Intel Xeon E5-2670 Deal and Price Tracking

[BackupProphet]

Yeah that is what I think too, maybe also a few more that build a lab for machine learning with budget constraints. Tensorflow scales nicely with many cpu cores.

In the other direction, cloud hosting has never been so fast and so cheap too. Hetzner recently released a nice cheap cloud for 2.5€ per month, that is with 2GB ram, 1 vcpu and damn fast storage (2500 iops with pg_test_fsync).

 

[MBastian]

I did made some plots a while back to help me decide. Compared to an E5-2667v2 (the most appealing for me) the top end parts aren't so much different if you add up the MHz per core. Bottom line for me: If you do not have very special requirements just go for the cheapest available. Imho you won't notice a +/- 15% difference in performance.
Btw.: Would short lived processes benefit from a higher base clock? I once read that turboing up ist not totally instantaneous, especially on older generations.

 

[alex_stief]

Btw.: Would short lived processes benefit from a higher base clock? I once read that turboing up ist not totally instantaneous, especially on older generations.

Unless you fiddled with the CPU frequency governor, a core will not run at base clock frequency while waiting for something to do. All CPUs of the same generation will sit around the same idle frequency of ~800MHz nowadays, so base clock is out of the equation for responsiveness.

 

[Joel]

Unless you fiddled with the CPU frequency governor, a core will not run at base clock frequency while waiting for something to do. All CPUs of the same generation will sit around the same idle frequency of ~800MHz nowadays, so base clock is out of the equation for responsiveness.

I'm pretty sure he was referring to this:




OP: Where "base" and "max turbo" come into play is when ALL cores are loaded. If just one or two are loaded, then max turbo is available.

 

[einstein]

If memory serves me right in CB15 a couple of E5-2670 will do 2k points, a couple of E5-2697 V2 3K. That's a 50% upgrade right there...in comparison the E5-2667 V2 won't reach 2.4K points... so I'm not so sure about the 15% difference against higher sku's in multi threaded scenarios.

I did made some plots a while back to help me decide. Compared to an E5-2667v2 (the most appealing for me) the top end parts aren't so much different if you add up the MHz per core. Bottom line for me: If you do not have very special requirements just go for the cheapest available. Imho you won't notice a +/- 15% difference in performance./

 

[Joel]

Sounds like core count is superior to clock speed for that workload, therefore a 2680v2 or even a 2660v2 would be a better choice than a 2667v2.

 

[MBastian]

If memory serves me right in CB15 a couple of E5-2670 will do 2k points, a couple of E5-2697 V2 3K. That's a 50% upgrade right there...in comparison the E5-2667 V2 won't reach 2.4K points... so I'm not so sure about the 15% difference against higher sku's in multi threaded scenarios.

In this case an E5-2697v2 would be 25% faster than an E5-2667v2, which sounds about right. I've ommited the low (10%) and high(25%) end as I count these as "special requirements".

 

[einstein]

Actually the E5-2667 V2 is somewhere in between those 2.

Sounds like core count is superior to clock speed for that workload, therefore a 2680v2 or even a 2660v2 would be a better choice than a 2667v2.

 

[Stereodude]

With my x264 test workload the E5-2687W v2 (8 cores) was ~25% faster than a E5-2670 v1 (8 cores). The E5-2690 v2 (10 cores) was 38.6% faster than a E5-2670 v1. The E5-2689 v1 was only 1.2% faster than the E5-2670 v1 (because the TDP held it back).

Can you get a similar performance boost for the money going v3/v4? Unless I missed something the v3/v4 E5 Xeons don't clock faster, they just have more cores. There aren't large generational IPC improvements.

At some point there are just too many cores to spread the workload across. 1080p encoding with x264 should be limited to 27 encoding threads due to the way the encoder works and the vertical resolution. x264 will spawn 1.5x the number of logical cores threads. So, a 10 core Xeon has 20 logical threads meaning x264 will spawn 30 threads, which is too many for optimal image quality. So you restrict it to 27 threads via the command line, but at some point it's just not going to run any faster by adding cores (with the threading restrictions). I'm not entirely sure where that point is. It's above 10 cores, but going wider the gains taper off.

 

[alex_stief]

I'm pretty sure he was referring to this:

I know. Hence my answer explaining why it is irrelevant.

 

[MBastian]

Sounds like core count is superior to clock speed for that workload, therefore a 2680v2 or even a 2660v2 would be a better choice than a 2667v2.

Sorry, I am not buying this. Apart from some extreme corner cases(3rd lvl cache size), heat or misconfigurations adding up the MHz per core should be the best case scenario.

