Haswell-EP and Broadwell-EP Xeons on time

[rtsurfer]

You do know that Haswell is around 75-80% faster than IB, when hitting that TDP, right?

Can you please rewrite
this. I think you meant something else.
Because the way it is right now.
It tells me that you are stating that at full load and max clockspeed. Haswell is 75-80% faster than Ivy.

Which is a load of BS.

Please keep your homemade FUD out of this thread.

It is not a home made FUD. It is a very likely possibility given their track record of screwing the constumers on the mainstream platform for two subsequent generations.

Unless you know, some Intel fanboys have trouble being in touch with reality.

 

[ShintaiDK]

When you run Linpack. You get around 100Gflops on a 3570. And you get around 180Gflops on a 4670. And the only way to get power consumption up on Haswell is via AVX. And guess what Linpack uses.

It will be intersting to see if the server variants can break 700Gflops DP on a single socket with Linpack.

 

[rtsurfer]

That is a specific scenerio. In the Haswell reviews Anand and other reviewers did, the power consumption did go up on other programs, not just linpack.

Also the performance went up due to the new AVX instruction sets in Haswell. Which is a specific use case scenerio as not all programs use AVX.

Btw a side fact, I get 158 GFLOPS on my 4770K @ 4.5. It does go up when I disable HT though.

 

[ShintaiDK]

That is a specific scenerio. In the Haswell reviews Anand and other reviewers did, the power consumption did go up on other programs, not just linpack.

Also the performance went up due to the new AVX instruction sets in Haswell. Which is a specific use case scenerio as not all programs use AVX.

Btw a side fact, I get 158 GFLOPS on my 4770K @ 4.5. It does go up when I disable HT though.

Prime95 uses AVX too. Not to mention MB variance, BIOS etc.

I got both a 3570 and a 4670 system, both using same PSU. And the Haswell system uses less power, idle and load.

If you only get 158gflops at 4.5ghz. Then something is very wrong with your OC. Assuming you run Linpack 11 through its tests.

 

[rtsurfer]

Dude.
Prime 95 & Linpack are stress test, not programs we use for daily work.

I am well aware that Haswell is more power efficient at idle and light loads.
But once you go full throttle, it is a small step behind Ivy.

Also, dont worry about my OC. It is perfectly fine. It has been discussed on various forms around the Internet that HT on gives a slight worse performance in Linpack.

I get 180+ too if I disable HT.

Its just that I expected Intel to do a little bit something about it as they too use Linpack in their IBT and the XTU stress test is the biggest joke ever.

 

[ShintaiDK]

Dude.
Prime 95 & Linpack are stress test, not programs we use for daily work.

I am well aware that Haswell is more power efficient at idle and light loads.
But once you go full throttle, it is a small step behind Ivy.

Also, dont worry about my OC. It is perfectly fine. It has been discussed on various forms around the Internet that HT on gives a slight worse performance in Linpack.

I get 180+ too if I disable HT.

Its just that I expected Intel to do a little bit something about it as they too use Linpack in their IBT and the XTU stress test is the biggest joke ever.

I get 180Gflops at 3.6Ghz. So if you get 158Gflops at 4.5Ghz, something is wrong.

 

[rtsurfer]

I get 180Gflops at 3.6Ghz. So if you get 158Gflops at 4.5Ghz, something is wrong.

Nope.
Nada.
I think it is in line with what other people get.
I will still rerun it at stock and at 4.5 with HT off later on during the day.

Currently I have a Physics final in a few hours. So later.

 

[JoeRambo]

Nope.
Nada.
I think it is in line with what other people get.
I will still rerun it at stock and at 4.5 with HT off later on during the day.

Currently I have a Physics final in a few hours. So later.

Need to limit Linpack to real cores with thread affinity, otherwise results will be low.

Still, Haswell CPU is WONDEROUS beast, I don't care about Linpacks, but in our company workload we have the following stats ( all on 4.5Ghz machines, built by ICC 14 /O3 for SB).

SB: 15.7s
IB: 14s
HSW: 12:3s

and shocker - when recompiled for Haswell AVX2/BMI set, our purely integer load is suddenly executing in 11s. Those new bit manipulation instructions and finetuned instruction scheduling for extracting more out of CPU work wonders.

So there is very real world and very important advance for us.

 

[Ferzerp]

Need to limit Linpack to real cores with thread affinity, otherwise results will be low.

Yes, linpack is one of the *very* few workloads that never have a thread stall and so hyperthreading is a net negative. In fact, I've never encountered any real application that is this optimized.

 

[ehume]

Yes, linpack is one of the *very* few workloads that never have a thread stall and so hyperthreading is a net negative. In fact, I've never encountered any real application that is this optimized.

Somewhere in the Intel stuff they tell you that HT impairs Linpack results. Something about making choices slows it up.