Intel Skylake Platform

[Thorburn]

They're low-power, not piss-poor. And if anything, that's a good thing: it means developers will have to optimise their engines for 8 threads in order to get good performance. This optimisation work may then carry over to the PC.

Developers don't have access to 8-cores on these systems, in both cases 2 cores are reserved for the OS and other background processes.

Just means we're going to see a complete lack of CPU intensive developments on this generation of consoles - they are a real disappointment IMO.

 

[Pressure]

Mainstream quad cores go back to Conroe, really. So:

Conroe, Penryn, Nehalem, Westmere, Sandy Bridge, Ivy Bridge, most likely Broadwell and apparently, Skylake.

So 8, although there were only 4 threads before Nehalem, and I don't think Intel ever bothered to actually make a 4C Westmere.

When you say Conroe, you really mean Kentsfield, right? And they were by no means mainstream considering their price ($851 at launch for the cheapest).

There were no quad-core Conroe parts.

 

[Alexko]

Developers don't have access to 8-cores on these systems, in both cases 2 cores are reserved for the OS and other background processes.

Just means we're going to see a complete lack of CPU intensive developments on this generation of consoles - they are a real disappointment IMO.

I didn't know about the reserved cores. Still, 6 Jaguar cores ought to provide a good bit of computing power, and some of the GPU's Compute Units can help too.

When you say Conroe, you really mean Kentsfield, right? And they were by no means mainstream considering their price ($851 at launch for the cheapest).

There were no quad-core Conroe parts.

Yes, wrong codename, my bad (it's been a while!). Kentsfield was expensive but it was "mainstream" insofar as it used the regular LGA775 platform. That's an admittedly loose definition of "mainstream" but then again it quickly dropped to less insane prices, especially considering that at the time AMD didn't have any quad cores to compete with it.

Spoiler

But when AMD did release one, at least it was a true quad-core!

 

[pjbliverpool]

Yes, wrong codename, my bad (it's been a while!). Kentsfield was expensive but it was "mainstream" insofar as it used the regular LGA775 platform. That's an admittedly loose definition of "mainstream" but then again it quickly dropped to less insane prices, especially considering that at the time AMD didn't have any quad cores to compete with it.

You also missed Haswell so it's actually 9.

 

[Albuquerque]

They're low-power, not piss-poor. And if anything, that's a good thing: it means developers will have to optimise their engines for 8 threads in order to get good performance. This optimisation work may then carry over to the PC.

This already exists, to a certain extent, in "modern" PC games ported from the current consoles. Skyrim is a prime example -- at very low clockspeds, you can demonstrate that it scales between dual and quad cores. However, at "normal" clockspeeds, the scaling disappears.

I expect to see similar things for similar reasons coming from the XBONE / PS4 ports to PC; the scaling will likely exist, but only at the lower speeds similar to the Jaguar.

 

[pjbliverpool]

8 core Skylakes at 600Mhz for the win!

 

[Thorburn]

I didn't know about the reserved cores. Still, 6 Jaguar cores ought to provide a good bit of computing power, and some of the GPU's Compute Units can help too.

It really isn't. I've been testing the A4-5000 (4 x 1.5GHz Jaguar cores) and it's somewhere between a Baytrail Z3740 and Z3770 in most CPU performance tests. Okay XBox One bumps that up the 1.75GHz and adds a couple more cores, but you're still talking less performance than say a Core i3 or i5 ULV chips, let alone any quad-core parts.

GPU Compute - do you really want to push more load on to the GPU - on the XBox One in-particular? They're already pretty underwhelming.

 

[homerdog]

It really isn't. I've been testing the A4-5000 (4 x 1.5GHz Jaguar cores) and it's somewhere between a Baytrail Z3740 and Z3770 in most CPU performance tests. Okay XBox One bumps that up the 1.75GHz and adds a couple more cores, but you're still talking less performance than say a Core i3 or i5 ULV chips, let alone any quad-core parts.

