[INQ] Performance RAM will damage your Nehalem

Printable View

Show 100 post(s) from this thread on one page

10-04-2008, 03:36 PM
adamsleath

...back to 'low latency' ram :hehe:
10-04-2008, 05:37 PM
zer0kewl

is it still worth it to upgrade to nehalem despite this actualy ?
10-04-2008, 05:54 PM
Mechromancer

Quote:

Originally Posted by zer0kewl

is it still worth it to upgrade to nehalem despite this actualy ?

It will have very high bandwidth for the QPI versions. Performance should be decent too, but if you already have a highend Intel Quad then you won't be missing anything for a year or two. I really wonder if these chips will be durable and robust enough for XS style enthusiasts too lol.
10-04-2008, 06:26 PM
zer0kewl

Quote:

Originally Posted by Mechromancer

It will have very high bandwidth for the QPI versions. Performance should be decent too, but if you already have a highend Intel Quad then you won't be missing anything for a year or two. I really wonder if these chips will be durable and robust enough for XS style enthusiasts too lol.

i got e6750 on p5k mainboard atm.
with masive dpc latency issues atm wich i'm sick off heck i want to upgrade to something decent for a change ;>
10-04-2008, 08:00 PM
Chad Boga

Quote:

Originally Posted by JumpingJack

:) :) A few reasons --- it's easy, it's free, and there exists a lot of data for comparisons.

Though I understand the rhetorical meaning of your question ... and I don't disagree. That is why the benchmark should be put into context. It is not a good benchmark to compare against other architectures as a general performance benchmark, thus because AMD does not do well at SP vs Intel does not lend to making a general statement about the two products.

What SuperPI does do is exercise just a few of the many facets of a CPU. What good is it then? Well, it does a few things -- it helps ascertain the scaling of a CPU and drives straight to isolate only key components. It is not BW sensitive, it is not cache sensitive (this 'it runs all in cache crud is a bunch of bunk when using it to analyze a CPU').

SuperPI is inherently recursive, thus a huge loop with a conditional clause of 'when to stop'. Intel does well in this benchmark because they have strong logic in branch predictors and loop detectors. Intel has mentioned specifically that they reworked the loop detection logic in Nehalem, so seeing this improvement clock for clock in Nehalem is a good indication that some of the core tweaks made an impact.

The second thing SuperPI is good at, because it is so simple and straight forward, is to validate CPUID and clock speed. SuperPI should scale well, regardless of the architecture tested. As such, with a good baseline of data, a CPU clock can be extrapolated from the SuperPI run -- if they don't agree (superPI and CPUID) then it is reasonable to suspect something is 'fishy'.

Thank you for such an extensive answer. :clap:
10-05-2008, 02:32 AM
naokaji

just seen this, 6x1GB DDR3 kit for nehalem as preorder, 1,7-2.0V... so highend ram manufacturers are going to stick to "high" volts for now as it seems.
10-05-2008, 08:32 AM
bingo13

Quote:

Originally Posted by BulldogPO

One question, has somebody really destroyed Nehalem with high Vdimm?

I have seen three of them destroyed with memory voltages running at 1.85V~2.0V over a two week period. ;)
10-05-2008, 08:42 AM
bingo13

Quote:

Originally Posted by Tony

QPI is essentially HT, its all based off the same technology and as such suffers from the same disadvantages when it comes to high vdimm. Phenom when first released had some serious limits to Vdimm, i7 has the same limits and for the same reasons...the thing you guys keep missing though is you may NOT need the ubber high vdimm and I for one will not be fricken interested in running 2000MHZ on it either, for me the thought of low latencies tri channel is way to tempting, 1333 6-6-6- or even 5-5-5- if the MCH supports cas5 is way to mouthwateringly good to pass up. Also we are done with TRD (thanks the Intel gods ;)) and the tricks motherboard manufacturers played to get boards doing silly FSB...as if FSB at 2400 matters :shakeshead:

I think we are dawning on a new fresh memory overclocking day with i7...yes some will push 2000mhz, some will think the read bandwidth is the be all and end off of benches, those who are clever will look to real tight access latency and huge write/copy speed.

1333 5-5-5- or 6-6-6- is where I hope to be, or even 6-5-5- ;) and 1.6V or so.

I have the boards, just need the CPU now ;)

To add on to Tony's message, all I can say is that properly setting up a i7/X58 system at DDR3-1554 CAS6 (1.65V, 6GB) resulted in memory copy and latency numbers that were significantly better than DDR3-1866 at CAS9 (1.7V, 6GB, CAS8 is not stable yet). In fact, CAS5 at 1333 is just stupid fast for the "majority" of users. :D
10-07-2008, 05:31 AM
Revv23

Quote:

Originally Posted by Shintai

http://www.dvhardware.net/article24148.html
http://www.newegg.com/Product/Produc...82E16820146873
http://www.newegg.com/Product/Produc...82E16820231185

Btw, bandwidth matters, not latency.

