Intel Launches Xeon E5 V3 Series Server CPUs With Up To 18 Cores

MojoKid writes Intel took the wraps off its Xeon E5 v3 server line-up today and the chip, based on Intel's Haswell-EP architecture, is looking impressive. Intel's previous generation Xeon E5 V2 chips, which were based on Ivy Bridge, topped out at 12 cores per socket. The new Xeon E5 v3 processors, in contrast, are going to push as high as 18 cores per socket — a 50% improvement. The TDP range is pushing slightly outwards in both directions; the E5 V2 family ranged from 50W to 150W, whereas the E5 V3 family will span 55W — 160W in a single workstation configuration. The core technologies Intel is introducing to the E5 V3 family pull from the Haswell architecture, including increased cache bandwidth, improved overall IPC, and new features like AVX2, which offers a theoretical near-doubling of floating point performance over the original AVX instructions. Full support for DDR4 DRAM memory is now included as well.

Comments

Why not 16?

[beltsbear]

I would have thought a 16 core config would be an efficient number.

Re:Why not 16?

[GoJays]

You’re on 16, all the way up, all the way up...Where can you go from there? Nowhere. What we do, is if we need that extra push over the cliff...Eighteen. Two more cores."

DiBergi: "Why don’t you just make 16 faster and make 16 be the top number, and make each core faster?"

Nigel (after taking a moment to let this sink in): "These ones go to 18 cores."

Re:Why not 16?

[ArcadeMan]

At first it was a 32-cores CPU but they had to scale it down because of budget cuts.

alternate reply: they calculated the number of cores on an old Pentium.

Re:Why not 16?

[gman003]

It's a weird die layout. They have four "columns" of cores - three columns of 4 cores, and one of 6. The memory interface and some other stuff takes up the room that would have been used for the other two cores on the first three columns. I guess they didn't need to be longer, so they used the extra room for two more cores.

compiling

[Anonymous Coward]

As a programmer I so want one of these.

$ make -j 18 # FTW!

Re:compiling

[Anonymous Coward]

Well, if you write a single non-recursive Makefile like you're supposed to[1], then GNU Make might be able to build thousands of targets in parallel, in which case it'd quickly grind to a halt under I/O pressure.

An easy rule of thumb for parallel jobs with mixed workloads is N+1, whether it's Make or something else entirely, where N is the number of CPUs.

[1] Usually that means using the include directive to build up a single dependency graph, which in turn means defining your recipes using absolute paths. For example:

CFLAGS_$(d) = $(CFLAGS) -fsomething
CPPFLAGS_$(d) = $(CPPFLAGS) -DFOO_BAR
$(d)/foo: $(d)/foo.c $(d)/bar.c:
                $(CC) $(CFLAGS_$(@D)) $(CPPFLAGS_$(@D)) -o $@ $^

You get used to it after while. Especially when you edit a single source file in a project 10x the size of the Linux kernel, and it only takes 5 seconds to recompile with `make all` instead of 5 minutes.

See http://evbergen.home.xs4all.nl/nonrecursive-make.html and http://make.paulandlesley.org/autodep.html for the typical approachs to non-recursive GNU Make.

Re: compiling

[LostMyBeaver]

HyperThreading plus branch prediction says you should use -j72

Re:Oooh Shiny

[armanox]

Meh. Oracle just announced a 32-core Sparc proc, with variable thread count (1-8 per core, depending on load need - if it needs better single thread performance is shuts off the other threads). AMD's move next.

Re:Oooh Shiny

[ArcadeMan]

AMD's move next.

And then, Atmel. Anyone up for a quad-core ATtiny85?

Earth

[ArcadeMan]

You know what would be awesome? A 42-cores CPU. To keep your cup of tea really, really hot.

Re:Earth

[ClickOnThis]

That would be almost, but not quite, entirely unlike a retro-gen GPU.

And highly improbable.

Re:Earth

[Chas]

You know what would be awesome? A 42-cores CPU. To keep your cup of tea really, really hot.

Unfortunately cores this large, powerful and complex all stand excellent chances of spontaneously self-triggering SMEF (Spontaneous Massive Existence Failure).

Re:Earth

[ClickOnThis]

Unfortunately cores this large, powerful and complex all stand excellent chances of spontaneously self-triggering SMEF (Spontaneous Massive Existence Failure).

