Intel Broadwell-E Core i7 chip to feature 10 cores, 20 threads

[AGTDenton]

6 cores will already give you double the performance of a Westmere

Not quite sure what you meant, but I've altered my original post as I have Gulftown which is a Westmere derivative with a max of 6 cores, as opposed to 10 with the Xeons (Westmere-EX).

[crossy]

Compilers need megabytes per task, and that isn't much these days when 32GB of ram isn't crazy (even my home PC has 16GB). So, with enough ram to cache all your source code I find performance doesn't drop until I compile around 120 copies of gcc running at the same time so I think my single hard drive can cope with about a magnitude more threads than the Xeon at work can deliver. [snipped]

120 compiles at the same time?
Think you're probably right about the caching setup - I'd assumed that you'd be hammering the disks which - unless you've got some fancy SAN or RAID'd SSD setup - could have been a problem. Always surprises me how "clever" the file caching algorithms are these days - AIX especially.

Perhaps with the VMs you are right. I have seen people running big Java based systems on VMs which have a stupid number of threads going none of which seem to do much apart from consume ram and CPU. That probably isn't too common though.

Reminds me of a conversation I had with a (harried) Websphere support guy many moons ago over some abysmal JVM performance.
Websphere guy: "yep, we'll just increase realmem by 50% and see what happens"
Me: "okay and what then if that doesn't work"
Websphere guy: "In that case increase JVM heap"
Me: "and if that doesn't work then bitch to the app vendor?"
Websphere guy: "nahh, they'll just tell us to increase realmem by 50% again, then increase heap if that doesn't work."
Me:

It's conversations like this one about your dev environment that bring it home how much compute power we have "on tap" these days. Especially the "it's no big deal" when someone casually talks about hosting a VM setup with half a dozen mixed *nix and Windows VM's on what essentially is a "home PC" rather than some monolithic datacentre-resident "server" unit.

Still wondering though whether to pick a 5820K (and save some money) or go 5830K for a bit more oomph in PCI-E lanes. Although in my case SLi/Crossfire wouldn't be the use - I'd want to use that extra capability for PCI-E based SSD. Plus that bit-tech review (and a couple of others) that suggest that the '30K can overclock higher than it's cheaper brother - if your cooling can handle the extra load.

[DanceswithUnix]

120 compiles at the same time?

lol.

make -j120

You need a *lot* of code to compile for that to work ofc. Linking tens of thousands of object files together at the end is what chugs

[ypsylon]

I'm fairly confident that # of lanes will remain at current level of 40. X99 boards don't offer much in terms of expandability anyway. There are no XL-ATX/HPTX boards with 8 slots or multiple SFF-8639 ports which is a bit puzzling considering that it is top of the line platform for pretty much everything. From browsing internet to nuclear decay computing/modeling.

[abaxas]

120 compiles at once will just mean loads of cache and memory flapping. Much better to have parallelism in actual cpu and memory systems than it is just to thrown it all into the mix and let it thrash.

[UKnoble]

Was only 2 years ago that I bought a fairly decent (at the time) PC and now I already want a replacement but tech advances so fast nowadays it's hard to do it.

[D-T]

Still waiting on the E5 v4 SKUs.

[KadisEtrama]

Will they run Crysis?

crysis 4 for sure

[DanceswithUnix]

120 compiles at once will just mean loads of cache and memory flapping. Much better to have parallelism in actual cpu and memory systems than it is just to thrown it all into the mix and let it thrash.

It really doesn't matter though, I measured it. Over 120 yeah the performance starts to tail off a little, otherwise I need at least -j20 to keep this 12 thread box busy and beyond that you are making sure that it can always find one job that is in memory that it can run and not blocking on I/O. Beyond 120, yes I think it must be hurting CPU cache or possibly the working set is so big it can't keep all the source code fs-cached in system ram.

So really you can just throw a ton of work at the CPU and let Linux sort it out, there isn't a noticeable drop off from too much work whereas not having enough work to keep the cores busy hurts throughput a lot so I err heavily on the side of giving it more than the cores can handle.

If I compile over NFS, then the extra filesystem latency means bigger job numbers are needed.

[w1ntergr33n]

CPU tech has slowed down a lot since AMD went non-competitive (at least at the high end) so hopefully this will move things on a bit...

[Arulmani]

10 cores ??!!!

[qasdfdsaq]

20 threads not much on a home VM farm - what the heck are they doing? I can run a Windows VM and three Linux VM's quite happily on a rinky-dink C2D and that's what ... 4 threads? Similarly in a non-shared setup wouldn't you hit storage/memory limitations on compiles before you used up your 20 threads? Not posing this as criticism merely asking the question, especially as I'm planning to up my sights from running a mere 3VM's simultaneously.

Not sure if sarcasm or trolling but the average home user has zero VMs. The high-end power users I know with home VM "farms" have Xeon workstation/server platforms for them, which cost less than the overpriced overclockable gaming platform you are talking about. If you want an overclockable gaming PC to run "sites" and a "VM farm" you are looking in the wrong place.

So the US$64,000,000* question ... is it worth holding off on a Haswell-E purchase on the basis of this "rumour"? Or, is it more sensible to get a "cheap" (stop-gap?) 5820K while waiting for that more capable Broadwell-E drop-in replacement? Speaking of drop-in, kudos to Intel (if the rumour is true) for allowing an upgrade path for X99 owners, rather than inventing yet another socket. Asking the purchasing question because I'm thinking of splurging on a 5930K for a VM/media-coding setup this Christmas.

