AlexZ wrote on Yesterday, 07:56:

Just to put things into context, here are Aida64 results for my s754 Athlon 64 3400+ and AM2+ Phenom II x4 955. Athlon 64 3400+ […]
Show full quote

Just to put things into context, here are Aida64 results for my s754 Athlon 64 3400+ and AM2+ Phenom II x4 955. Athlon 64 3400+ memory timings haven't been tuned and it runs on auto and 2T. Phenom II x4 955 doesn't work with 1T command rate.

What is interesting is how slow L2 cache was in Clawhammer. DDR2 gives us bandwidth not far away from that L2 cache except for latency.

L2 cache in Phenom II is really fast compared to s754 and s939. L3 cache is faster than L2 cache in Clawhammer.

It would be great if Archer57 could post his results for Windsor at 3Ghz.

DDR timings tuning guide https://www.elektroda.com/rtvforum/topic1203122.html

Cache and memory:

And also run this, just for fun, as it contains CPU benchmarks too:

Curious how some results are higher for GPU on S939, i have no explanation why.

Also thought this would be interesting:

Slow L2 on 754? 😁

When comparing results of CPUs with different core counts the L1 and L2 cache bandwidth has to be divided by the number of cores. I tested on my Phenom II X6 that when running it as X4 vs X6 cache bandwidth goes up by 50%. L3 cache bandwidth scales much less with additional cores as it is shared.

L2 cache on s754 is faster than on Athlon XP, but it also has much worse latency.

Memory read performance on Windsor looks weird, it should be about the same as write performance. Unfortunately I didn't run aida64 on my testing configurations so we cannot verify whether it is normal.

On Phenom II memory read is slightly faster than write. I verified in on both X4 and X6 CPUs. Interestingly, with DDR3 the discrepancy is even larger than DDR2.

These benchmarks confirm one needs to have a fast AM2 CPU with DDR2 800 to benefit from DDR2 memory.

AlexZ wrote on Yesterday, 11:06:

Memory read performance on Windsor looks weird, it should be about the same as write performance. Unfortunately I didn't run aida64 on my testing configurations so we cannot verify whether it is normal.

Yep, but i ran this benchmark in different variations for like 20 times at this point, the results are repeatable and still there with 12 CPU multiplier which makes memory work at 800 and not weird 753.

What's curious on S939 increasing CR to 2T significantly affects write speed in this benchmark. What you are seeing on S754 is probably a good example of that - if you were to set it to 1T write and read would probably become ~the same as they are on S939 in my benchmarks. At 2T write results are somwere below 5000 on this system.

Not sure if this is a quirk of specific benchmark or a quirk of memory controller and if it represents actual performance difference.

Here's what got on different board :

AlexZ wrote on Yesterday, 11:06:

These benchmarks confirm one needs to have a fast AM2 CPU with DDR2 800 to benefit from DDR2 memory.

it was always the same, DDR3 1333 its not much faster than DDR2 800, DDR4 3000 than DDR3 1866, etc

the tradeoff is that you need higher latencies to gain that much higher frecuencies, thus, the gain is minimal overall. The main beneficed in this is the IGPU.

ontopic: nah, cant feel the retro vibe yet. Many people using it as an everyday computer. That might change with the end of 10 since no AM2 processor supports 11 24h2.

I did some experiments with my s754 Athlon 64 DDR400 memory timings. There is no Command Rate setting in BIOS. I always changed one value and ran memtest86 to avoid corrupting Windows XP. I once accidently booted Windows XP with tRCD=2 but fortunately it rebooted even before Windows XP logo was shown.

CAS Latency (tCL) 2. On auto 2.5 was used which works the same as 2 and results in 66.8ns latency we saw on screenshot. With CL 3 I got 71ns memory latency.
RAS to CAS delay (tRCD) 3T. Not stable with lower values.
Row precharge time (tRP) 2T
Row to Row delay (Trrd) 2T
Min RAS active time (tRAS) 6T tRAS = tCL + tRCD + tRP (+/- 1) - various formulas exist on the internet
Row cycle time (tRC) 8T tRC >= tRAS + tRP
Read-to-Write time (Trwt) 2T
Write Recovery Time (Twr) 2 bus clocks
Write-to-Read Delay (Twtr) 2 bus clocks
Read Preamble value 5ns. See https://archive.techarp.com/showFreeBOG60c6.h … ang=0&bogno=394
Async Latency value 6ns. See https://archive.techarp.com/showFreeBOG7eac.h … ang=0&bogno=399

Decreasing "Async Latency value " allowed memory latency to drop further to 61ns. Memory read and write performance remained exactly the same. Without decreasing "Async Latency value", there would have been no measurable benefit to memory timings tuning. 2 DIMM slots closest to CPU are occupied therefore settings for 2 DIMM slots were used.

