• Welcome to Overclockers Forums! Join us to reply in threads, receive reduced ads, and to customize your site experience!

Question about A64 939 and running RAM Asynch

Overclockers is supported by our readers. When you click a link to make a purchase, we may earn a commission. Learn More.
There are two facets to this "debate" which I believe is causing this confusion.

Firstly, does using a divider explicitly decrease performance? No.

What I mean is, is 9x305 with a 166 ( CPU/11) divider slower than 11x250 "1:1" (200MHz or CPU/11 divider) ? The answer is no.

Using a different set of multipliers with dividers nets you the same performance as long as the memory speed remains constant


Does running memory slower take a performance hit? Of course it does.


When people say that dividers take no performance hit, it resonates with older platforms (P4, AXP...) that had an explicit latency increase simply by using a divider, which the A64 does not have.

Bottom line, run your memory as fast as you can, using any possible combo. Which combo of multipliers and dividers it is doesn't matter, just like Gregory said.

Gregory_WE said:
What I'm gathering is if your RAM speed is still the same synch and asynch, the performance would be virtually the same - which is what I was specifically wondering about. Obviously, a system at 300 1:1 will be faster than a system at 300 5:6.
 
Gautam said:
There are two facets to this "debate" which I believe is causing this confusion.

Firstly, does using a divider explicitly decrease performance? No.

What I mean is, is 9x305 with a 166 ( CPU/11) divider slower than 11x250 "1:1" (200MHz or CPU/11 divider) ? The answer is no.

Using a different set of multipliers with dividers nets you the same performance as long as the memory speed remains constant


Does running memory slower take a performance hit? Of course it does.


When people say that dividers take no performance hit, it resonates with older platforms (P4, AXP...) that had an explicit latency increase simply by using a divider, which the A64 does not have.

Bottom line, run your memory as fast as you can, using any possible combo. Which combo of multipliers and dividers it is doesn't matter, just like Gregory said.


:clap: :attn:
 
deception`` said:
Sorry, but anyone who thinks that there is no performance penalty for using dividers is flat out wrong. While it is nowhere near as significant as on an Athlon XP system, performance differences do in fact exist. Here's a good article that should clear everything up:

http://www.insanetek.com/index.php?page=overclocka64

deception``
Comparing the first two columns it would seem to indicate the the one with the 5:6 divider runs better performance wise than the one running 1:1 - this despite having a LOWER ram frequency!! The problem however is that the memory timings differ and also one is at 1T whereas the 1:1 is at 2T so this comparison does not really have any basis quite frankly and should reallybe ignored.
What is required is both at either 1T or 2T, similar timings, with identical ram and cpu frequencies as in the scenario I gave.
 
glock19owner said:
OC Dect:

If you know of a good way to run a 3200 like you stated with the 3500 above...let me know and I will rum some benchmarks with latency test included...

Sure - stock settings.
First up 10 x 200 at 1:1, then at 8 x 250 with a 5:6 divider on (RAM:HTT). That way both cpus are at 2GHz and the ram frequencies are both 200Mhz. That would clear it up once and for all - but remember for it to be statistically significant it would have to exceed the limits of error (whatever they are in the likes of 3DM01 etc).
 
I've run these tests and posted them on a thread around here for everyone. The simple fact is that memory dividers on the A64 DO NOT create a performance hit in and of themselves.

The results are here for anyone that's interested:

http://www.ocforums.com/showthread.php?t=362703
 
Last edited:
OC Detective said:
Simple way to prove it is to get someone to test a 3500+ run it at 229 x 11 at 1:1 and 280 x 9 at 5:6 that way the cpu and ram frequencies are virtually identical and if running async does give a performance hit then it would show up in benchmarks.


That is not a fair comparsion because the fsb are different. Of course they would perform viturally the same because the cpu and bandwith end up about the same, however try running 266x9 1:1 and 266x9 with "166" divider, youll lose alot of ram bandwith and performance. This is what im trying to say that async takes a big performance hit where youd need 150MHz higher cpu clocks to make up for the loss of ram bandwith.