That acronym is a fine example of Onomatopoeia.

But is that really true? GPUs have far more than 42 cores, and they don't SMEF very often, do they?

Re:Earth

[Chas]

Ah. But that's where you're overthinking it.

How may multi-core GPUs have EXACTLY 42 (Forty Two) cores?

  =)

  =D

The plot thickens!

Re:F**k Everything! We're Doing Nineteen Cores!

[0xG]

An oldie, but one of the greats...http://www.theonion.com/articles/fuck-everything-were-doing-five-blades,11056/

What happened to the core-wars?

[MindPrison]

I remember 8-16 cores being announced YEARS ago, but they never ever appeared in regular desktop computers (well, not at your cheap online stores or mainstream street stores either).

As a hobbyist 3D modeller, the more cores the merrier (and more memory + cache of course). But I'm kind of disappointed about where we're headed. Announcements of new processor with an astonishing amount of cores appear all the time, but they never appear in the actual stores, are they too expensive or something?

I remember the good times when I built my own computers, just going to the local computer shop and purchasing the needed components, aahh...the times we could overclock our cheapo AMD and then Intel came back with a much more overclock-able CPU etc. I've been stuck with my 4 core cpu for the last 6-7 years now and the only thing that has improved my rendering is the NVIDIA GPUs...but they do have their limitations, you can't render everything at improved speeds with these, some things just have to be done with a regular CPU.

What's up with that?

Re:What happened to the core-wars?

[hypergreatthing]

i think you can get up to 8 cores in a standard desktop.
These are server cpus. You can put a server cpu in a desktop pc, but these xeon cpus can be found in single, dual or quad variants. The price for one of these 18 core cpus? I figure it should be around 2800$. The motherboards for these will be around 500$. It's not something you'll normally see on someone's desk.

Re:What happened to the core-wars?

[Fwipp]

Most people don't even make use of 4 cores on their desktop, so it doesn't make sense for them to push 16-core consumer chips. If you want to do server-like highly-parallel tasks, maybe you should buy a server CPU.

Re:What happened to the core-wars?

[MindPrison]

i think you can get up to 8 cores in a standard desktop. These are server cpus. You can put a server cpu in a desktop pc, but these xeon cpus can be found in single, dual or quad variants. The price for one of these 18 core cpus? I figure it should be around 2800$. The motherboards for these will be around 500$. It's not something you'll normally see on someone's desk.

You know, kids buy 500-1000$ graphics cards, just for gaming. So to me, this isn't that far fetched. But of course you're right if we're talking about the Hendersons (Regular office Joe and Jenny Surf)...it would make NO sense for them to purchase an Alienware monster or similar.

Re:What happened to the core-wars?

[alen]

that went to mobile where the fandroids cream their shorts over any mention of cores and megahertz

Re:What happened to the core-wars?

[TechyImmigrant]

It's a server chip. You'll need a server motherboard, but you can certainly buy one and use it for your computing needs.

Re:What happened to the core-wars?

[hypergreatthing]

Yeah but these server chips aren't for gaming. Typically to put in that many cores and keep that amount of TDP, you're going to make sacrifices. Specifically in top speed. You may have 18 cores, but they're definitely not running at 4ghz like the i7-4790K. Try more like 2.7ghz. For gaming, single threaded performance is still king.
Also for gaming, it's fairly rare today to be cpu locked (cpu being the bottleneck) and instead being the graphics card, memory bus, etc.
These cpus are first and foremost used for virtualization/servers, then multi-core/threaded apps like 3d rendering, video editing, etc. The focus on server components is on stability, not tweaking performance.

Re:What happened to the core-wars?

[NatasRevol]

Exactly. These are for VMware farms.

Instead of 50 VM per core, now you can fit 75 VMs per core.

Which your QA's will stretch to the max anyway, so this is at least saving you money by not having to buy 50% more physical servers.

Re:What happened to the core-wars?

[Kjella]

I've been stuck with my 4 core cpu for the last 6-7 years now and the only thing that has improved my rendering is the NVIDIA GPUs

6-7 years ago, that's like a Q6600 or so? Have you actually looked at benchmarks like Q6600 vs 4790K because current top of the line quad-cores are 3-4 times faster than that.