It may be possible that the lower write performance is caused by Command Rate 2. From what I have read with 2 modules per memory controller CR1 rarely works. This is probably why Archer57 read that people were considering 2x 512MB for dual channel for Vista - to be able to use CR1. Without any memory timings tuning, 71ns memory latency is normal on Athlon 64.

I also somehow gained about 1000MB/s in L2 cache read speed. Remaining values are basically the same. No voltage adjustment was done.

I will do further experiments on my AM2+ build.

I also wonder if there exists a stronger steel backplate for socket AM2. I have the standard steel one and it bent about 1mm after fully tightening Scythe Ninja 5 (see https://www.techpowerup.com/review/scythe-ninja-5/ ). I tightened it fully as otherwise applying equal pressure can be difficult. Many screwdriver revolutions are needed. Plastic backplates are only suitable for stock AMD coolers. I expected the steel backplate to be stronger. I would prefer not to flex the motherboard.

It looks like this

I'm using only one fan on Scythe Ninja 5 as there is no room for another. 2x top 14cm fans and 1x 12cm fan next to IO panel keep it really cool.

Living wrote on Yesterday, 19:25:

ontopic: nah, cant feel the retro vibe yet. Many people using it as an everyday computer. That might change with the end of 10 since no AM2 processor supports 11 24h2.

Same issue with lga 775. We are early explorers as the parts are quite cheap now. I would recommend to everyone to get stocked on AM2 now unless they prefer lga 775. As demonstrated, AM2 can also simulate 939 quite accurately in case one can't get a 939 X2 2.4Ghz CPU. 939 has a major problem with highend X2 CPU availability. To do a 939 build one has to get the CPU first. AM2 can struggle in Windows Vista, but handles Windows XP fine.

AlexZ wrote on Yesterday, 20:18:

I did some experiments with my s754 Athlon 64 DDR400 memory timings. There is no Command Rate setting in BIOS. I always changed […]
Show full quote

I did some experiments with my s754 Athlon 64 DDR400 memory timings. There is no Command Rate setting in BIOS. I always changed one value and ran memtest86 to avoid corrupting Windows XP. I once accidently booted Windows XP with tRCD=2 but fortunately it rebooted even before Windows XP logo was shown.

CAS Latency (tCL) 2. On auto 2.5 was used which works the same as 2 and results in 66.8ns latency we saw on screenshot. With CL 3 I got 71ns memory latency.
RAS to CAS delay (tRCD) 3T. Not stable with lower values.
Row precharge time (tRP) 2T
Row to Row delay (Trrd) 2T
Min RAS active time (tRAS) 6T tRAS = tCL + tRCD + tRP (+/- 1) - various formulas exist on the internet
Row cycle time (tRC) 8T tRC >= tRAS + tRP
Read-to-Write time (Trwt) 2T
Write Recovery Time (Twr) 2 bus clocks
Write-to-Read Delay (Twtr) 2 bus clocks
Read Preamble value 5ns. See https://archive.techarp.com/showFreeBOG60c6.h … ang=0&bogno=394
Async Latency value 6ns. See https://archive.techarp.com/showFreeBOG7eac.h … ang=0&bogno=399

Decreasing "Async Latency value " allowed memory latency to drop further to 61ns. Memory read and write performance remained exactly the same. Without decreasing "Async Latency value", there would have been no measurable benefit to memory timings tuning. 2 DIMM slots closest to CPU are occupied therefore settings for 2 DIMM slots were used.

It may be possible that the lower write performance is caused by Command Rate 2. From what I have read with 2 modules per memory controller CR1 rarely works. This is probably why Archer57 read that people were considering 2x 512MB for dual channel for Vista - to be able to use CR1. Without any memory timings tuning, 71ns memory latency is normal on Athlon 64.

I also somehow gained about 1000MB/s in L2 cache read speed. Remaining values are basically the same. No voltage adjustment was done.

I will do further experiments on my AM2+ build.

This may sound stupid but i'd mention it anyway - you are aware of "ctrl+f1", right? Also may have to dig through everything in bios, sometimes CR is in obscure places like CPU settings instead of memory settings and can be named a bit weirdly too.

Otherwise yeah, this is the reason i did not bother messing with memory too much. Likely need fancy "DDR500" or something + more voltage, i've seen people mention 2.9v and such.

I like to fool around with OC, but for this old hardware... at least for me running it mostly stock feels like a good compromise. At least for stuff that's rare/expensive.

Also situation with CR means that 754 is practically limited to 1GB, 939 - 2GB without losing some performance...

