldn't have been happy, but honestly I wouldn't have cared too much.

 

[Brad Grenz]

Asymmetry versus symmetry has always been a sideshow. People have invented these elaborate theories to say one is better than the other simply because it's the most easily identified difference between a DS3 and the 360 controller. I suspect with the improved ergonomics in the DS4 people will forget they ever thought stick placement was such an important factor. A 360 controller feeling better in the hand has nothing to do with stick placement and everything to do with its more comfortable shape. If the DS4's shape is just as comfortable it won't matter anymore. Personally I find a stick is easier to operate accurately in the lower position than a d-pad, but that could just be another ergonomic factor the Xbox One controller is already correcting.