I was aboout to say... 6 jaguar cores at 1.6GHz ain't very much more powerful than the Atom CPU in my Asus tablet. Comparing it to a modern LGA115X CPU is ridiculous. Even the dual core Celerons should outrun it.

 

[Thorburn]

I was aboout to say... 6 jaguar cores at 1.6GHz ain't very much more powerful than the Atom CPU in my Asus tablet. Comparing it to a modern LGA115X CPU is ridiculous. Even the dual core Celerons should outrun it.

Baytrail Z3740 will outpace the Sandy Bridge Pentium 987 in some stuff, so HSW Celeron DC would probably be close.

 

[Wynix]

The power of the consoles cores are irrelevant; the point is that virtually every single console game engine will be programmed to use more than 4 cores.

Obviously a quad core Intel CPU can handle the load without issue, that is not where my problems are.
My problem is that we will now see a surge of games supporting 5+ cores and Intel showing no interest to update the status-quo more than just an IPC or perf/watt increase like we have been stuck with for ~8 consecutive generations.

22nm brought fin-fets; 14nm should bring an increase in core counts.

i7 = 6c/12t
i5 = 6c/6t
i3 = 4c/4t
pentium = 2c/4t

Give us more cores and the software will be updated to take advantage of it; Stick with 4 cores and it's obvious why programs will only take advantage of 4 cores.

 

[Thorburn]

The power of the consoles cores are irrelevant; the point is that virtually every single console game engine will be programmed to use more than 4 cores.

Obviously a quad core Intel CPU can handle the load without issue, that is not where my problems are.
My problem is that we will now see a surge of games supporting 5+ cores and Intel showing no interest to update the status-quo more than just an IPC or perf/watt increase like we have been stuck with for ~8 consecutive generations.

But why? What performance benefit will it bring?

If the games are written to run over 6 rather low performance cores then they'll run quite nicely over 4 cores with 3-4x the overall performance, particularly with Hyper-threading.

22nm brought fin-fets; 14nm should bring an increase in core counts.

I'm pretty sure for the majority of users the extra die space is better spent on beefing up the integrated graphics than on more CPU cores which are already hard to fully occupy with typical consumer workloads.

i7 = 6c/12t
i5 = 6c/6t
i3 = 4c/4t
pentium = 2c/4t

Give us more cores and the software will be updated to take advantage of it; Stick with 4 cores and it's obvious why programs will only take advantage of 4 cores.

In some ways I'd actually quite like to see that (add Celeron to the bottom of the stack as 2c/2t), although I'd probably envisage the i5 as being 4c/8t. Problem is you'd have to consider the actual performance delivered within the TDP for mainstream parts. Would a 6c/12t i7 actually deliver any benefits for consumers inside an 84/95w power envelope?

 

[Albuquerque]

The power of the consoles cores are irrelevant; the point is that virtually every single console game engine will be programmed to use more than 4 cores.

Great, so any modern Intel quad core that is equipped with hyperthreading will get benefit -- after all, there are eight logical processors available for assignment of tasks.

Every time this conversation has come up in the past about "ZOMG why doesn't the PC have as many cores as the consoles?!?", it boils down to "It doesn't have to, because the cores that it does have a are whoppingly faster than on the consoles."

Even more so on this generation. There's simply no point, the workloads that any "normal" consumer uses do not use anything like this many (logical or physical) cores simultaneously, and the niche corner case uber-gamer isn't going to dictate an entire product stack for a fortune 200 organization like Intel.

 

[mczak]

It really isn't. I've been testing the A4-5000 (4 x 1.5GHz Jaguar cores) and it's somewhere between a Baytrail Z3740 and Z3770 in most CPU performance tests. Okay XBox One bumps that up the 1.75GHz and adds a couple more cores, but you're still talking less performance than say a Core i3 or i5 ULV chips, let alone any quad-core parts.