Do you really think you'll benefit from 10 cores over 8? If so, then hold off. Otherwise, the performance improvement is miniscule. Frankly, I have corporate VM farms that have trouble loading up 8 cores simultaneously let alone 10.

Not quite sure what you meant, but I've altered my original post as I have Gulftown which is a Westmere derivative with a max of 6 cores, as opposed to 10 with the Xeons (Westmere-EX).

I meant 6-cores of Skylake at ~4Ghz would give double the performance of 6-cores of Gulftown at ~3.2 Ghz. You wouldn't need 12 to get a doubling of performance (although, admittedly, you'd need 12 to get a doubling of cores)

I'm fairly confident that # of lanes will remain at current level of 40. X99 boards don't offer much in terms of expandability anyway. There are no XL-ATX/HPTX boards with 8 slots or multiple SFF-8639 ports which is a bit puzzling considering that it is top of the line platform for pretty much everything. From browsing internet to nuclear decay computing/modeling.

That's because X99 board are gaming boards and neither of those features are relevant to gamers. That's about as useful as saying there are no Xeon E7 boards with Iris graphics.

There are plenty of Haswell boards with 7+ slots or multiple SFF-8639 ports but they don't use X99.

Still waiting on the E5 v4 SKUs.

They'll be out around the same time. Seeing as they're the same platform.

[crossy]

Not sure if sarcasm or trolling but the average home user has zero VMs.

Erm, neither sarcasm nor trolling - if you'd bothered to actually READ the post, rather than knee-jerking then you'd have seen I said "Not posing this as criticism". Strangely enough Danceswithunix (to whom the comment was addressed) knew exactly what I was getting at.

The high-end power users I know with home VM "farms" have Xeon workstation/server platforms for them, which cost less than the overpriced overclockable gaming platform you are talking about. If you want an overclockable gaming PC to run "sites" and a "VM farm" you are looking in the wrong place.

Okay, so disregarding the flannel in your comment above what you're basically saying is "don't bother with i7 for VM, Xeon is better". Which if that was your intention then I'd obviously agree. Unfortunately, all the b'marks I've seen show that Xeon's lower clock (lower IPC too??) make them not a good choice for gaming and some media work. Downside of Xeon is that the overall costs are higher than something based on the "mainstream" processors. Ideal setup would be to have a dedicated (Xeon-based) headless server (running a hypervisor) and a proper gaming setup. Unfortunately that's a lot of money and space - neither of which I have at the moment.

PS, merely saying "you're wrong" but not either saying why or - better still - suggesting a "better" alternative isn't helpful. That's one of the things I like about Hexus - if you goof at least some friendly voice will usually point out your mistake and try to steer you right.

Do you really think you'll benefit from 10 cores over 8? If so, then hold off. Otherwise, the performance improvement is miniscule. Frankly, I have corporate VM farms that have trouble loading up 8 cores simultaneously let alone 10.

I was assuming that B-E chips will also come with some architectural improvements too. Personally speaking hexacore is fine for me - and I've seen times with my home system when all six of those cores are being hammered, (80%+ utilisation). Oh, and sizing in a corporate environment is an "art" rather than a science - and it gets worse with the increasing number of cores, e.g. it's "fun" to have to size when you're talking about a potential of 128-256 processor cores.

[qasdfdsaq]

what you're basically saying is "don't bother with i7 for VM, Xeon is better". Which if that was your intention then I'd obviously agree. Unfortunately, all the b'marks I've seen show that Xeon's lower clock (lower IPC too??) make them not a good choice for gaming and some media work. Downside of Xeon is that the overall costs are higher than something based on the "mainstream" processors. Ideal setup would be to have a dedicated (Xeon-based) headless server (running a hypervisor) and a proper gaming setup. Unfortunately that's a lot of money and space - neither of which I have at the moment.

Xeons have the exact same IPC and are cheaper. The only thing you're paying for when buying an "Extreme Edition" i7 is the "Extreme" label. They fit in the exact same mainboards, the exact same sockets, and support the exact same peripherals as the HEDT platform, so the "overall costs" outside the processor are exactly the same. They're the same chip built of the same fab with the exact same architecture, just with less features disabled and programmed without an "i7 Extreme" string in their model number. The only reason you'd not buy a HEDT i7 is if you're a hobbyist overclocker or just want the label.

I was assuming that B-E chips will also come with some architectural improvements too. Personally speaking hexacore is fine for me - and I've seen times with my home system when all six of those cores are being hammered, (80%+ utilisation). Oh, and sizing in a corporate environment is an "art" rather than a science - and it gets worse with the increasing number of cores, e.g. it's "fun" to have to size when you're talking about a potential of 128-256 processor cores.

The core architecture has always been identical. The peripheral support systems (i.e. "uncore") often have improvements, necessitated by the need for coherent communication between multiple cores and/or sockets, but the core logic itself is not going to have any architectural improvements. The only time I've seen an "improvement" in the core itself was in the case of Haswell's broken TSX implementation in the earlier non-E editions but that's not so much an improvement as a bugfix also found in later steppings of the non-E.