AlexZ wrote on Yesterday, 20:18:

Same issue with lga 775. We are early explorers as the parts are quite cheap now. I would recommend to everyone to get stocked on AM2 now unless they prefer lga 775. As demonstrated, AM2 can also simulate 939 quite accurately in case one can't get a 939 X2 2.4Ghz CPU. 939 has a major problem with highend X2 CPU availability. To do a 939 build one has to get the CPU first. AM2 can struggle in Windows Vista, but handles Windows XP fine.

Dual core S939 is, IMO, strictly a collector's thing. For those who specifically want it for some reason. From practical point of view dual core CPUs are too expensive. I've seen 3800+ for relatively reasonable amount of money, but that's low end and i am not sure it will be good for practical use.

Single core, on the other hand, may make sense. Venice is relatively common and given whole line is pretty much "cost-reduced" and does not have high-end options - pretty cheap. May be a viable alternative to other late single core stuff, for example S754, with all the advantages like dual channel memory and pci-e/agp choice.

AM2 is really, really easy in comparison. CPUs are plentiful and cheap, there is huge choice of motherboards, etc. So for someone who wants K8 CPU and does not care for specific generations/sockets - it makes much more sense. IMO instead of both 754 and 939.

Archer57 wrote on Yesterday, 21:01:

This may sound stupid but i'd mention it anyway - you are aware of "ctrl+f1", right? Also may have to dig through everything in bios, sometimes CR is in obscure places like CPU settings instead of memory settings and can be named a bit weirdly too.

Actually I didn't know about "ctrl+f1". I had Epox motherboard with Athlon XP that didn't need it and then switched to Intel and didn't OC anymore.

On Gigabyte GA-K8NE it reveals Advanced Chipset Features menu and new items in other menus. Notably:
- HDD Smart Capability - BIOS always said it is off but there was no way to enable it
- PCI Latency Timer in PnP/PCI Configurations
- multiple Spread Spectrum settings in MB Intelligent Tweaker
- 1T/2T DRAM Timing in Advanced Chipset Features - this will be the Command Rate

Archer57 wrote on Yesterday, 21:01:

Otherwise yeah, this is the reason i did not bother messing with memory too much. Likely need fancy "DDR500" or something + more voltage, i've seen people mention 2.9v and such.

I like to fool around with OC, but for this old hardware... at least for me running it mostly stock feels like a good compromise. At least for stuff that's rare/expensive.

Also situation with CR means that 754 is practically limited to 1GB, 939 - 2GB without losing some performance...

I try to avoid more voltage. s754 will not be getting more and AM2+ will be tested with lower DRAM voltage than 2V. DRAM will not die from tighter timings.

Yeah CR1 and dual channel memory controller is a major advantage of 939 over 754 for single core.

AlexZ wrote on Yesterday, 21:36:

Actually I didn't know about "ctrl+f1". I had Epox motherboard with Athlon XP that didn't need it and then switched to Intel and […]
Show full quote

Actually I didn't know about "ctrl+f1". I had Epox motherboard with Athlon XP that didn't need it and then switched to Intel and didn't OC anymore.

On Gigabyte GA-K8NE it reveals Advanced Chipset Features menu and new items in other menus. Notably:
- HDD Smart Capability - BIOS always said it is off but there was no way to enable it
- PCI Latency Timer in PnP/PCI Configurations
- multiple Spread Spectrum settings in MB Intelligent Tweaker
- 1T/2T DRAM Timing in Advanced Chipset Features - this will be the Command Rate

Nice. This usually exists on gigabyte boards, sometimes on ones from other manufacturers like ECS. Often makes all the difference between mediocre and great set of options.

AlexZ wrote on Yesterday, 21:36:

I try to avoid more voltage. s754 will not be getting more and AM2+ will be tested with lower DRAM voltage than 2V. DRAM will not die from tighter timings.

Yeah CR1 and dual channel memory controller is a major advantage of 939 over 754 for single core.

Yeah, tighter timings or slight overclock without changing voltage is safe. The issue is - it is rarely if ever possible to significantly reduce timings without either having binned memory or higher voltage. And i hate memory errors. Setting it right on the edge will inevitably result in some very infrequent errors. So i tend to leave some headroom, always did with memory.

And killing memory is not what worries me, DDR1 is not rare or expensive. Killing the memory controller, which will be getting that increased voltage and is inside the CPU is a bigger issue...

Also is 1GB vs 2GB very significant for XP and stuff relevant for single core system? I've used 2GB on all the systems i fooled around with and i've never seen even 1GB used - even running crysis + MSI afterburner (which wastes like 300MB). May be reasonable to have 1GB on 754 with CR1.