 

[einstein]

Not sure if he is trying to sell you anything other than stating the obvious. MHz per core it's pretty much the worst case scenario in a program like Cinema 4D...a couple of E5-2697 V2 will have 24 cores/48 threads of ivy bridge goodness.... that's the same core count of 3 E5-2667 V2...lol but within the tdp of just 2

 

[Robert Fontaine]

I think the problem really is the squeeze on ram. People are making do with what the have. If ram hadn't gotten out hand I would have switched to ddr4 ages ago. I bought the motherboard and it has been sitting in a box for almost a year.

 

[Evan]

I think the problem really is the squeeze on ram. People are making due with what the have. If ram hadn't gotten out hand I would have switched to ddr4 ages ago. I bought the motherboard and it has been sitting in a box for almost a year.

Could well be something to do with it, people holding on to what the have a bit longer due to the next gen costs, also for example I have a good amount of e5 v2 which were all confined with 512gb ram so going scalable is probably only increasing to 768gb which is not a big % increase

 

[MBastian]

...a couple of E5-2697 V2 will have 24 cores/48 threads of ivy bridge goodness.... that's the same core count of 3 E5-2667 V2...lol but within the tdp of just 2

E5-2697 v2 has an all core turbo of 3.0 GHZ, an E5-2667v2 one of 3.6GHz. So a 50% increase in cores gives you a 25% increase in GHz. If cores would be enough everyone would go for the low cost E5-2651v2 from ebay...
Unless your workload is IO bound you should be able to utilize all cores at 100%. If it's IO bound it should run better with less cores ... if the application is CPU aware and/or properly configured that is.

Don't get me wrong, I do not want to start an argument "because I am right" but I do not see any tangible evidence to the contrary so far.

 

[wildpig1234]

I wonder if all the interest in crypto currency is creating greater demand for cpu"s and servers that's pushing up prices these days? Or perhaps lack of flood of parts from end of service contracts? Kind of weird that old technology like 2011 socket cpu"s and ddr3 ram is actually going up in price.

Its mostly supply side dependent. As well as people like me who buy cpu just to test it out and run benchmarks...lol.....

 

[Joel]

Sorry, I am not buying this. Apart from some extreme corner cases(3rd lvl cache size), heat or misconfigurations adding up the MHz per core should be the best case scenario.

"Best case scenario" doesn't always equate to real world performance.

Also, simply taking MHz * core count doesn't take TDP ceilings into account. Basically, you'll hit the TDP ceiling before all cores are pegged at 100%. Even if the processor isn't too hot, it'll throttle so it doesn't draw more power than it is designed to.

Note that I said:

for that workload...

If all workloads were the same, then Intel could cut it's number of Xeon SKUs in half.

Personally, I like the 2667v2 better than a 2697v2 for workstation usage, because it has higher clock speeds, which still matter for many programs that aren't written to take advantage of high core counts.

Example:

Adobe Photoshop CC Multi Core Performance

Note how many tasks scale linearly up to 3-4 cores and then hit a very flat plateau.

 

[skelleton]

With my x264 test workload the E5-2687W v2 (8 cores) was ~25% faster than a E5-2670 v1 (8 cores). The E5-2690 v2 (10 cores) was 38.6% faster than a E5-2670 v1. The E5-2689 v1 was only 1.2% faster than the E5-2670 v1 (because the TDP held it back).

Can you get a similar performance boost for the money going v3/v4? Unless I missed something the v3/v4 E5 Xeons don't clock faster, they just have more cores. There aren't large generational IPC improvements.

At some point there are just too many cores to spread the workload across. 1080p encoding with x264 should be limited to 27 encoding threads due to the way the encoder works and the vertical resolution. x264 will spawn 1.5x the number of logical cores threads. So, a 10 core Xeon has 20 logical threads meaning x264 will spawn 30 threads, which is too many for optimal image quality. So you restrict it to 27 threads via the command line, but at some point it's just not going to run any faster by adding cores (with the threading restrictions). I'm not entirely sure where that point is. It's above 10 cores, but going wider the gains taper off.

I work around that by running a bunch of Proxmox containers as encoding workers each with 8 cores.
I tend to encode BluRays of TV shows for my media center, so i still get a speed improvement with more cores.

 

[Stereodude]

I work around that by running a bunch of Proxmox containers as encoding workers each with 8 cores.
I tend to encode BluRays of TV shows for my media center, so i still get a speed improvement with more cores.

I've done similar things running 4 copies of handbrake concurrently using NUMA node and affinity controls to limit the portion of the hardware it runs on and how many threads it spawns. Still, I'm not always trying to convert 4 things at once.

 

[Ales]

Alright, so what's the current best bang for buck build, if doing a pretty beefy colocation (single 4u) web server?

I know there are a lot of variables to consider here... But just a general consensus.