"First up 10 x 200 at 1:1, then at 8 x 250 with a 5:6 divider on (RAM:HTT). That way both cpus are at 2GHz and the ram frequencies are both 200Mhz. That would clear it up once and for all - but remember for it to be statistically significant it would have to exceed the limits of error (whatever they are in the likes of 3DM01 etc)."


once again, HTT doesnt matter, cpu clocks are equal and ram bandwith is equal. However 250x8 1:1 would crush 200x10 1:1. Take my case, 262x9 1:1 is about as fast as 280x9 with 166 divider. What happened there? My ram bandwith decreased because running async lowered my final ram MHz below 262. If there was a divider where I could run 280x9 with the ram at 262 then yes it would have the same ram bandwith as 262x9 but the extra cpu MHz would pull me quite ahead. Problem is, there is only a handful of dividers, mine has 100, 133, 166 and 200. some mobos have a 183 but NOT mine!
 
Gregory_WE said:
What I'm gathering is if your RAM speed is still the same synch and asynch, the performance would be virtually the same - which is what I was specifically wondering about. Obviously, a system at 300 1:1 will be faster than a system at 300 5:6.
Not neccesarraly true. On A64s higher multis also net better bandwith. So overall its about balance just like it was back in the AXP days
 
Overclocker550 said:
That is not a fair comparsion because the fsb are different. Of course they would perform viturally the same because the cpu and bandwith end up about the same, however try running 266x9 1:1 and 266x9 with "166" divider, youll lose alot of ram bandwith and performance. This is what im trying to say that async takes a big performance hit where youd need 150MHz higher cpu clocks to make up for the loss of ram bandwith.


"First up 10 x 200 at 1:1, then at 8 x 250 with a 5:6 divider on (RAM:HTT). That way both cpus are at 2GHz and the ram frequencies are both 200Mhz. That would clear it up once and for all - but remember for it to be statistically significant it would have to exceed the limits of error (whatever they are in the likes of 3DM01 etc)."


once again, HTT doesnt matter, cpu clocks are equal and ram bandwith is equal. However 250x8 1:1 would crush 200x10 1:1. Take my case, 262x9 1:1 is about as fast as 280x9 with 166 divider. What happened there? My ram bandwith decreased because running async lowered my final ram MHz below 262. If there was a divider where I could run 280x9 with the ram at 262 then yes it would have the same ram bandwith as 262x9 but the extra cpu MHz would pull me quite ahead. Problem is, there is only a handful of dividers, mine has 100, 133, 166 and 200. some mobos have a 183 but NOT mine!

WOW...just WOW

Okay...I'm better now. :)

Bro in every single one of the cases you cited it's the lower memory speed damaging ram bandwidth and not the divider itself. Saying that running async hurts memory bandwidth is false because it's not the divider itself that's hurting memory performance it's the lower speed.

280x9 w/ a 166 memlock means memory is running at 229Mhz. I hardly think it's a big shock that bandwidth is lower than 262x9 1:1. Now what happens if you compare bandwidth at 315x8 with the same 166 memory lock? Same performance? Yup. So the divider didn't hurt performance at all.
 
Last edited:
OC550 you have a great ability to make me laugh! As for your comment about performance on same HTT (NOT FSB) and multi running at 1:1 v 5:6 may I just quote Yeha on this matter. DUH!!
The whole point of this thread was when the cpu and RAM were the same does the divider have an impact (as it would on NF2's). The answer is seemingly no!
 
OC Detective said:
The whole point of this thread was when the cpu and RAM were the same does the divider have an inpact (as it would on NF2's). The answer is seemingly no!
Other than the multi issue yea. The main vice for A64s is the multiplier. I forget exactally why but there is a sizable bandwith increase with a higher multiplier. Almost identical to the hit taken by going asynch on an AXP.
 
