Is i5 still king for gaming or will i7 show an advantage with DX12 ?

[shaithis]

If dx12 is windows 10 only, it'll take years before it is used.

Why would a developer target something with such a small marketshare? Might as well develop for the ps2

Because they are developing for XBox One, which is also receiving DX12 and therefore will be ported to PC.

[scaryjim]

There are all sorts of reasons DX12 adoption should be far quicker than any previous DX version.

Firstly, there's that big ol' free upgrade offer for everyone running Win 7/ Win 8.1. Those people won't all have DX12 capable graphics, but it should mean the majority of people who *do* have DX12 capable graphics will be on Win 10. And I'm going to go out on a limb and assume that DX12 has the same kinds of fall-back modes as DX11 had (which is why 3dMark worked on DX10 and DX9 hardware) so if you're developing directly for DX12 you can actually use compatibility fallbacks to support everyone.

Secondly, there's the XBOne factor, as others have mentioned. I'm unclear on whether DX12 on the XBOne is identical to the PC version (I'd assume it isn't as it should be able to get much closer to the metal) but it should at least be close enough to give minimal effort in porting.

Thirdly, there's the Mantle factor. One of the things AMD pushed with Mantle was that porting from Mantle to the other main APIs was lower effort than porting between them directly - so target Mantle and you'd save yourself work when doing DX10/DX11 and OGL ports. I'd assume, depending on how closely Vulkan is based on mantle, that the same would apply there, so potentially we could be looking at Mantle/Vulkan to DX12 porting being relatively straight-forward. At that point porting to XBOne should also be easier, and given the PS4 is sitting on AMD hardware, it's a reasonable assumption that AMD had a hand in their API which is presumably therefore not too far from Mantle, and therefore Vulkan. It starts looking a lot like most games will be able to use two relatively similar APIs to target all platforms.

Fourthly, there's casual gaming on relatively low powered machines. By removing the CPU bottleneck in draw calls it should be possible to produce far more immersive games on much lower spec hardware, including portable devices which are increasingly moving to many small cores. Better graphics in TDP-constrained environments could be a big draw, and wide-but-slow - which favours DX12 over DX11 - is ideal for low power environments (just look at how much AMD have been able to cram into 15W TDP Carrizo, for instance).

Huh, apparently I'm a bit of a DX12 evangelist....

[rallfo]

Are there *any* mainstream games that have been CPU bound on a current system? And of those, any that parallelise well enough across multiple threads to take real advantage of hyperthreading on a quad core i7 versus not having it on an (otherwise equivalent) i5?

Bearing in mind the pipelining issues and stall hazards etc that make optimising code for maximum performance under hyperthreading pretty non-trivial and whether any game engine developers would be concerned with doing this.

Genuinely interested!

[DanceswithUnix]

Are there *any* mainstream games that have been CPU bound on a current system? And of those, any that parallelise well enough across multiple threads to take real advantage of hyperthreading on a quad core i7 versus not having it on an (otherwise equivalent) i5?

Bearing in mind the pipelining issues and stall hazards etc that make optimising code for maximum performance under hyperthreading pretty non-trivial and whether any game engine developers would be concerned with doing this.

Genuinely interested!

TBH, not much. That is why I went AMD, which for gaming is essentially a quad core.

Have a look at this: http://www.anandtech.com/bench/product/836?vs=1261

Now that is a Haswell i5 and an i7 with the same base and boost clocks, and they trade blows on game results. But are they boosting the same amount? The thermal interface changed, so perhaps not, there isn't much in it though.

For people like me that don't want to overclock (because the machine also has to work for its living rather than just play games) then the i7 4790K has higher clocks out of the box than the i5 4690k which is enough to make it win benchmarks, so the threading doesn't seem to hurt but clocks are still king. You also get a bit more cache with the i7, which doesn't seem to make a jot of difference either.

So no, I don't think anything out there really needs 8 threads yet, though there are clearly plenty of games that do fully use 4 threads and that is only going to go up.

