i7 5820K @4.4 slower than i5 3570K@4.4

[sjerra]

Hi there,
First post in over 10 years of reading Anandtech. This is how baffled I am with what happened to me.

I have an I5 3570K in a little P8Z77i with 16Gb 1600 CL8 running at 4.6Ghz with SSD. It's been running stable ever since I bought it in 2013.
I don't game. I run 1D simulation and Solidworks & Ansys FEA (Maxwell 3D). I recently ran out of memory on a job.
So I decided to buy a new rig.
After reading a lot I settled on a X99-Deluxe with an i7-5820K and 8x16 DDR4 2400 CL16 with SSD.
After some tuning I managed to run it stable at 4.4Ghz. I could probably go a little higher but I need a reliable 24x7 rig.
Anyway, when I ran the simulation I wasn't impressed at all. Ok, I didn't have memory issues, but it didn't feel as fast as my old rig. So I did a test. Both PC's side by side, same model, press start button at the same time.......and yes, my I5-3570K is faster -though after a hour it throws a memory error-.
How much faster I can' say I didn't time it, but enough for me to feel it from the start so I estimate something like 7-10%.

That's an Ivy Bridge @ 4.6Ghz vs an Haswell-E @ 4.4Ghz.
So I tuned the Ivy bridge down to 4.4Ghz and it is still faster. Now, the margin has gone down to something I'm willing to suffer (2-3%)considering the memory availability but this is very weird. I'm not expecting miracles considering this is a tock generation, but at the same clock this rig should have beaten my previous gen i5.

All my bios settings are on auto, except the memory, that's been set to xmp (so 2400 CL16) and processor clock speed. The voltage profile is set in AIsuite 3.

Is there anybody who can confirm or deny this type of behaviour especially considering one has 1600 CL8 and the other one 'only' 2400 CL16? I know from the many discussions out there that memory speed seldom makes a measurable difference. But might this be the use case where it does?
Are there any tips on why I might not get the performance I should expect?

Thanks
sjerra

 

[Yuriman]

If you believe memory is the main difference between the two, why not swap the memory in the two systems and run your benches again? There are cases where cas latency is more important than bandwidth, but latency also decreases with frequency.

http://blog.design-point.com/blog/2012/march/hardware-requirements-for-solidworks.aspx

^ This blog makes the claim that Solidworks is not (very?) multithreaded, except when you have multiple drawing views, so when running that particular program, 10-11 of your 12 threads are likely idle. Haswell is slightly faster (by a few percent) per clock than Ivy Bridge in almost all cases, but the clockspeed advantage of the i5 probably makes up for that, so in most cases I'd expect the new and old system to run about the same, all else being equal.

https://forum.solidworks.com/thread/56989

Modeling is very linear, you have a feature tree, functions are solved one by one, so only one thread, some sub-task can benefit from some multithread optimization, but they are rare. So, one cpu, the quickest possible.
Rendering makes use of the most number of cores and most number of thread, rendering is a very very parallel task, one task can be easily cut into sub tasks, and then, sub sub tasks, then sub sub sub tasks, all independants between each other... so the more computing units (let it be logical or physical), the best. (You can buy Modo too, Anna Wood has a good rendering farm setup, and it's a more clever way to spend money)
Flow Simulation is the same as Rendering (minus the modo comment).
Simulation is the same as modeling. Some simulation benefits of better multithread optimization, but as far as the simulation I am doing (nonlinear with contacts), it is monothreaded.

Do buy lower latencies memory if you can afford them. Latency is expressed in cycle, the lower it is, the faster the cells can refresh before it can read/write a new value.

 

[Burpo]

You didn't need a new rig. You could have just added a Radeon HD7970..

http://www.tomshardware.com/reviews/geforce-gtx-titan-opencl-cuda-workstation,3474-8.html

 

[Ranulf]

I don't think a memory swap between the two systems will work.

 

[Flapdrol1337]

Did you monitor clock speeds? If you leave stuff in the bios on auto you might be hitting a powerconsumption limiter that automatically clocks back the cpu.

 

[RaistlinZ]

Noob question, but does Z77 only support 16GB RAM max, or was that just the limitation of the motherboard? Seems like the OP upgraded to a platform without researching if it would make sense for what he was looking to do with it.

