1090T and DDR2

[lopri]

I've seen others curious about this and I myself was the same prior to acquiring my own Thuban CPU: How does Thuban perfom when paired with DDR2 sticks?

Short answer is "very well", but not "the very best". However, I'm exceptionally satisfied with this upgrade and I couldn't ask for more. You see, I'm biased - I haven't purchased a stick of DDR3 yet.. In 2010! And now thanks to AMD's new hexa-core CPUs, I feel set until SB and BD hit the market. Considering I bought my DDR2 sticks 3~5 years ago, I'm glad that I'm squeezing the last drop out of them. (and then some)

As I previously posted in the X6 Overclocking Thread, I established my 1090T was good for 4.0 GHz with minimal vCore increase. I was testing the chip itself, and benchmark results were not the focus. DDR2-667 sticks (which requires 2.0V just to run @DDR2-800) and 785G board, for instance, produced ~67 GFlops in Linpack.

Now I've moved the CPU into my main system, and set the memory @healthy DDR2-1000 along with the CPU-NB @2500MHz. Followings are various test results.

1090T @4.0 GHz (16x250) | CPU-NB @2500 MHz | DDR2-1000 (5-5-5)
Note: Ignore the strange frequencies and voltages. C1E is acting up all over the place in current BIOS.

Linpack




Everest


Cinebench R10


Cinebench 11.5


3DMark06 (Radeon HD 5800 @default)


x264 Bench v3.0

Results for x264.exe r1342 
-------------------------- 
encoded 1442 frames, 84.03 fps, 3898.70 kb/s
encoded 1442 frames, 84.18 fps, 3898.70 kb/s
encoded 1442 frames, 84.03 fps, 3898.70 kb/s
encoded 1442 frames, 84.18 fps, 3898.70 kb/s
encoded 1442 frames, 37.23 fps, 3970.06 kb/s
encoded 1442 frames, 37.02 fps, 3969.41 kb/s
encoded 1442 frames, 36.96 fps, 3969.50 kb/s
encoded 1442 frames, 37.23 fps, 3969.03 kb/s

 

[lopri]

I'm using this reserved spot to post DDR3 results for comparison. Take a look at this guy's 1035T.

http://www.xtremesystems.org/forums/showthread.php?t=252114

He is using 890FX chipset and DDR3. Of course bus speed is way higher since 1035T has max multiplier of 13.5. Also note that he has set the CPU-NB @3.0 GHz. (and whatever tweaks he applied) From the looks of it, I think benchmarking is his living so these results are likely at the high end of the spectrum.

1035T @4.05 GHz (13.5x300) | CPU-NB @3000 MHz | DDR3-1600 (7-7-7-1T)

Linpack (80 GFlops!?)


Cinebench R10


X264 Bench V3.0

Results for x264.exe r1342 
-------------------------- 
encoded 1442 frames, 92.52 fps, 3898.70 kb/s
encoded 1442 frames, 92.35 fps, 3898.70 kb/s
encoded 1442 frames, 92.34 fps, 3898.70 kb/s
encoded 1442 frames, 92.44 fps, 3898.70 kb/s
encoded 1442 frames, 38.37 fps, 3971.80 kb/s
encoded 1442 frames, 38.51 fps, 3972.09 kb/s
encoded 1442 frames, 38.58 fps, 3951.22 kb/s
encoded 1442 frames, 38.69 fps, 3968.90 kb/s

 

[BD231]

Nice, though there's really not much to question on DDR2 vs DDR3.

The general parity between top end DDR2 and typical ddr3 1600 should leave most looking at nothing more than power consumption/additional gaming performance. DDR3 is really just DDR2 with some new tricks and more bandwidth which equate to a lower power draw at the same performance levels.

Short of an older DDR2 AMD chipset performing worse than their newer DDR3 chipsets (highly unlikely, AMD has a pretty no frills line of chipsets) there isn't much you're missing out on with good DDR2 memory.

 

[Gikaseixas]

I might sell my 940BE and get a 1090. Good to know this, thanks for the very informative post.

 

[richierich1212]

yeah, thanks lopri. You're not going to lose much performance using DDR2 vs. DDR3 with AMD platform as long as you can keep timings tight.

 

[lopri]

There is one thing strange though: In certain single-threaded task, Thuban seems slower than Deneb clock-per-clock. It shows only in benchmarks almost as if it's a margin of error, but I think there is a trend. I suspect it has to do with L3/NB and whatever AMD had to do in order to pack 6 cores in the same power envelope. (looser internal timings for NB/L3?) I could be totally wrong on this, however.

 

[VirtualLarry]

There is one thing strange though: In certain single-threaded task, Thuban seems slower than Deneb clock-per-clock..

Could this be related to why Thuban seems slower in some games than Deneb, even with Turbocore?

 

[richierich1212]

Well Thuban is slower because L3 cache is shared among 6 cores instead of 4

 

[lopri]

I was thinking more of the opposite: Thuban's NB/L3 is tuned rather for DDR3 than DDR2, and it's more pronounced when the cache is not saturated (i.e. single-threaded workload). When cache is fully loaded and constantly swapping data, its limited capacity applies to both DDR2 and DDR3.

Well, then again that doesn't make much sense because the reviews comparing Deneb and Thuban both use DDR3.. So I don't have exact understanding yet. (nor concrete proof that single-threaded performance is not the same between the two)

Edit: Or it could be the hardware C1E.

 

[nyker96]

first of all nice benchmark runs. I also asked the same question weeks back regarding ddr2 vs ddr3 for thuban. (i got a 785G w/ ddr2 as well) but from your benchmarks, it seems it does fairly well.