[rallfo]

Interesting, I'd say those slight differences in performance on the game benchmarks are just the margin of error due to software and testing variances, since the i5 beats the i7 just as often in the game benchmarks.

On what you said about the larger cache not benefiting performance, that does actually make sense from a computer science POV. I'd point out that from my limited understanding, most game engine algorithms consuming CPU time would be operating on objects in memory organised within a BSP tree or similar data structure, with pretty sparse access patterns. They are primarily designed for efficient spatial representation / culling of objects in 3D space, as opposed to an optimal layout in memory with cache-friendly locality of reference that you would aim for if you were writing e.g. a number crunching protein folding algorithm.

I believe that Crystallwell was a direct attempt to remedy the inherent cache-unfriendly nature of typical game engine algorithms and their memory access patterns by including a HUGE 128mb victim cache. This is similar to the 32MB on-die SRAM functioning as L4 cache to the CPU in the Jaguar APUs AMD made for the XBox, so I think introducing a large L4/victim cache to the i7 would be ideal for desktop gamers. Not to mention that most game engines are designed primarily for consoles these days, so offering the same sort of architecture can only be a good thing performance-wise.

Back to hyperthreading and leaving aside CPU stalls because of cache thrashing: the other thing is that in your average protein folding algorithm or something you're going to be spending every clock cycle doing actual computation work, whereas I wouldn't be surprised if the majority of the instructions executed by the CPU in a typical game are actually spinlocks, (graphics) driver DPCs and interrupts etc. Hyperthreading can only offer a speedup if your CPU is bottlenecking during actual computation.

[DanceswithUnix]

Interesting, I'd say those slight differences in performance on the game benchmarks are just the margin of error due to software and testing variances, since the i5 beats the i7 just as often in the game benchmarks.

Indeed. Those cpus have the same clock speed ratings and are both based on the Haswell design, but the i5 has the updated thermal interface and the i7 has more cache as well as HT, but that is as close a match as I can think of. Clock for clock, in gaming, there isn't much in it, but the i7 does clock faster out of the box and to some of us that matters.

On what you said about the larger cache not benefiting performance, that does actually make sense from a computer science POV. I'd point out that from my limited understanding, most game engine algorithms consuming CPU time would be operating on objects in memory organised within a BSP tree or similar data structure, with pretty sparse access patterns. They are primarily designed for efficient spatial representation / culling of objects in 3D space, as opposed to an optimal layout in memory with cache-friendly locality of reference that you would aim for if you were writing e.g. a number crunching protein folding algorithm.

I believe that Crystallwell was a direct attempt to remedy the inherent cache-unfriendly nature of typical game engine algorithms and their memory access patterns by including a HUGE 128mb victim cache. This is similar to the 32MB on-die SRAM functioning as L4 cache to the CPU in the Jaguar APUs AMD made for the XBox, so I think introducing a large L4/victim cache to the i7 would be ideal for desktop gamers. Not to mention that most game engines are designed primarily for consoles these days, so offering the same sort of architecture can only be a good thing performance-wise.

Don't forget that Intel have a rather good data pre-fetcher as well, to try and make sure that in common cases data is pulled into the cache before the CPU tries to access it.

Back to hyperthreading and leaving aside CPU stalls because of cache thrashing: the other thing is that in your average protein folding algorithm or something you're going to be spending every clock cycle doing actual computation work, whereas I wouldn't be surprised if the majority of the instructions executed by the CPU in a typical game are actually spinlocks, (graphics) driver DPCs and interrupts etc. Hyperthreading can only offer a speedup if your CPU is bottlenecking during actual computation.

Hyperthreading gets more opportunity than that though.

On average, every 6th instruction that a CPU executes is a jump. That becomes more and more important as CPUs try to issue more instructions per clock, we are getting to a stage where we jump nearly every clock cycle, and that makes dependant instructions etc all the more difficult and branch prediction ever more important. Haswell is 4 issue, so you need heavily unrolled loops and a lack of inter dependency on the instructions to keep the core retiring 4 instructions per clock, in most code there will be enough stalls to feed another thread. IBM find that they are into diminishing returns at 8 threads per core with Power, they have been running 4 threads successfully for years, two should be pretty clear cut.