 

[JDG1980]

If you believe memory is the main difference between the two, why not swap the memory in the two systems and run your benches again?https://forum.solidworks.com/thread/56989

He can't. The Ivy Bridge platform requires DDR3, whereas Haswell-E uses DDR4.

 

[Dave2150]

He can't. The Ivy Bridge platform requires DDR3, whereas Haswell-E uses DDR4.

Indeed, that poster is clueless.

 

[Magic Carpet]

You didn't need a new rig. You could have just added a Radeon HD7970..

http://www.tomshardware.com/reviews/geforce-gtx-titan-opencl-cuda-workstation,3474-8.html

Daaamn, that poor Titan. This app must be seriously GCN optimized.

 

[Dave2150]

Daaamn, that poor Titan. This app must be seriously GCN optimized.

Nope, AMD cards are just great for Compute.

 

[Hulk]

Assuming no GPU assist in either system the only reason I can think as to why the Ivy would be faster is because it is not hyperthreaded. On some applications hyperthreading can actually slow things down.

Turn off hyperthreading in the BIOS of the Haswell rig and run the test again to confirm this theory.

Outside of this or some strange anomaly with the Haswell rig, Haswell should beat Ivy every time in every application at equal clocks.

 

[chubbyfatazn]

Noob question, but does Z77 only support 16GB RAM max, or was that just the limitation of the motherboard? Seems like the OP upgraded to a platform without researching if it would make sense for what he was looking to do with it.

His board only had 2 slots (mITX), hence max 16GB. I used 32GB with Z68 just fine for awhile.

Odd that he went from mITX all the way to ATX. Obviously money isn't an issue for him, so whatever I guess...

I don't even know where he got 8x16GB for his system, I haven't seen any consumer 16GB DDR4 sticks out yet.

 

[crashtech]

There's no real data in the OP's post, just an anecdote about how he thinks his old rig is faster. Let's see some benchmarks, OP.

 

[Dave2150]

His board only had 2 slots (mITX), hence max 16GB. I used 32GB with Z68 just fine for awhile.

Odd that he went from mITX all the way to ATX. Obviously money isn't an issue for him, so whatever I guess...

I don't even know where he got 8x16GB for his system, I haven't seen any consumer 16GB DDR4 sticks out yet.

5820k supports 64GB of RAM maximum. So OP is not using 8X16GB DIMMS, even if he believes so. Probably just a typo.

 

[myocardia]

5820k supports 64GB of RAM maximum. So OP is not using 8X16GB DIMMS, even if he believes so. Probably just a typo.

Indeed, not to mention that the X99 chipset supports a maximum DIMM size of 8GB, and a maximum of 8 DIMMS.

 

[Burpo]

Oh really? Forget i7.. Think Xeon
Max Memory Size (dependent on memory type) 768 GB
http://ark.intel.com/products/83352/Intel-Xeon-Processor-E5-2620-v3-15M-Cache-2_40-GHz

 

[crashtech]

http://ark.intel.com/products/82932/Intel-Core-i7-5820K-Processor-15M-Cache-up-to-3_60-GHz?q=5820K

 

[myocardia]

Oh really?
Max Memory Size (dependent on memory type) 768 GB
http://ark.intel.com/products/83352/Intel-Xeon-Processor-E5-2620-v3-15M-Cache-2_40-GHz

That's a CPU. The X99 chipset supports up to a maximum of 8 DIMMS @ up to 8GB size.

 

[Burpo]

Each slot supports 16GB memory module, thus delivering a maximum capacity up to 128GB. ... ECC is supported with Intel® Xeon® processors.

http://www.asrock.com/mb/Intel/X99 Extreme4/

And that's a single CPU..

 

[myocardia]

Each slot supports 16GB memory module, thus delivering a maximum capacity up to 128GB. ... ECC is supported with Intel® Xeon® processors.

While everything you said here is true...it has nothing to do with the X99 chipset. None of that is true, when the Xeon you linked is running on the X99 chipset.

edit: I'm glad to see that someone has finally made a motherboard with that chipset, that supports what the 2011-3 Xeons can do. Strange that Asus did it with an ASRock board, though.

edit #2: It also doesn't apply to the OP, since his motherboard only supports up to 8x8GB DIMMS, even with Xeons.