Haswell ULV i3 chips can't beat the 6 Kabini cores in XBone (best you can get is 1.7Ghz, no turbo), if you just consider perfectly scaling multithreaded code. The i5 and i7 ones (no big difference there) might be real close though. And obviously quad-cores (no ULV parts there) and higher TDP dual-core (say desktop i3) can do it rather easily.

 

[zed]

This is backwards. Apps don't use 4 cores today, why would they use 8 tomorrow?

The single most used app by far uses all cores (internet browser)
eg chrome uses all 4 cores of my i5 haswell

 

[Thorburn]

Haswell ULV i3 chips can't beat the 6 Kabini cores in XBone (best you can get is 1.7Ghz, no turbo), if you just consider perfectly scaling multithreaded code. The i5 and i7 ones (no big difference there) might be real close though. And obviously quad-cores (no ULV parts there) and higher TDP dual-core (say desktop i3) can do it rather easily.

Errrm, based on what?

 

[mczak]

Errrm, based on what?

You can find the benchmarks yourselves (though not much has been published, but at least some A5-5000 numbers, just multiply them by 1.8 to account for 6 cores and higher clocks).
For a dual-core haswell to beat that 6 core Kabini it would need to have 3 times the IPC if they run at the same clock.

 

[Thorburn]

You can find the benchmarks yourselves (though not much has been published, but at least some A5-5000 numbers, just multiply them by 1.8 to account for 6 cores and higher clocks).

From the testing I've done you'd need a 2x increase in performance, if not more, to match it and as you say that assumes PERFECT core and clock speed scaling.

For a dual-core haswell to beat that 6 core Kabini it would need to have 3 times the IPC if they run at the same clock.

Doesn't sound too far out?

 

[pjbliverpool]

You can find the benchmarks yourselves (though not much has been published, but at least some A5-5000 numbers, just multiply them by 1.8 to account for 6 cores and higher clocks).
For a dual-core haswell to beat that 6 core Kabini it would need to have 3 times the IPC if they run at the same clock.

1.75 to be exact.

 

[mczak]

Doesn't sound too far out?

Looks more like 2x, maybe 2.5 IPC to me.
Though i3 haswell ulv reviews are extremely hard to find (just like Kabini ones).
The i3-4010u (1.7Ghz) does ~1.9 in Cinebench (here: http://www.notebookcheck.net/Intel-Core-i3-4010U-Notebook-Processor.93567.0.html), whereas the A4-5000 does around 1.5. So actually in this benchmark haswell has 2.4 times the IPC.
That pattern seems to hold quite true overall (http://www.notebookcheck.net/AMD-A-Series-A4-5000-Notebook-Processor.92867.0.html has the same benchmarks for both chips), the A4-5000 nearly catches that i3 in 3dmark06 cpu and in sisoft sandra as well if you care about that.
From those limited set of benchmarks physics in vantage would be the one where the i3 fares best and even there it wouldn't be up to the task matching 6 kabini cores clocked roughly the same.
Now these may well be crappy benchmarks, if you can find some others which show the i3 doing better fine I want to see them. I think just from a pure theoretical aspect 3 times IPC in general doesn't make sense however. Haswell is mostly two times Kabini (e.g. 4 instructions per clock instead of 2) in raw throughput. Yes it has way more OoO resources thrown at the problem, but this is more or less just needed to actually be able to exploit its theoretical twice higher IPC (if you'd use a chip with disabled HT for comparison purposes I think you'd have trouble seeing it reach twice IPC anywhere actually).
A4-5000 isn't THAT crappy really in raw numbers. I think reviews focusing 99% on the higher end cpus might lead to some false impressions, cause in that price range those "1.5Ghz" intel chips run at nearly twice that clock most of the time (at least in workloads not using the gpu too - fwiw that's true for Silvermont too).

 

[pjbliverpool]

You have to remember that the bigger i7's have a lot more cache than the ULV's per core as well so performance should scale beyond pure clock and core increases.