Intel Launches Xeon E7-8800 and E7-4800 V3 Processor Families 46
MojoKid writes: Intel is taking the wraps off of its latest processors for enterprise server and pro workstation applications today, dubbed the Xeon E7-8800 / 4800 v3. Like its high-end desktop processors, the Xeon E7-8800 / 4800 v3 product families are based on the Haswell-EX CPU core. These new Xeons, however, offer a plethora of other enhancements and are packing significantly more cores than any current desktop processor. The highest-end Xeon E7-8800 series processors, for example, are 18 core chips. Previous generation Xeon E7 v2 processors were based on the Ivy Bridge-EX core, while the new E7 v3 parts are based on Haswell-EX, though both are manufactured on Intel's 22nm process node. Next generation Broadwell-EX based Xeons will make the move to 14nm. Xeon E7-8800 / 4800 v3 series processors have 32-lanes of PCIe 3.0 connectivity per socket, TSX is enabled in all SKUs, they offer support for both DDR3 and DDR4 memory (though, not simultaneously), and can address up to 6TB of memory in a 4-socket configuration or 12TB in an 8-socket setup. Intel has also goosed the chip's QPI interface speeds to 9.6GT/s.
Good for EM simulation (Score:1)
Re:Good for EM simulation (Score:5, Informative)
Looks good for my EM simulation needs. Too bad the licensing to take advantage of all those cores is very expensive.
Intel has 2 parts just for you: E7-8891v3 and E7-8893v3, maximum clock frequency, fewer cores. Seriously, they are designed and marketed specifically for "Lower per-core software license fee costs Higher per-core performance".
Re: (Score:2, Insightful)
Don't worry. I'm sure Oracle already has a plan to screw you with these processors too.
I hear their next licensing program is called the 'Vinny special'. In this scheme you turn over your ledger, and in return Larry's goons will tell you how much of your gross you need to fork over to stay "in compliance" and how much more you need to pay "keep your kneecaps".
Re: (Score:3)
Microsoft and Oracle already know how to deal with this: it's called "nominal cores". The vendor comes up with a list of CPU models that are assigned "Weighting scores"; cores of certain CPU models are worth 150% of a core each, for licensing purposes, certain cores are worth 75% of a core each for licensing purposes, etc.
Re: (Score:1)
pay-per-cpu-cycle
Re: (Score:2)
Intel has 2 parts just for you: E7-8891v3 and E7-8893v3, maximum clock frequency, fewer cores.
This is a good idea for single-threaded tasks which can't be sped up anymore by distributing that task over multiple cores....... maximum single-core speed is ideal for certain workloads, and the extra cores would just be wasted.
Re: (Score:2)
Might be able to maintain 5 FPS in Dwarf Fortress.
Re: (Score:2)
This is a good idea for single-threaded tasks which can't be sped up anymore by distributing that task over multiple cores....... maximum single-core speed is ideal for certain workloads, and the extra cores would just be wasted.
To be clear, those two are 4-core and 10-core parts...
Re: (Score:2)
Oh, I don't know, servers? Virtualisation hosts?
Why would anyone use a Xeon with that many cores in a desktop? o_O
Re:Ah, 18 cores (Score:4, Interesting)
Why would anyone use a Xeon with that many cores in a desktop? o_O
For my compiler, you insensitive clod. Now that we have SSDs that can feed the beast fast enough to keep it busy, I would be delighted to have one of these on my desk.
Re: (Score:2)
Exactly. Not that it would likely be worth it for my employer, even with salary accounted for, but my work project has 25 subprojects that need to be built but a max depth of 5 I think in the dependency graph. So could reduce my full builds from ~1min 5-10 times a day to say 10s 5-10 times a day. But usually I can multitask: do another once over the pending changes, etc. So the time isn't usually wasted anyways.
Re: (Score:1)
Oh it must be so traumatic... back in my day we use parallelism to get our build time down to 30 minutes. Except on the Cray T3D where, ironically, it took 24 hours.
Re: (Score:3)
No it isn't tramatic just a waste of time. Especially when a couple people come over to your desk and want you to check something on another branch. Checkout, deploy database. Then build. Then when switching back need to build again. So 3 people sitting around waiting for a progress bar.