 

[chubbyfatazn]

ECC is supported with Intel® Xeon® processors.

Yeah, but not with the consumer Haswell-E chips.

http://ark.intel.com/products/82932/Intel-Core-i7-5820K-Processor-15M-Cache-up-to-3_60-GHz

 

[Burpo]

The memory controller is on the CPU.. Just like X58 & X79, you are able to exceed stated maximum memory.. X58 max was 24Gb, yet may people are running 48Gb.. It's only dependent on the size of the sticks of ram you're using, and if your CPU will see it. Has little to do with the chipset..

"Gigabyte GA-7TESM Dual LGA1366 is a server board & holds the most memory ever for X58 (288Gb)"

http://www.legitreviews.com/gigabyte-ga-7tesm-server-board-supports-288gb-of-memory-new-record_11487

 

[sjerra]

Thanks for the many replies.
First. I made a mistake. I now have 8x8Gb=64Gb, not 8x16. It was late. sorry.
I'm also using a Firepro W5000 cause anything else (exept a quadro) gives me poor performance with any of the software tools I'm using. I don't think a Geforce would do me any good.
The Z77 board is a mitx indeed so I was limited to 2 slots. At the time I thought 16Gb would be ample. For solidworks modelling it is. It's just the ANSYS FEA that's a mem hog.
I ran out of memory so I halved my design. Then I still ran out of memory. That's how I estimated 32Gb would still not be enough. That's why I settled on a board with 8 memory slots otherwise I would have just bought a matx or atx version of the Z77 and dropped in another 16Gb of the same memory.
Anyway, my simulation finished overnight and it turns out it's using 74Gb so I still end up swapping to my ssd's. But I guess it's now manageable.

I will check if maybe the bios is limiting the power as mentioned by someone but I did monitor the frequency going up to 4.4GHz. I've already disabled hyperthreading as it doesn't offer significant benefits in Ansys Maxwell anyway.

I'm under the impression, it might actually be the memory as Yuriman mentioned. Maybe I should run a syntactic benchmark or something to verify. It would just be a little frustrating that dispite the many advices on the net it does actually make a significant difference. I should have gone for an older X79 based board. Not that it would have been any cheaper though.
How about the whole quadchannel story? Shouldn't this offset against the extra latency? or is this just a salespitch then?

Ah well, it least with this board I'm a bit futureproof.......Can't wait for that DDR4 3000 CL10!

Thanks again,

sjerra

 

[chubbyfatazn]

The memory controller is on the CPU.. Just like X58 & X79.. You are able to exceed stated maximum memory.. X58 max was 24Gb, yet may people are running 48Gb.. It's only dependent on the size of the sticks of ram you're using, and if your CPU will see it. Has little to do with the chipset..

"Gigabyte GA-7TESM Dual LGA1366 is a server board & holds the most memory ever for X58 (288Gb)"

http://forums.anandtech.com/showpost.php?p=37252364&postcount=2100

http://www.legitreviews.com/gigabyte-ga-7tesm-server-board-supports-288gb-of-memory-new-record_11487

The regular Bloomfield i7s had a 36-bit address bus for a theoretical max of 64GB. The Xeons had a 40-bit address bus for a theoretical max of 1TB.

 

[sjerra]

The passmark verdict is in:
Numbers are written first Ivy, then Haswell-E (both @4.4GHz)
system: 2861 - 2949
proc: 8935 - 12655
Mem: 2731 - 2193
Disk: 4529 - 5113

But, the devil is in the detail:
CPU-single threaded: 2434 - 1991 (Millions of operations per second)
So in single threading, my Ivy is a good bit faster then my Haswell-E. This seems to be a known and has been reported on the net (including Anandtech). Unfortunately I have not paid enough attention to it. My simulations run multi-threaded, but only the actual matrix solver. Large parts of it, like setting up the solver and matrix is all single threaded.

It is what it is! I needed the 64Gb memory support.
Anyway thanks for all the input and if there is anybody in a similar situation: Lessons learned!

sjerra