Intel Broadwell for desktops

[Alexko]

There's always LGA 2011. The bottom-line for Intel is unfortunately simple, however: if people have less than $300 to spend Intel doesn't need to bother, because they'll get an i7 anyway. If they have more to spend Intel doesn't need to bother either, because they'll get an LGA 2011 Core i7 or a Xeon.

That said, come 2015 AMD could release a 20nm CPU with 8~16 Excavator cores, which would shake things up a bit. Unfortunately I'm far from convinced that this is going to happen.

 

[Rodéric]

I want 6/8 cores because I don't want a GPU I won't use embedded in the CPU.

That said, if the OS let apps use both GPU and all CPU then it's quite interesting as load can be spread amongst more units and you wouldn't have to care too much about increasing display latency...

 

[eastmen]

There's always LGA 2011. The bottom-line for Intel is unfortunately simple, however: if people have less than $300 to spend Intel doesn't need to bother, because they'll get an i7 anyway. If they have more to spend Intel doesn't need to bother either, because they'll get an LGA 2011 Core i7 or a Xeon.

That said, come 2015 AMD could release a 20nm CPU with 8~16 Excavator cores, which would shake things up a bit. Unfortunately I'm far from convinced that this is going to happen.

I don't see how amd can ever compete again on the cpu side. in 2015 intel will be releasing their second 13nm cpu.

 

[fellix]

Well, if Mantle catches on, at least some games won't need the power of a single-threaded guzzler.

 

[Alexko]

I don't see how amd can ever compete again on the cpu side. in 2015 intel will be releasing their second 13nm cpu.

If the main competition is a 4-core Skylake CPU, it's not impossible. And 14nm vs. 20nm is just one node ahead—not great for AMD, but nothing unusual.

With, say, 6 Excavator modules (12 cores) it should be an easy victory for AMD in highly threaded workloads. The real question is whether there will be a market for that.

 

[pjbliverpool]

Yeah I agree. Intel have been far too passive on the performance front in recent years. The IPC difference between Sandybridge and Haswell is ridiculously small, maybe 20% at best. It's amazing AMD still hasn't managed to catch up given this opportunity but Steamroller certainly looks promising and if Intel contnues it's pathetic IPC increased through Broadwell and Skylake and AMD makes Excavator a decent upgrade from Steamroller then a 6 or 8 module version should have no problem handling a quad core Skylake in threaded applications. Intel could easily respond with an 8 core Skylake of course - but that's what we want!

 

[I.S.T.]

Well, Broadwell is basically just a die shrink+better graphics anyway. It's not really gonna improve on the CPU side unless Intel releases like 4.3 GHZ versions.

 

[Alexko]

I wouldn't call Intel's IPC improvements pathetic, in fact I think they're quite impressive, it's just that Sandy Bridge was pretty damn good to begin with. The law of diminishing returns is hard to fight, and at this point every bit of IPC improvement that doesn't come at the expense of power efficiency is quite a feat.

On the other hand, Bulldozer sucks, and Piledriver is just a bit less crappy, so AMD should have an easier time making substantial improvements for at least a couple of generations.

 

[Wynix]

I wouldn't call Intel's IPC improvements pathetic, in fact I think they're quite impressive

You have to be kidding, IPC improvements have been small; The only good progress has been with power consumption, something which is not that important to me.

As i suggested in another thread, give us a couple more cores, we deserve them for getting such minuscule IPC improvements for the last 2 revisions.

 

[Laurent06]

You have to be kidding, IPC improvements have been small; The only good progress has been with power consumption, something which is not that important to me.

As i suggested in another thread, give us a couple more cores, we deserve them for getting such minuscule IPC improvements for the last 2 revisions.

What Intel has achieved is impressive no matter what you think and it's very likely they are getting close to the point where they'll have to start from scratch to keep on getting higher IPC.

 

[Alexko]

You have to be kidding, IPC improvements have been small; The only good progress has been with power consumption, something which is not that important to me.

As i suggested in another thread, give us a couple more cores, we deserve them for getting such minuscule IPC improvements for the last 2 revisions.

I'm not saying the improvements have been large (though Haswell did get ~10%), I'm saying they're impressive considering the starting point. Improving IPC is harder when it's already high. And if improving IPC in itself is not necessarily very hard even in this case, improving it without increasing power consumption is very hard, and necessary.

If you just enlarge a few buffers, OoO windows, caches, etc., you may get +5% IPC, but if that comes with +10% power, you have to reduce clock speeds by ~10% in order to stay within your TDP envelope, and actual performance goes down. So yeah, Intel's improvements are impressive. It's not an accident that no design can touch Haswell in IPC.

 

[Thorburn]

If you just enlarge a few buffers, OoO windows, caches, etc., you may get +5% IPC, but if that comes with +10% power, you have to reduce clock speeds by ~10% in order to stay within your TDP envelope, and actual performance goes down. So yeah, Intel's improvements are impressive. It's not an accident that no design can touch Haswell in IPC.

IIRC the design criteria is for every 2% you increase performance you can only raise power consumption by 1%.

 

[homerdog]

IIRC the design criteria is for every 2% you increase performance you can only raise power consumption by 1%.

That's been the case since the original Core2 days I think. They've definitely stuck with it since then.

Of course on the desktop we have TDP to spare so I wouldn't mind a 1 to 1 increase in perf/power up to ~100W. Even under full load my i5-3550 just sips power considering the IGP is idle.