I've lived the HPC game too. In those days for me at least it was 5% dev and 95% reading/writing journal articles, books, either triggering or automating compute job configurations etc. Very little code but running for 200k+
Re: (Score:2)
But, that is what shared build machines are for. Not everyone is going to be building at once, and with a 5-20 member team its a lot less expensive to buy a machine with 40 cores and a bunch of midrange desktops/laptops, than buy all the developers high end workstations.
Re: (Score:2)
We have "build" machines too. They get hijacked to run CI workloads. Typically 8-16 jobs running concurrently 24-7. They build -> then trigger smoke, other, security, migration etc testing.
Often we compile several times locally in the process of getting things working. Sadly highly coupled code between C# and tsql/db code. You need to migrate and build/run tests locally to have any hope of pushing something to a CI/build server that has a hope in hell of not breaking everything.
Anyways, I like having a h
I've played with a 144-thread x86 beast. (Score:2)
It's *beautiful* for simulating a cluster on a single machine.
Re: (Score:3)
Why would anyone use a Xeon with that many cores in a desktop?
I can think of quite a few specialised but realistic applications: CAD/CAM/CAE, rendering/pre-viz, high-end audio or video mixing work, simulation, and running modern web apps as fast as their traditional desktop equivalents used to run on a Pentium II.
Re:Most CAD Modeling still uses one core. (Score:5, Informative)
Get the i7-4790k which runs up to 4.4ghz and is unlocked so you can overclock it to 4.8ghz or so.
It's probably hundreds of times faster than the shitty pentium 4 you mentioned to boot......
Re: (Score:2)
How fast are they with 1/4 to 1/8th of the memory bandwidth?
Re: (Score:2)
Pre-Haswell Core i7 dual core laptop chips run roughly 4 times as fast as a 3.8GHz P4 single core -- per thread. And that is without overclocking. Haswell cores are even faster, and the desktop chips run at much higher clock rates than the laptop chips, so I've no doubt you could see 5-6 times the performance on a current generation CPU.
Re:Mainframe era? (Score:4, Interesting)
Uh, it's nearly as much CPU power (141 cores at 5.2GHz, but even more CISC that x86) as the current mainframe, zSeries hasn't been about brute CPU in decades, it's about balanced CPU and I/O combined with high QoS and absolute stability. As an example the Z13 has nearly 1GB of L4 cache in the I/O coprocessors.
Re: (Score:2)
It's breathtaking to use systems that are explicitly designed for multiple workloads. I wonder how much energy and space datacenters could save if everyone and their brother hadn't gone whole-hog on an architecture that is only marginally suitable for multiprocessing. It keeps getting better, but x86-64 is just not designed for this.
Re: (Score:2)
Discoveries are made very irregularly, and are hard to plan. Therefore, its better for a company in the long run to "sandbag".
Re: (Score:2)
Sun back in the day, not sure if they still do it, used to offer servers with say half of the cores disabled. You could buy licenses and they'd turn on the extra cores without even needing a reboot. They offer different price points even if it means they waste some good cores disabling them to make an artificial performance different its still better for them than to have to make a different design/fab for each step in the process. The lifetime of a chip architecture/manufacturing process is so short that b
Snap. (Score:1)
still on Haswell (Score:2)
So they're not even based on Broadwell yet; when are they expecting to get Xeons based on Skylake out - next year?
Re: (Score:2)
It's been like this since the P4/Foster Xeon:
Desktop->workstation/1P server->DP server->MP server.
This is the final spin of the Haswell line, Broadwell-EP Xeon's are set to launch at the end of the summer.
Seems to still be architecturaly-gimped (Score:4, Interesting)
It appears that they didn't do much to the QPI besides boost the speed a bit. That's not going to fare well in HPC stuff. The reason I didn't use the V2 E7-8*** line was because due to how gimped the memory architecture was, you could run 2 socket 4 GPU, 4 socket 2 GPU, but not 4/4.
It was cheaper, and just as effective, to go with the E5 instead, and make multiple node systems into a single box, instead. 8 socket, 12 GPU. Fuck yea.
Re:Seems to still be architecturaly-gimped (Score:4, Interesting)
Quick clarification: Not the memory controller was gimped, but how processors communicated and shared stuff out of their memory to other processors was gimped. And the E7 v3 looks to have the same limitation. Pumping up QPI speed might help alleviate that SOME but nowhere near what's needed for multiple socket multiple GPU configs in a single non-nodal system.