I remember 8-16 cores being announced YEARS ago, but they never ever appeared in regular desktop computers

No, because of a couple things:
1) Single-threaded performance is still huge and often the bottleneck in interactive work - big multithreaded jobs just decide how long a coffee break you get.
2) Lots of cores means big die means big costs and poor yields meaning they aren't really interested in selling it at consumer prices.
3) Companies would no doubt try to use these as cheap servers or whatever and they don't want enterprise users buying anything but Xeon.
4) You can now get i7-5960x in an "enthusiast" system with 8 cores at least, though it'll cost you $1000. Or you can buy AMDs marketing and get an "8-core" FX processor...

Re:What happened to the core-wars?

[apraetor]

That should read "per processor", yes?

Re:What happened to the core-wars?

[armanox]

And how much more does it cost you in power to run that AMD? Plus add on cost for a more powerful PSU.

Re: What happened to the core-wars?

[LostMyBeaver]

I think the changes in v3 probably are able to make the claim of more VMs per for due to memory bandwidth. Of course, then one could say 50% more cores at 20% greater capacity per core (my guess) which would be 65% more VMs per physical server.

I suppose Cisco will release a quad socket, 6TB RAM blade soon. I hope they'll loan me one to try. I was able to run entire companies with VDI on 3 60-core blade with V2, a little slowly, but effectively. 3 72-core V3 blades should be better.

The real problem comes that whether you're using Cisco with 640Gb/sec Ethernet or some aweful Infiniband on HP/Dell solution, I don't see how to get the data in and out of the blade quickly enough to use the cores. :( I wonder if the systems will even have enough PCIe bandwidth to support it.

I am hoping that with these servers, Cisco will support quad-SLI nVidia at some point. VDI is much better when it's 3D accelerated and adding massive numbers of GPU cores is helpful. But if they do, I doubt there will be enough PCIe lanes to handle storage and video.

So... New bottlenecks?

so

[hypergreatthing]

These shiny new processor having working TSX instruction sets? The ones that are supposed to help with virtualization?

Re:so

[viperidaenz]

Isn't a TSX an American branded Japanese Honda Accord?

Re:so

[kirk444]

Isn't a TSX an American branded Japanese Honda Accord?

Not anymore, it was axed, along with the TL, and they were both "combined" to make the TLX.

Re:so

[alen]

we have accords here
the acuras are overpriced honda's for dummies who want "luxury"

Re:so

[viperidaenz]

I know you do, but American Accords are not the same as Japanese Accords.

It's how many they could pack in

[Sycraft-fu]

The way they design their CPUs it is easy to have pretty much any number that is divisible by 2. It isn't a big deal to have something that is any particular amount more or less. So then it comes down to power, thermal, and die size limits.

Apparently 18 cores is what they cap out at, this time around. I'm sure you'll be able to get 16 core, and less, chips, that is just the most they could stuff in there before exceeding whatever design limitations they'd set.

at least they have 4 and 8 core models as well

[alen]

SQL server went to per core pricing a few years back and are looking at around 8 cores per server now when we buy new hardware. more cores won't do much for us except send more money to microsoft

Re:at least they have 4 and 8 core models as well

[0xG]

Yes, this is a real concern - I am worried that 4 & 6 core servers will not be available much longer.
And I am pretty sure that in a db environment, 2 sockets with 6 cores each will handily outperform a single socket with 12 cores...

Re:at least they have 4 and 8 core models as well

[Anonymous Coward]

Postgres is calling you ... 0xG joooiiiinnn ussss ...

(in some sort of ghostly voice I guess)

Re:at least they have 4 and 8 core models as well

[Anonymous Coward]

I'm wondering how long before everyone demands that licensed software only make use of the hardware it's licensed for, which may only be part of the resources available to the server as a whole. For example, a 2-core license for SQL Server could run just fine on a 12-core server, but would only use 2 cores at a time. Make that happen, Microsoft, then we'll take your stupid per-core licensing seriously.

On a related side note, today I was greeted back to work on a Monday morning by a BizTalk re-licensing issue. Two years ago, we got the 1-processor license and built out an 8-core, single-processor box to run it on. Now they want $30k+ for the Software Assurance (read: rental software) update to make it an 8-core license because Microsoft is a bunch of greedy fucks. We don't need an upgrade, but SA-licensed software must be periodically re-licensed in order to stay legal if you continue to use it.

