Intel Launches Xeon E7-8800, 4800 v4 Broadwell-EX Family

Status
Not open for further replies.


The real crazy thing is that servers can and will benefit from every aspect of this unlike consumer desktops. Servers are just better at utilizing cores and performance enhancements than consumer desktops are.
 

bit_user

Polypheme
Ambassador
That's certainly what I thought.

I'm guessing they repurposed the pins for 8 PCIe lanes for the extra QPI channel. Even if you don't get all 3 QPI channels per slot, it's still pretty impressive that existing boards can support these with only a BIOS upgrade.

The article also states that these CPUs have 8 memory channels, but the specs on ark.intel.com indicate only 4.
 

bit_user

Polypheme
Ambassador
At $6841.00, I wonder if the E7-8893 v4 is their most expensive quad-core, ever...
 

PaulAlcorn

Managing Editor: News and Emerging Technology
Editor
Feb 24, 2015
858
315
19,360


Yes, only 32 PCIe lanes on the E7 v4.

The move up to eight memory channels comes via the Brickland platform, which allows up to eight DDR3 or DDR4 channels per socket. I added a graphic in the article (and a bit of text) to help clear it up. The platform links the on-die memory controllers to memory buffers, codenamed Jordan Creek, through four of the Intel SMI Gen 2 channels, thus inflating the memory channel count. Very exciting tech, it helps Intel compete with the impressive per-socket RAM capacity on the Power8 architecture.

12.png
[/URL][/img]
 

Haravikk

Distinguished
Sep 14, 2013
317
0
18,790
Servers are just better at utilizing cores and performance enhancements than consumer desktops are.

Handbrake can use all 16 of my i7 threads just as well as a xeon could
I suspect they meant in terms of total workload; even with Handbrake you're probably not encoding 100% of the time and most consumer desktops aren't used for anything near as demanding as that. Traditionally if you require a machine for encoding and similar tasks you buy a workstation, but there's certainly less need for it these days.

For example my main working machine used to be a 2008 Mac Pro with 2x quad core 3.2ghz xeon CPUs and a Nvidia 8800 GT. I've since replaced it with a tiny system using an i7-4790T (quad core, eight threads, 45W with HD4600) which actually outperforms the Mac Pro at most tasks, and in areas where the Mac Pro does still have an edge it's only a small one. But that's the nature of the game these days; processors are getting so efficient you can do a lot with a system made from consumer grade components. That said, I'd hesitate to call an i7 desktop as "consumer" desktop, more like "prosumer" as you don't spend the money for an i7 unless you think you're going to need that kind of performance; even in gaming systems it's usually a waste of money, so things like video encoding are more suitable tasks for it, but they're not really regular consumer tasks.

But ehm… yeah, for the kinds of systems these xeon processors are for you're looking at close to maximum utilisation at all times, probably using some kind of virtualisation to run as many jobs as possible on as many systems are required to meet demand. That usually means pushing the minimum number of servers to their limits so you can let others sleep until demand increases.

Anyway, as much as I'd love a 24 core (48 thread) CPU I think the price is a little steep for me ;)

I wonder, is the decrease in PCIe lanes a result of the increased memory lanes?
 

bit_user

Polypheme
Ambassador
Thanks for clarifying that. So, it's an off-chip de-interleave, more or less. I'm guessing memory transactions can get reordered, in the "Performance Mode", making it something other than a strict interleave (hence the name Scalable Memory Buffer).

I think it'd still be bottlenecked by the memory controller interface on the CPU, however. That would make this mostly about capacity, and not the level of performance improvement you'd expect to see by going from 4 -> 8 channels. I wonder how badly it affects latency.
 

bit_user

Polypheme
Ambassador
The question is: what would you do with such a thing? Realtime raytraced VR is one idea that comes to mind.

And why 24 GPUs? There would only be enough x16 slots for 8, and that assumes you've got nothing else on the CPU's PCIe lanes. If you drop to x8 links, I could see having 12 GPUs and keeping 8 lanes per CPU free for things like SSDs and 10/40/100 Gigabit NICs.

The good news is that you could probably do it, if you can manage to find the $$$. But, once you account for cooling and power, it'd end up being the size of a refrigerator. And the electricity bills...
 

hatib

Reputable
Jun 21, 2015
722
0
5,060

for video editing and simulation

 
oh dang only 24tb of ram now how am i suppose to have my entire file server on a ramdisk? i would like to see someone challenge intel though. power 8 has been competitive but not nearly enough to win back the monopoly intel has going.
 

bit_user

Polypheme
Ambassador
By video editing, perhaps you have in mind some complicated motion graphics effects and rendering? In any case, I think just a single-CPU box with a lot of RAM, fast storage, and a fast GPU would be enough for everything in video editing/production, except big rendering jobs.

It's actually a little tricky to think of tasks that would benefit from having all this in one box, but not suffer from the NUMA topology. Keep in mind that the memory & PCIe busses are partitioned amongst the CPUs, and that accessing a device or memory attached to another CPU means going through QPI. So, that would become a bottleneck, for certain workloads.

These systems have a sweet spot. For a task to run well on these systems, it would need to be latency and/or bandwidth sensitive beyond the limits of 10/40/100 gigabit ethernet, but not so much that QPI would become a bottleneck.

Many simulation & server tasks would perform similarly on cheaper, single-CPU boxes networked with 10/40/100 gigabit, which has the added benefit of scaling better. There's a nice writeup on nextplatform.com of how Facebook migrated from dual-E5 Xeons to single-CPU Xeon D systems, for better efficiency. For them, it seems QPI was just an overhead (in terms of power), since they had separable tasks which didn't need a lot of intercommunication.

That's why I go to realtime (i.e. highly realistic VR), as a use case.
 


i use the best i can get and at the moment xeons are the best by a significant amount. however that dosnt change the fact that i would like to see some pressure on intel to improve and if something is better by enough i wouldn't hesitate to give it a try. i was actually trying to get my hands on a power 8 rig for abit but gave up due to the insainly high price and now meh performance compared to the competition.
 


Agreed. I'm the same way, but with NVidia. :D Competition only breeds something good.
 


But in the workstation/server environments they do improve. It is just the consumer space that the improvements are not seen nor would they be seen. Servers will make use of 24 core CPUs happily and more memory since every normal IT setup has a VM setup where one server acts as multiple roles but each is a VM (VM for file server, VM for print server, VM for whatever server etc).

In fact it wasn't until more recently that Intel has been able to compete in some of the higher performance spaces against IBMs Power CPUs. If anything IBM, AMD and the other HPC CPU manufactures need to start trying to improve more.
 


i currently run 4 vms on my main rig 2 gaming/workstation ones that need a high single core to be possible (the e5-2670's come within 10-20% of a skylake i7, i have a small oc on them) a file server and a web server. so far these have been able to work very well for what i do however im still looking for a cpu that could potentially lead me to upgrade. i dont believe power cpus will come close with any shred of single core prefomence. however i am completely unsure of that as i can find anyone who has tested it and the only info i know is they are suposibly %20 faster in that category then power 6. and last time i checked the sparc m7 was still the king @ 32cores 256threads (the max most os will touch) and 4.1 ghz however i will never be able to touch one of those within the next decade.
 
Status
Not open for further replies.