Tell that to my BH-5 ram.
10-07-2008, 06:52 AM
Tony

Quote:

Originally Posted by bingo13

To add on to Tony's message, all I can say is that properly setting up a i7/X58 system at DDR3-1554 CAS6 (1.65V, 6GB) resulted in memory copy and latency numbers that were significantly better than DDR3-1866 at CAS9 (1.7V, 6GB, CAS8 is not stable yet). In fact, CAS5 at 1333 is just stupid fast for the "majority" of users. :D

Yup and the majority of users is where the money will be.

Im advising OCZ to push for 1.6V cas5 or cas6 parts at 1333, easy to support, easy for end users to set up, and easier on the wallet as they should be easier to bin.

there will of course be those who want to show off, i just do not really want OCZ promoting an easy way to kill a cpu.
10-07-2008, 12:58 PM
zads27

Quote:

Originally Posted by Kasparz

There are lots of chips that can do 1800MHz under 1.7Vdimm. And yes, Nehalem IMC will bring better memory speed at same voltage, so nothing to worry about. Also by setting higher cpu voltage you can set higher Vdimm too. I don't see why is anyone shouting about this. Many people killed their CPU's back in K8 days with BH5.

Yes, definitely.
The problem is, with current Core 2 systems and chipsets, the enthusiast wants the highest speed possible, no matter the voltage required.
There is no market for the JEDEC voltage, overclocked yet not the fastest memory.
That's why Wintec Industries has been holding off on the release of a 1GB 1800 CL9 1.5V module since July..
We will likely release this as a tri-channel kit with the Nehalem release.

And really, the memory bandwidth (all read/write/copy/latency) on Nehalem is a huge leap forward, even at slow speeds.
Running tri-channel DDR3-1066 CL6 nets much higher bandwidth numbers (20-50%) and lower latencies than dual-channel DDR3-2000 CL8.
10-07-2008, 01:02 PM
Astratuner

Quote:

Originally Posted by zads27

Running tri-channel DDR3-1066 CL6 nets much higher bandwidth numbers (20-50%) and lower latencies than dual-channel DDR3-2000 CL8.

Can you please explain that to me hehe

tks
10-07-2008, 01:09 PM
zads27

1 Attachment(s)

Quote:

Originally Posted by Astratuner

Can you please explain that to me hehe

tks

Take these memory bandwidth numbers and multiply by ~1.2 or 1.5,
that's what tri-channel DDR3-1066 CL6 is doing.
10-07-2008, 01:33 PM
Blauhung

Quote:

Originally Posted by Astratuner

Can you please explain that to me hehe

tks

Bandwidth isn't as easy as 2*2000>3*1067. That's maximum theoretical bandwidth. First off, the memory controller on bearlake (P35/X48) and eaglelake (P45) isn't setup to be able to use the maximum bandwidth made available by running dual channel at 2GHz on the memory as the design was never intended to ever use that much information. The DDR2 design was changed slightly to be able to interface with DDR3. I'm sure someone has the #'s available, but I believe that at some point even at current achievable DDR2 speeds the FSB ends up saturated with data and all that extra bandwidth is going to waste

On Nehalem though, the memory controller was specifically designed to run 3 channels and be able to feed near the maximum bandwidth to the rest of the chip. The interface between the controller and the cache is now much much wider then it was before when all the information had to travel along the FSB.

This is also how latencies have improved vastly. With FSB based chipsets, the path for a memory call takes 4 hops. From the CPU to the NB, NB to memory, memory to NB, NB to CPU. Each hop requires information generated and sent to the next stop, the information must then be read and then sent on. The FSB is also somewhat laggy in its nature. On nehalem, the CPU uses dedicated pathways to the memory controller to request information stored in memory. These paths are probably on the order of 100-1,000 times shorter then the ones from CPU to memory controller on the NB and can operate at much lower voltages and maintain much higher signal integrity allowing them to run much faster and wider. The interface to the memory is now dedicated to carrying only read or write requests, where the FSB was the channel for all information to or from the CPU.
10-07-2008, 01:50 PM
Astratuner

thanks for the piece of information guys!
10-08-2008, 01:41 AM
Nedjo

Quote:

Originally Posted by Tony

Yup and the majority of users is where the money will be.

Im advising OCZ to push for 1.6V cas5 or cas6 parts at 1333, easy to support, easy for end users to set up, and easier on the wallet as they should be easier to bin.

there will of course be those who want to show off, i just do not really want OCZ promoting an easy way to kill a cpu.