I'm considering sending BizTalk to the scrap-heap anyway. It's cheaper to just build a .Net CLI app to do what it does for us (which is file format translation for idiotic old mainframe formats... think ANSI X12, but don't think it too long or it'll start to burn).

Re:at least they have 4 and 8 core models as well

[mlts]

Oracle and Sybase as well have this type of licensing, unless something has changed.

IBM addressed this with POWER7 and newer in a fairly innovative way. They have an option called TurboCore mode which turns off half the cores. The ones still running can use the disabled core's caches, and because of the space available for heat dissipation, clock speed could be bumped up. The result was half the cores, but almost the same performance due to the faster clock and cache available.

Re:at least they have 4 and 8 core models as well

[tlhIngan]

IBM addressed this with POWER7 and newer in a fairly innovative way. They have an option called TurboCore mode which turns off half the cores. The ones still running can use the disabled core's caches, and because of the space available for heat dissipation, clock speed could be bumped up. The result was half the cores, but almost the same performance due to the faster clock and cache available.

Intel has it too - it's called "TurboBoost". Basically if the CPU is not under thermal restraint, and the other core is basically idle, then it will go into a practical shut-down state while the clock on the active core is boosted. This is designed less to save per-core licensing and more for single thread performance - if the workload consists of one thread, then rather than have the other core stay there idling for threads that aren't coming, it can be put in low power and the one running real code can be boosted to speed up

Takes into account a lot of PC workloads are still single-threaded but there are a few that can benefit from more cores.

Re:at least they have 4 and 8 core models as well

[mlts]

Of course, tossing in virtualization in the mix is fun as well. For example, if I'm sitting on two boxes with 36 cores, and run a relatively small Oracle instance for VMWare vCenter with one vCPU in fault-tolerant mode, I'm on the hook for 72 cores for the Oracle license. With the cost being around $60,000 per core for the enterprise tier, this can add up. Add to this something like vMotion HA where the license has to include every machine that -could- run the DB, and it can get painful even in the enterprise.

Re:at least they have 4 and 8 core models as well

[alen]

the single 12 core socket will have more on board cache and should outperform 2 sockets. plus the timing issues of running data between 2 physical CPU's

Re:at least they have 4 and 8 core models as well

[alen]

that would actually be nice since the higher core CPU's have more cache which has been shown to increase performance for decades now. i'll take a top end CPU and disable cores any day over buying a lower end CPU just to save on licensing

Re:at least they have 4 and 8 core models as well

[alen]

sounds like if i bent over my ass wouldn't be big enough for oracle

Re:at least they have 4 and 8 core models as well

[mlts]

Sybase is exactly the same. You can license it for development by the number of users, or production by the number of cores.

It can get so expensive due to the licensing model they use, that buying a POWER or SPARC machine actually saves money compared to putting it in a VM environment, just because of whatever the DB -can- touch for CPU cores has to be licensed.

I'm not sure about MS SQL server, but from what I read, it is pretty similar.

Re:at least they have 4 and 8 core models as well

[Ixpath]

How precious, you think MS SQL Server is a "real SQL server".

Re:at least they have 4 and 8 core models as well

[Ixpath]

I do have an arm chair. That must be the reason no one is using MS SQL for anything serious at Google, Apple, Facebook, Yahoo, Netflix, Fitbit, Twitter, etc. Too many arm chairs here in the Valley.

Re:at least they have 4 and 8 core models as well

[apraetor]

Having multiple sockets will only profit you in SOME cases; the rest of the time it will hurt performance -- execution units on the same die can share cached data. Threads which don't rely on passing large amounts of data between themselves (but which move a lot of data individually) would potentially get a bump in throughput using 2 sockets, assuming the bus isn't the bottleneck.

Re:at least they have 4 and 8 core models as well

[armanox]

I'm a fan of the way Oracle is doing the licensing on the Oracle DB Appliance - the box comes with lots of cores, but you license them on demand.

Re: at least they have 4 and 8 core models as well

[armanox]

Until Microsoft moves to IBM style licensing where you pay for the cores on the hardware, regardless of what you are assigning to the virtual machine.