Manufacturers cannot test every module for 10 hours with memtest86 like I did with DDR400 on s754 so they need to use some headroom. They do not want to risk errors in case the modules run in a hot region like Mexico/India. I also plan running prime95 for 10 hours as memtest86 may not detect all problems.

Aida64 memory benchmark is not a great way to test effects of memory timings as there is negligible impact on burst/vector loads. The only value that matters for games (they do not benefit much from SSE memory bandwidth) is latency measured by Aida64, which is our workload on these systems.

My AM2+ Gigabyte GA-MA770-UD3 board also reveals more settings with ctrl+F1, but nothing significant, mostly obscure settings you never heard of. It doesn't expose "Async Latency value" setting which lowered memory latency further on s754. This means we pretty much cannot get below 66ns easily on AM2+. tCL, tRCD, tRP are already low so no major gains can be made. I checked manual of Gigabyte GA-K8N-SLI (s939) and it does have "Async Latency value" setting.

I prefer 2GB RAM for Windows XP so that I can turn off swap file. The only game that comes close to consuming it is IL-2 Sturmovik 1946. Similarly, on AM2+ I prefer 8GB RAM (4x2GB) with CR2 over 4GB with CR1 as it can also cover Windows 7 era. I do not want to use extreme voltage or buy expensive modules to make CR1 work. Someone building AM2 just for Windows XP will be fine with 4GB RAM and CR1. This was also my recommendation for Windsor.

AM2+ Phenom II 3.2Ghz turned out to tolerate 1.9V for DDR2 800 and tight timings, which is also the value used by the motherboard when voltage is set to Normal. Standard DDR2 voltage is 1.8V. I'm happy with 1.9V in the long term.

It is stable with 4-4-4-11 timings (CL-tRCD-tRP-tRAS). tRC=15. There are various other timings that were set to either 2T or 4T. Command Rate 2T. I didn't try more aggresive timings until things would start to fail.

We didn't gain any memory bandwidth as expected but saw a small drop in latency about 2ns.

To improve write performance we would probably have to use 2 memory modules only to allow Command Rate 1T.

Athlon 64 3400+ (2.2Ghz) gets about half L1/L2 cache performance of Athlon 64 X2 4800+ (s939, 2.4Ghz) which is expected as Aida64 score depends on core count. If it was downclocked to 2.2Ghz then the score should be 2x higher as these CPUs are clearly very similar.

RoberMC wrote on 2020-02-13, 19:13:

not so many choices

what?

there are tons of processors for am2, almost as 775 (250 different processors counting am2+). The fact that you can go from a Sempron 3000+ to a Phenom II x6 1100T is wild to me.

and besides, motherboards for AM2 are much more interesting than 775, IMO

The fact that you can go from a Sempron 3000+ to a Phenom II x6 1100T is wild to me.

I don't recall any pure AM2 board supporting Phenom II x6, at least officially. And there's not many of those that support Phenom II x4 to begin with.

The Serpent Rider wrote on Today, 16:19:

The fact that you can go from a Sempron 3000+ to a Phenom II x6 1100T is wild to me.

I don't recall any pure AM2 board supporting Phenom II x6, at least officially. And there's not many of those that support Phenom II x4 to begin with.

you need a 125watts am2+ motherboard for that, and there are only amd chipsets for that.

AlexZ wrote on Today, 07:59:

I prefer 2GB RAM for Windows XP so that I can turn off swap file. The only game that comes close to consuming it is IL-2 Sturmovik 1946. Similarly, on AM2+ I prefer 8GB RAM (4x2GB) with CR2 over 4GB with CR1 as it can also cover Windows 7 era. I do not want to use extreme voltage or buy expensive modules to make CR1 work. Someone building AM2 just for Windows XP will be fine with 4GB RAM and CR1. This was also my recommendation for Windsor.

We all have our preferences, nothing wrong with that. I like as much RAM as possible too. But...

Disabling swap is a bad idea. Regardless of amount of RAM, simply because of how memory management works.

We look at this numbers nowadays and they seem small, but back then 256MB was decent for XP, 512MB was very good, more than that - an overkill for most use cases. All this systems, like S462 with 2GB, are simply a product of RAM being cheap to a point of being practically free. As long as it does not affect anything negatively i do it too, but as soon as there is a measurable performance impact... well, that's the reason i left S939 with 2GB.

My opinion - there will be no performance difference for games between 1GB and 2GB on XP, and difference between 512MB and 1GB will be limited to very late stuff which belongs to vista/7 more than XP anyway.

8GB on AM2? If used for modern tasks it makes a lot of sense. For vista and period correct games? It will not be utilized.