 

[Thorburn]

That's been the case since the original Core2 days I think. They've definitely stuck with it since then.

Of course on the desktop we have TDP to spare so I wouldn't mind a 1 to 1 increase in perf/power up to ~100W. Even under full load my i5-3550 just sips power considering the IGP is idle.

Problem is you are sharing the same microarchitecture with low power parts where you can't afford to push the TDP up. A 1 to 1 perf/power increase leaves them no faster than before.

 

[zed]

You have to be kidding, IPC improvements have been small;

yep they have been terrible, I wonder due to not having competition if theyre deliberately holding back

 

[DavidC]

I may be wrong, but I had the impression 2MB of the L3 was reserved in GT3e SKUs to manage the 128MB eDRAM L4, rather than being artificially fused off.

Edit: I am probably wrong. According to http://extremespec.net/intel-announced-8-types-crystals-haswell/ Haswell dies including GT3 graphics only physically have 6MB of L3.

No, but GT3e SKUs have the tag for the eDRAM, which takes a fair bit of space. One might say 2MB L3 equivalent.

 

[homerdog]

yep they have been terrible, I wonder due to not having competition if theyre deliberately holding back

They are definitely holding back at the high end, where the platform and CPU architecture is running a generation behind and prices are through the roof. They are not even coy about this. If AMD were in the picture I'm sure this wouldn't be the case.

For the smaller sockets they want higher IPC to enable higher performance at lower clockspeeds and power envelopes. In those markets there is competition a plenty (or at least there will be in the very near future).

As mentioned already in the thread, with each generation it gets harder and harder to sensibly improve IPC.

 

[Silent_Buddha]

They are definitely holding back at the high end, where the platform and CPU architecture is running a generation behind and prices are through the roof. They are not even coy about this. If AMD were in the picture I'm sure this wouldn't be the case.

For the smaller sockets they want higher IPC to enable higher performance at lower clockspeeds and power envelopes. In those markets there is competition a plenty (or at least there will be in the very near future).

As mentioned already in the thread, with each generation it gets harder and harder to sensibly improve IPC.

A 1:1 increase doesn't even make financial sense in the high end. The only consumer that might be interested in such a thing is enthusiast gamers. HPC/Server market wants high Perf/watt. If power increases inline with performance increase then why use any new processor from Intel for those markets? For general home consumers they likely don't need that power increase anyway. For businesses, we again get into a situation where perf/watt is highly desireable.

The market for high performance combined with high power consumption is shrinking every year. There's no reason to invest in it.

Regards,
SB

 

[Andrew Lauritzen]

They are definitely holding back at the high end, where the platform and CPU architecture is running a generation behind and prices are through the roof.

i.e. IVB-E/Xeon? There are some legitimate reasons why that stuff takes longer to design, but I agree in principal that there's less pressure to do it quickly with the lack of competition. That said, being a generation "behind" there in terms of micro-architecture isn't the end of the world. I'm still happily using a SNB-E and I doubt I'll feel too compelled to upgrade until HSW-E at least (mmm AVX2/FMA ).

As mentioned already in the thread, with each generation it gets harder and harder to sensibly improve IPC.

Yeah I sort of wonder how much ILP people really think is still on the table in the applications that they figure should get faster... I imagine few people have really measured or thought about it much though.

As far as enthusiast gamers go, if you don't care about TDP, have adequate cooling and so forth and want that extra last boost in performance regardless of the efficiency, that's what overclocking is for. It's pretty well supported and risk-free these days, so I'm not really sure why anyone needs Intel to do it for them (and presumably pay a premium for an even higher clocked SKU).

Well, if Mantle catches on, at least some games won't need the power of a single-threaded guzzler.

Well there's still, you know, the rest of the game to run Graphics stuff isn't even the worst single-threaded offenders in most games... it's amazing how poorly all of the gameplay and AI stuff typically runs.

 

[homerdog]

A 1:1 increase doesn't even make financial sense in the high end. The only consumer that might be interested in such a thing is enthusiast gamers. HPC/Server market wants high Perf/watt. If power increases inline with performance increase then why use any new processor from Intel for those markets? For general home consumers they likely don't need that power increase anyway. For businesses, we again get into a situation where perf/watt is highly desireable.

The market for high performance combined with high power consumption is shrinking every year. There's no reason to invest in it.

Regards,
SB

No doubt. I understand and agree 100%.

i.e. IVB-E/Xeon? There are some legitimate reasons why that stuff takes longer to design, but I agree in principal that there's less pressure to do it quickly with the lack of competition. That said, being a generation "behind" there in terms of micro-architecture isn't the end of the world. I'm still happily using a SNB-E and I doubt I'll feel too compelled to upgrade until HSW-E at least (mmm AVX2/FMA ).

Right, Intel has little need to stay on the cutting edge for the LGA2011 platform because even the SNB-E stuff is way ahead of the competition

As far as enthusiast gamers go, if you don't care about TDP, have adequate cooling and so forth and want that extra last boost in performance regardless of the efficiency, that's what overclocking is for. It's pretty well supported and risk-free these days, so I'm not really sure why anyone needs Intel to do it for them (and presumably pay a premium for an even higher clocked SKU).

I dislike the K series because they actually lack some of the functionality of the non-K chips but are more expensive. Haswell K-series actually disables TSX that was a really dumb and pointless move IMO.