Re:at least they have 4 and 8 core models as well

[0xG]

Well, in the "real world" we run numerous applications because we have numerous LOBs.
And when the developers of the software (that the LOB wants) says "we only support db 'x'"...
...you get on with it.
We're not big enough to "force" the developers to port their applications, just for us.
We're not rich enough to develop all of our own apps.

So, like many (most) IT organizations, you have to ignore the idealists and Utopian views, and just serve the business.

Comments like "real SQL server" are unconstructive at best.

Re:How can it be 50% improvement if

[NatasRevol]

It's unpossible! You math fail!

Re: Oooh Shiny

[binarylarry]

You lost me at Oracle.

Re:How can it be 50% improvement if

[jsm300]

You're correct that it is not a 50% improvement, but because of your bad math (1/3 of 12 is 4, not 6). The real fail is the article which got it wrong with respect to Xeon E5 V2. You can get Xeon E5 V2 chips with 15 cores, so the V3 chips are really only a 20% improvement.

Re:How can it be 50% improvement if

[jsm300]

That was supposed to be "but not because of your bad math ..."

Re: TSX

[plcurechax]

> These shiny new processor having working TSX instruction sets? The ones that are supposed to help with virtualization?

TSX is not for virtualization, but for transactional synchronization, it provides efficient transaction locking for multi-threaded applications. Not necessarily virtualization, although it can benefit from efficient locking as well

No, as far as I know, these have TSX disabled, or will be with a microcode update, as TSX isn't expected to be fixed until 2015 in Broadwell or Haswell-EX Xeons (not Haswell-EP which these are).

Re:Oooh Shiny

[jon3k]

And how many $ (and watts) per MIPS/GFLOPS/whatever-metric.

And the winner will still be...

[pigiron]

the architecture that can achieve the fastest speed on complex relational joins.

You remember normalized tables and joins of course because they aren't going away since they are the only program constructs that are remotely built on the solid foundation of real math set theory

AND they aren't even Turing machines!!! LOL.

Re:And the winner will still be...

[pigiron]

Partitioning a table over multiple drives has allowed RDMS's to utilize multiple processors for a single join ever since Teradata pioneered the hardware for this 30 years ago. The sky is the limit.

YAY! Lots of cores

[FlyingGuy]

This is really great for doing cpu intensive jobs but what seems to never get any love is moving massive amounts of data around. We need to put as much effort into buss bandwidth as we are with cpu's.

Re: Oooh Shiny

[caveqat101]

Sorry, Amd is needed to be kept afloat by Intel. Otherwise where is the entry electronics going to come from? Yeah, you and the megas can afford the super sized shit, that needs lox cooling and nuke plant power cables, but johnny on his homework needs something to use at home, grandma to call the kids, needs clean easy to use programs, with decent usage controls. Yes the fancy chip was needed,yes, but I feel, Intel's lost grip of the public market. About two hundred is what grandma's will spend. As the price of tablets drops, watch video conferencing, see if grandmas and johnnies are connecting, the look at price points. AMD is there, not Intel.

Re:Oooh Shiny

[armanox]

No idea yet. I'd expect AMD still rules on performance per $, and am interested in seeing the performance per watt and performance per core metrics. Interested in seeing how it lines up against POWER 8 and Intel's new Xeon (I personally wish AMD could compete in performance per watt against Intel. I'll keep wishing).

... not only earth ...

[Taco Cowboy]

"The answer to the ultimate question of life, the universe and everything is 42"
--- The Hitchhiker's Guide to the Galaxy

Re:... not only earth ...

[ArcadeMan]

http://x2.fjcdn.com/thumbnails...

Re:"AVX frequencies"

[WMD_88]

Per Anandtech, the 2.3GHz 18-core model has an AVX base frequency of 1.9GHz. That's the only hard number I've come across. It seems to simply act as a warning that, under heavy usage of AVX, the chip may clock down below the advertised base frequency. It could still run higher, of course, if only a few cores are being used.

That's what i'm talking bout

[fa2k]

Is there any reason for enthusiasts to choose the Core i7 over this?

Leave your nerd card at the door

[ArcadeMan]

If you don't know about 42, you don't belong here.

Re:Oooh Shiny

[sjames]

Will it actually play the cha-ching sound when the extra features self-enable?