Hmmm...I can't say as I've really noticed any change in bandwidth due to multiplyers, but then I've never really tested that either.
 
Sentential if that were true then you would see a difference in the two options I showed for the 3200 and 3500+ correct?
For 3200 that would mean 10 x 200 (1:1) being better than 8 x 250 (5:6). After all the only way you can compare them correctly is so that the cpu and RAM frequencies are the same is to have a lower HTT on the higher multi.
 
OC Detective said:
Sentential if that were true then you would see a difference in the two options I showed for the 3200 and 3500+ correct?
For 3200 that would mean 10 x 200 (1:1) being better than 8 x 250 (5:6). After all the only way you can compare them correctly is so that the cpu and RAM frequencies are the same is to have a lower HTT on the higher multi.
afaik and im being serious, might have to dig up the article it is true. Here is basically how it went.

The multi is directally related to the memory controler just like the CPU mhz is. So when the multi is cranked down you lose bandwith / efficiency. Its a given.

So yes. Bandwith of a 3500+ IS better than a 3200+. They compared all of the CPUs <believe it was anandtech / toms> and it was most noticable ebtween the 3800+ vs the 3000+.

Its not much but its still a good 200MB/s
 
I have heard that for cpu frequency bandwidth efficiency improves as it increases for any given ram frequency - is that what you mean?
 
Last edited:
1. For the same CPU frequency and memory frequency, the performance is about the same, independent of the memory_HTT_ratio or max memclock setting.

2. The memory bandwidth efficiency varies with the CPU_memory_divider,
where CPU_memory_divider = CPU_frequency / memory_frequency

E.g. based on some measurements,
memory = CPU / 9, memory_bandwidth_efficiency = 81%
memory = CPU / 10, memory_bandwidth_efficiency = 85%
memory = CPU / 11, memory_bandwidth_efficiency = 90%
etc
The higher the CPU_memory_divider, i.e. the higher the ratio of CPU_frequency to memory_frequency, the higher the memory_bandwidth_efficiency.

Since CPU_memory_divider depends on CPU_multiplier and memory_HTT_ratio and is given by

CPU_memory_divider = celing(CPU_multiplier / memory_HTT_ratio)

so it looks as if memory_bandwidth_efficiency depends on CPU_multiplier. It is only true when the memory_HTT_ratio is kept constant.

Indeed, at default memory_HTT_ratio of 1:1, the memory_bandwidth_efficiency of a 3500+ with x11 is higher than a 3200+ with x10 is higher than a 3000+ with x9.

But for a CPU with lower multiplier (e.g. 3000+), one can lower the memory_HTT_ratio to achieve the same CPU_memory_divider as a CPU with higher multiplier (e.g. 3200+, 3500+).

hitechjb1 said:
Possible explanation:

When the CPU is clocked faster (consuming faster), the memory controller is not fast enough to keep pace and provide enough data I/O with the L2 cache running in sync with the processor clock, hence resulting in more cache wait states relative to the memory controller and in turn lower efficiency (the actual bandwidth is still higher, but efficiency which is bandwidth per memory clock is reduced).

Will the revision E0 correct/improve this?
 
Last edited:
I'm sure we all agree that memory speed affects performance.

But the statements I quoted earlier were completely irrelevant to the subject, as nobody would have their memory running slower than necessary.

That "no performance hit in using different HTT / RAM-ratios on A64s" –phrase that some of us keep on stating, is absolutely not about lowering the memory speed for no good reason, despite that some people make it sound like that. The point is not the speed itself, but the non-existent latency hit in using different ratios, and there are many different ways to get a given memory speed within a certain range. This is damn obvious.

Furthermore I strongly feel that there was no reason to "correct" the first replies, as there were nothing really misleading about them. Someone seems to have no idea as to why one would want to run the RAM at a different speed than the HTT. That possibility has no negative aspects to it, hence there was no need to contradict what Sentential and I had stated initially.

