Interesting tidbit about i7 uncore

[cyclone3d]

So I have decided to do a little bit of experimenting today to see what affects the RAM throughput and latency on an LGA1366 setup.

Everything will be tested in Triple channel mode with the rig in my sig... currenltly running at 20x200 instead of 21x200.

So far, it looks like uncore speed has a pretty decent sway on this.

I did an Everest test with the RAM at 1600 and the uncore at 3200.

I then raised the uncore to 4000 without changing anything else.

All of the throughputs went up a decent amount and latency went down from 47.8ns to 46.1-2ns

I can't really get an good figure on the throughputs in everest because you can run the test multiple times and results will vary quite a bit.

I will probably start using Sisoft Sandra as the results are a lot more steady as well as a lot more sensitive to small timing changes.

 

[cyclone3d]

So here is a more complete analysis using Sisoft Sandra.

Uncore makes a lot more difference in latency than does tightening up the timings.

Going from 1600 to 2000 gives you way more throughput than just tightening up the timings at 1600.

------------------------------------------------------------------------------------------------------------------------
Default timings 1T

8-9-9-28-5-88-14-8-24-4 | 1N-56-58-58 | 7-7-18-9-9-9-76-4-7-7-4

1600/3200
-----------------------------
27GB/S

Latency
Overall : 58ns
Speed Factor 58.30

1-16kB 1.0ns/4.0clocks
64kB 2.5ns/10.0
256kB 2.6ns/10.2
1MB 11.6ns/46.6
4MB 13.3ns/53.0
16MB 55.3ns/221.2
64MB 58.3ns/233.3
-----------------------------

1600/4000
-----------------------------
28GB/s

Latency
Overall : 55ns
Speed Factor 55.20

1-16kB 1.0ns/4.0clocks
64kB 2.5ns/10.0
256kB 2.6ns/10.2
1MB 10.3ns/41.3
4MB 12.0ns/47.9
16MB 52.5ns/209.9
64MB 55.3ns/221.1
-----------------------------

2000 tightened base timings - timings I can run when at DDR3-2000
8-9-7-23-5-100-6-5-18-4

1600/3200
-----------------------------
Aggregate 27.36GB/s
Integer B/F 27.4GB/s
Float B/F 27.32GB/s

Latency
Overall : 58ns
Speed Factor 58.30

1-16kB 1.0ns/4.0clocks
64kB 2.5ns/10.0
256kB 2.6ns/10.2
1MB 11.7ns/46.6
4MB 13.3ns/53.1
16MB 55.2ns/220.9
64MB 58.3ns/233.4
-----------------------------

1600/4000
-----------------------------
Aggregate 28.39GB/s
Integer B/F 28.37GB/s
Float B/F 28.4GB/s

Latency
Overall : 55ns
Speed Factor 55.10

1-16kB 1.0ns/4.0clocks
64kB 2.5ns/10.0
256kB 2.6ns/10.2
1MB 10.3ns/41.3
4MB 12.0ns/47.9
16MB 52.2ns/208.7
64MB 55.2ns/220.7
-----------------------------

1600 tightened base timings - timings I can run when at DDR3-1600
6-7-5-18-3-82-4-3-16-4

1600/3200
-----------------------------
28GB/s

Latency
Overall : 54ns
Speed Factor 54.20

1-16kB 1.0ns/4.0clocks
64kB 2.5ns/10.0
256kB 2.6ns/10.2
1MB 11.8ns/47.1
4MB 13.3ns/53.2
16MB 51.3ns/205.3
64MB 54.3ns/217.1
-----------------------------

1600/4000
-----------------------------
29GB/s

Latency
Overall : 51ns
Speed Factor 50.90

1-16kB 1.0ns/4.0clocks
64kB 2.5ns/10.0
256kB 2.6ns/10.2
1MB 10.3ns/41.3
4MB 12.0ns/47.9
16MB 48.0ns/192.2
64MB 50.9ns/203.7
-----------------------------

2000/4000 tightened base timings
-----------------------------
Aggregate 34.00GB/s
Integer B/F 33.87GB/s
Float B/F 34.00GB/s

Latency
Overall : 49ns
Speed Factor 49.30

1-16kB 1.0ns/4.0clocks
64kB 2.5ns/10.0
256kB 2.6ns/10.2
1MB 10.3ns/41.3
4MB 12.0ns/48.0
16MB 46.6ns/186.4
64MB 49.3ns/197.2
-----------------------------

2000/4000 - tightened base and sub-timings
-----------------------------
Aggregate 34.35GB/s
Integer B/F 34.34GB/s
Float B/F 34.35GB/s

Latency
Overall : 49ns
Speed Factor 49.10

1-16kB 1.0ns/4.0clocks
64kB 2.5ns/10.0
256kB 2.6ns/10.2
1MB 10.3ns/41.3
4MB 12.0ns/47.9
16MB 46.5ns/186.1
64MB 49.2ns/196.7
-----------------------------

 

[Chillernr1]

algemeen geweten feit voor uncore hoger te zetten voor mer performance,

zeker bij benchen .

 

[cyclone3d]

algemeen geweten feit voor uncore hoger te zetten voor mer performance,

zeker bij benchen .

"generally known fact uncore for turning up for EIA performance,

especially when benching."

For those of use who don't speak Dutch...

Well, I hadn't come across anything showing this before and I was curious.

 

[faiz23]

interesting posts and screen shots

 

[sn0man]

Well, I hadn't come across anything showing this before and I was curious.

Eva at i4memory has done a ton of testing with this.

Increases in Uncore speed alone account for quite a bit of performance (benchmark-wise, at least). A lot of the performance gain you see from using high speed memory is simply because the Uncore speed has to be raised along with it.