Tony, how viable is DDR3-1600 @ 1.6(5) say CL8/9?
10-08-2008, 09:12 AM
zads27

Quote:

Originally Posted by Nedjo

Tony, how viable is DDR3-1600 @ 1.6(5) say CL8/9?

Depends.
If you're talking 1GB DIMM, like I said earlier, Wintec Industries already has an unreleased part that does DDR3-1800 CL9 @ 1.5V.
If you're talking 2GB DIMM, we have a few products that will do DDR3-1600 CL9 at 1.5V, but as far as I recall, no 1600 CL8 product @ 1.5V/1.6V. We haven't been actively testing for this spec yet, but as Nehalem launch gets nearer, we'll create one and release it concurrently with all the other nehalem stuff.
10-08-2008, 09:24 AM
Nedjo

thank you zads looking forward to yours product.... although on Deneb ;)
10-08-2008, 10:18 AM
T_Flight

Quote:

Originally Posted by Tony

Yup and the majority of users is where the money will be.

Im advising OCZ to push for 1.6V cas5 or cas6 parts at 1333, easy to support, easy for end users to set up, and easier on the wallet as they should be easier to bin.

there will of course be those who want to show off, i just do not really want OCZ promoting an easy way to kill a cpu.

I've got a new saying Tony... "It's not Xtreme if you kill it." hehe :D

Me no kill a i7 CPU...me make CPU go weal fasssst. :yepp:
10-08-2008, 10:26 AM
Astratuner

2 asus motherboards have a fix for this and allow the user to boost the vdimm up to 2.4V :D
10-08-2008, 10:37 AM
savantu

Quote:

Originally Posted by Tony

QPI is essentially HT, its all based off the same technology and as such suffers from the same disadvantages when it comes to high vdimm. ...
...

?? They are really different.What are u referring to ?
10-08-2008, 11:16 AM
Kasparz

Quote:

Originally Posted by naokaji

just seen this, 6x1GB DDR3 kit for nehalem as preorder, 1,7-2.0V... so highend ram manufacturers are going to stick to "high" volts for now as it seems.

It has noting to do with Cellshock sticking to this or that, these are old sticks sold as 6pcs. Manufacturer can put higher volts, as long as they do caution about potential harm while using Vcore at stock, or using this kit in motherboards that doesn't support separate Vdimm feed.
Honestly, we can release TriKits even today if you wish, but we will test boards from various manufacturers to ensure XMP is working properly, and how well it overclocks with various boards. Even now on X38/X48/P45 many boards read XMP wrong, and this cause many problems in some cases to even boot up system. Of course, you can skip the trouble and make just JEDEC tables without extra XMP, but that's not an excuse to make proper enthusiast product.
Tony is right(as he always is) about low latency. Remember when AM2 came out, it was slower than s939 due to good old low latency BH5. Memory controller inside CPU loves low latency and this is the main goal.

What comes to heatspreaders, in some boards like Intel Smackover X58, ram slots are next to each other. In this case you can't put FlexXLC or ReaperX modules with heatspreaders. Thats why you have to be more careful by designing both engineering side of memory itself and cooling.
I believe, Qimonda and Elpida is great for high freq, low voltage, but if it's CL9 at 1800MHz, it is useless. 1333MHz CL6/1600Mhz CL7 would be way faster.
10-08-2008, 03:29 PM
Tony

we have cas9 1800 parts 1.5ish volts...1400ish cas5/6 -5-4- is faster...so you can run at 1800 and be slow or run 1400 to 1500 with tight latency and be fast...choice is yours ;)

No tRD in this one boys, remember this...its all about hitting the sweet spot now and finding what is fast.

We are over FSB, now dawns a new age of low bus speeds and ram tweaking a whole new way...im just waiting for all the Intel fanbois saying its like clocking ram on AMD boards...LOL
10-08-2008, 03:31 PM
zlojack

Well, I'm going to be stuck in the FSB age for a while yet :D

So I'll be after high-clocking modules for the time being.
10-08-2008, 03:34 PM
Tony

Quote:

Originally Posted by savantu

?? They are really different.What are u referring to ?

LOL the name is yes...the tech...well lets wait for the reviews and whether they mention where the tech actually comes from.And before you ask it is NOT AMD, you need to think where they got it from.

Can you really see Intel joining the HT consortium with AMD and NVidia??...me neither ;)

Quote:

2 asus motherboards have a fix for this and allow the user to boost the vdimm up to 2.4V

well lets see you buy a 965 and run 2.4V to the ram then, feeling brave??

You can strengthen the CPU just as we did with AMD's where you increase vdimm to strengthen the diodes, but to much and time will break them down and then you are out of pocket.
Plus why inject all that heat when you can run 1.6V ish and be just as fast??

Show 100 post(s) from this thread on one page