It was explained to the threadstarter how he should take advantage of these possibilities. He was advised to experiment so that he could find the best settings for his rig. Nobody told him he had to reduce his memory ratio unless facing stability issues, or wanting tighter timings. The option of independently adjusting these settings without hurting memory performance is a significant advantage unique to the A64-platform, and there could be plenty of scenarios where this would be the best solution.

Take a CPU that has a max multi of 9, and RAM that maxes out at DDR500. If the HTT and RAM had to be at the same rate (if to avoid serious latency penalties seen on certain other platforms), the memory would've held back the entire OC. Luckily this is not the case, as it is possible to maintain memory speed and stability, yet at the same time keep on increasing the HTT / CPU further, without being held back by the RAM:


Example 1:
HTT | CPU speed: 250 x 9 = 2.25 GHz Memory speed: 250MHz = DDR500

Example 2:
HTT | CPU speed: 300 x 9 = 2.7 GHz Memory speed: 250MHz = DDR500


These are just examples and it may be that those exact combinations aren't available, but the point is still valid. It should be clear to everyone that rig 2 would be the fastest of the two. Its bandwidth efficiency would be at least equal to the one of example 1. There would be absolutely no performance hit due to the so-called "async-mode" itself, or anything else for that matter. For the sake of the memory, DDR400 in one given setup performs pretty much the same as DDR400 achieved via different HTT / RAM-settings, because the memory doesn't use the HTT.

While the RAM's speed and timings will obviously affect overall system performance, as long as the memory is tweaked to its full potencial there's no need to worry about the memory running at a different rate than the HTT (faster or slower). First, one can find the max HTT / CPU-speed using a very low memory speed and LDT, and then start to overclock the RAM to the max. These factors are of course advantages in the sense that they allow fine-tuning of the system.

Obviously "the faster the better" applies to RAM as well as other components, but what I'm explaining here is something entirely positive.

In some cases it could be a good idea to try different memory dividers and latencies even if the RAM runs perfectly stable when "in sync" with the HTT. Some OCers prefer high clock frequency @ loose timings, others prefer lower clock frequency @ tight timings.


It may be that this isn't terribly relevant for many overclockers ATM, but that has nothing to do with fact nor theory. This whole thing about running the memory slower than the HTT will perhaps become more relevant as the platforms mature, because more OCers may reach HTT-clocks too high for their RAM to match. Also, the higher the amount of RAM, the more stressful to the controller. Many OCers seem to have trouble with 1GB, and with 2GB the RAM would probably be very hard to OC. Again, thanks to the RAM's independence of the HTT, the CPU could be equally overclocked.
 
Miguita said:
On the platform in question though, one has the option of adjusting the memory speed and HTT independently of each other, without losing performance due to latencies, as the RAM doesn't use the HTT. I'm not saying one should lower the memory speed just for fun. But if the memory cannot run as fast as the HTT, it can be adjusted so that it still runs stable, without losing performance due to the latency-issues that are seen on platforms that have an FSB. Thus allowing higher CPU-clock (on a locked CPU). And again, there is no such thing as "running in sync" with A64s.
glock19owner said:
So Based on what you just stated...how is running at say, 320FSB/HT with 5:6 divider and 3x HT be faster/equal then running like my setup at 275 1:1 with 4x HT and tighter timings...
The "latency issue" I was referring to, has nothing to do with the timings that can be adjusted in the BIOS (like CAS, tRP, etc). Those could be the same in both scenarios. What I meant was that the memory controller on an A64 will communicate equally efficient with the RAM, no matter the HTT / MEM-ratio. Lowering the ratio is a good idea if your memory is holding back the CPU while using the 200-ratio. And besides, if your CPU / cooler makes it possible to OC further, 320 HTT would certainly be better than 275.
 
Back