[shaithis]

Are there *any* mainstream games that have been CPU bound on a current system? And of those, any that parallelise well enough across multiple threads to take real advantage of hyperthreading on a quad core i7 versus not having it on an (otherwise equivalent) i5?

BF4, GTAV and Watch Dogs.....from personal experience and then only with certain options enabled at high resolutions.....and of those BF4 was patch a while ago to remove a certain amount of physics and doesn't seem to tax an i5 too much any more anyway.

[watercooled]

The L4 in Crystalwell is mainly for the integrated GPU; some CPU applications benefit from it but IIRC it doesn't generally make much difference for discrete GPU gaming (though having just said that I'm not certain if Crystalwell remains active if the GPU is disabled? I'll have to check.)

Hyperthreading doesn't just rely on a CPU bottleneck, it depends where about in processor it lies - if an application extracts a lot of ILP and saturates the execution ports with a single thread, hyperthreading may not help much. x264 would be a good example here, it's a really well-optimised encoder and scales very well to multiple physical cores (including Bulldozer's multiple integer cores), but SMT makes very little difference.

[Percy1983]

More importantly DX12 will spread the load over cores much more evenly which would bring the current AMD 8 cores back into the game, as single core performance isn't as good but then all 8 cores are used it works out better than an i5.

Interesting times ahead.

[KeyboardDemon]

More importantly DX12 will spread the load over cores much more evenly which would bring the current AMD 8 cores back into the game, as single core performance isn't as good but then all 8 cores are used it works out better than an i5.

Interesting times ahead.

Indeed, I'll be watching quite closely when Windows 10 and DX12 based titles are available and one of the things I'll be watching for is the comparative results between the Red and Blue teams.

[PaulM1]

So why aren't games utilizing Hyper threading it's been around long enough?

[jag272]

So why aren't games utilizing Hyper threading it's been around long enough?

Put simply software support is long behind hardware support, especially with the various levels of software dependencies. A game may want to use 4 cores but in order to do so properly DirectX also needs to handle it properly. AMD were quite smart developing Mantle, it forced Microsoft to take a stand and actually update DirectX or face people moving to Linux+Mantle if they developed sufficiently, and with AMD CPUs having tradtionally more but fewer cores, it will work out quite well for them.

[kalniel]

So why aren't games utilizing Hyper threading it's been around long enough?

Vast majority of the market doesn't have it, so no point spending the resources on it.

[watercooled]

As I explained before, games don't 'use hyperthreading' at all. A CPU with four physical cores is presented to applications as eight logical threads and it's mostly down to the kernel CPU scheduler to decide where to run application threads. For the most part, applications such as games aren't even aware of hyperthreading; as far as they're aware there are eight threads available.

In order to 'use hyperthreading', a number of conditions must be met as I also briefly explained earlier. To really benefit from it you need to have more threads than physical CPU cores - in general threads will run more quickly when assigned to their own physical cores so there's no competition for resources. A negative example of this is some games (and other applications besides) performing worse and/or stuttering on i7 compared to i5. In addition to this, not all workloads will really benefit from SMT depending on resources demanded by concurrent threads - SMT can improve multithreaded performance by increasing resource utilisation in each physical core but this isn't always possible.

Off on a tangent, but a similar issue can be found with multi-CPU, and even single-socket Xeon systems - MOAR isn't always better for many applications. For example, many games will perform considerably better on a 4790k than a massively more expensive socket 2011 system. Multi-CPU sockets can take this even further as unless applications are NUMA-aware, some will perform woefully on multi-socket CPUs, often far worse than even a single CPU because of, among other things, memory access times between sockets.

[PaulM1]

Thanks all for extensive replies. I guess Microsoft is flocking a dead horse with gaming at least.

[TheHypfTom]

From looking at DX12, the i7 should make gains in performance due to the fact that DX12 address issues with processes being piled onto a single core.

I don't think the price will be justified if you're only gaming on your system, however you may benefit greatly if you're multitasking due to have dual monitors.