The oc potential for this chip is also impressive, I got my x4 to just 3,25ghz with 1.4v while you are at 4ghz with same voltage. Don't know how amd tweaked the 45nm process to make the chip so ocable but they sure squeezed a ton out of it. very nice!

 

[BD231]

Introducing two additional cores is a lot of added traffic/delegating for an unaltered L3 and memory controller. The constant added hits to cache and memory even while some cores are idle should account for the discrepancy.

 

[thilanliyan]

The oc potential for this chip is also impressive, I got my x4 to just 3,25ghz with 1.4v while you are at 4ghz with same voltage. Don't know how amd tweaked the 45nm process to make the chip so ocable but they sure squeezed a ton out of it. very nice!

Check out the OCing thread:
http://forums.anandtech.com/showthread.php?t=2071268

FYI, lopri's chip is probably the best in terms of voltage for those in the database so don't necessarily expect you will have similar results. That said, the 1090T chips seem to require less voltage to get to 4GHz compared to the 1055T chips.

 

[chubbyfatazn]

It's 4:30 in the morning, but I'm pretty sure my eyes are seeing an "X6 1095T" in CPU-Z. Am I seeing things correctly? What's up with that?

 

[thilanliyan]

^Yeah I saw other people with a 1090 were seeing it as a 1095 in CPUz...just a bug in CPUz I'm guessing.

 

[lopri]

I posted some highly overclocked 1035T results (stole from XS) using DDR3 and very high NB frequency in the post #2. So while results using DDR2 are respectable, DDR3 along with high NB frequencies (1:2~1:2.5 seems ideal) gives the best performance out of Thubans. So the simple lesson to draw is: If you'd like to extend the life of your existing AM2+ platform, Thubans doesn't disappoint. However, if you're building from scratch for the best performing Thuban system, DDR3 and the new 8 series chipset are the way to go.

 

[n7]

That's an amazing system considering you're still using DDR2!

Nice results

 

[lopri]

Introducing two additional cores is a lot of added traffic/delegating for an unaltered L3 and memory controller. The constant added hits to cache and memory even while some cores are idle should account for the discrepancy.

I think this theory is quite sound. Is there a way for me to verify it?

That's an amazing system considering you're still using DDR2!

Nice results

Thank you. I did not know you had an i7-930, btw? Thought you had switched to i7-860 from i5-750.

 

[lopri]

Anyhow, I managed to convince someone who never overclocks that he does a favor for me swapping his C3 955BE with my C2, and ran some testing on previous 785G board. (This C3 revision 955BE turned out to be a good clocker on its own. My old 955BE C2 can't do 4.0 GHz on 64-bit Windows)

Single-threaded performance comparison @4.0 GHz (16x250, DDR2-1000):

Cinebench R10
Deneb: 4512
Thuban:4282

Cinebench 11.5
Deneb: 1.18
Thuban: 1.13

Note: I'm aware that each Cinebench run varies and tried to pick the most representative values. But as appears, I could never reach anywhere close to single-core score of 1.18 in Cinebench 11.5 using the 1090T @4.0 GHz

From this short testing, Thuban outperforms Deneb when multi-threaded memory performance matters. And Thuban's single-threaded performance, when paired with DDR2, seems a tad lower than Deneb + DDR2. And that situation is rectified by DDR3 if I take a look at the benches carried out by others. So my tentative hypo with regarded to single-threaded performance is:

Thuban + DDR2 < Deneb + DDR2 < Thuban + DDR3

Below are Deneb @4.0GHz screenshots.

Everest


Cinebench R10


Cinebench 11.5


Linpack


X264 Bench V3.0

Results for x264.exe r1342 
-------------------------- 
encoded 1442 frames, 81.23 fps, 3900.68 kb/s
encoded 1442 frames, 80.10 fps, 3900.68 kb/s
encoded 1442 frames, 81.16 fps, 3901.21 kb/s
encoded 1442 frames, 79.69 fps, 3901.21 kb/s
encoded 1442 frames, 25.77 fps, 3971.70 kb/s
encoded 1442 frames, 25.70 fps, 3970.66 kb/s
encoded 1442 frames, 25.72 fps, 3972.03 kb/s
encoded 1442 frames, 25.69 fps, 3971.59 kb/s

 

[Arkaign]

DDR3 will eventually run away from DDR2 when process technology and clock speeds ramp up (if it's not replaced by another standard by then), but for now it's just not a big difference whatsoever.

This reminds me of continual overlap in memory performance before the new gen really gets cooking :

PC133 Sdram was really no slower than DDR1-266 (compare KT133A vs. KT266 on top Athlon XP of the day), but then DDR-333+ started to run away with it when combined with more mature DDR chipsets and higher performance cpus.

DDR1 400 was really no slower than DDR2-533/667, particularly on the common high-latency modules that were widely available at the time, DDR400 1T CL2 modules actually outperformed most DDR2 south of 800mhz. But of course DDR2 eventually crashed in price along with getting better latency chips, then 1066+ hit, and it was cheaper to mfg anyway.

DDR2 to DDR3 is mirroring the DDR1 to DDR2 transition, driven mostly by Intel and OEM volume pushing, not much performance difference in the beginning, but I think we're starting to get to the tipping point soon where DDR2 will be a fading memory (lol). AMD will be the last stand, and these hex-cores might just be the EOL for CPU support now that ODMC is standard for everyone.

OP, fantastic work and post, thank you!

 

[lopri]

FYI, I've found the most comprehensive examination on Thuban's NB/L3/memory scaling here. I largely agree with his findings.

AMD Thuban NB and Memory Performance Scaling