New chess rig

[cma6]

The total system consists of:

AMD TR-3970X
Asus Prime TRX40 Pro main-board.
128 GB Corsair Vengeance LPX (2x CMK64GX4M4B3200C16, Nanya chips), just added the second 64 GB this week.
1x Samsung 970 EVO 1TB and 1x Samsung 970 EVO 2TB. plus a 4TB WD Red for bulk storage.
1x Asus RTX-2060 super turbo (waiting for a better card to arrive by the end of this year).
Noctua NH-U14S TR4-SP3 CPU cooler with a second NF-A15 fan in push-pull configuration.
Seasonic Prime TX 850 power-supply (one of the best).
Fractal Design R6 computer case.

The TR-3970X is not very suitable for overclocking because it runs very hot if you do, even with water-cooling.
I use standard settings without PBO, 3700 MHz. base and 4500 MHz. PB, XMP memory setting at 3200 MHz., at these settings the processor draws > 300W with all cores loaded and already gets 80 deg. C.

Memory speed is not very important, the difference in speed I measured between 2133 MT/s and 3200 MT/s memory is at max. 5%, buying very expensive G.Skill memory with low latency to gain maybe 1 or 2% is IMHO complete nonsense.
[/quote]

Joost,
Thanks for the info. So I could buy slower RAM than 3200 MT/s for the TR-3970X?
But if you are running at 80C with Noctua NH-U14S TR4-SP3 CPU cooler, that is supposed to be too hot. I believe that AMD's recommendation for long runs is 68C.
[/quote]

What I actually meant to say is that the performance of the TR-3970X doesn't depend much upon main-memory speed because it has very large caches, so I would buy standard 3200 MT/s memory with a latency of 10 ns. If you want to have utmost performance (and you have deep pockets) you can buy very expensive G.Skill memory with lower latency, but in practice you won't notice much difference.

Underneath the Noctua NH-U14S TR3-SP4 cooler there is not much room, with this cooler you need low-profile memory, that's why I use Corsair Vengeance LPX. DDR4 DRAM is very cheap at the moment, 128 GB Corsair Vengeance LPX 3200 MT/s costs < 600 euro.

I don't like the high temperature of 80 C. too, and I don't want to go for a noisy water-cooling solution, in practice I turn PB off and let the system run at 3700 MHz. max. which keeps the temperature below 65 C. at full load, the performance of the TR-3970X at 3700 MHz. is still amazing. For an occasional tournament or when I need max. performance I turn PB on.
[/quote]

Joost,
Thanks for the clarification on RAM. "Low-profile memory, Corsair Vengeance LPX. DDR4 DRAM is very cheap at the moment, 128 GB Corsair Vengeance LPX 3200 MT/s costs < 600 euro." That is exactly what I have been looking for.

"I turn PB off". I assume that PB is some kind of O/C? So one can run CPU at stock speed of 3.7 GHz and stay below 65C? That is fantastic and good enough to stay with air cooler, either Noctua NH-U14S TR3-SP4 or Dark Rock Pro TR4.

[corres]

The total system consists of:

AMD TR-3970X
Asus Prime TRX40 Pro main-board.
128 GB Corsair Vengeance LPX (2x CMK64GX4M4B3200C16, Nanya chips), just added the second 64 GB this week.
1x Samsung 970 EVO 1TB and 1x Samsung 970 EVO 2TB. plus a 4TB WD Red for bulk storage.
1x Asus RTX-2060 super turbo (waiting for a better card to arrive by the end of this year).
Noctua NH-U14S TR4-SP3 CPU cooler with a second NF-A15 fan in push-pull configuration.
Seasonic Prime TX 850 power-supply (one of the best).
Fractal Design R6 computer case.

The TR-3970X is not very suitable for overclocking because it runs very hot if you do, even with water-cooling.
I use standard settings without PBO, 3700 MHz. base and 4500 MHz. PB, XMP memory setting at 3200 MHz., at these settings the processor draws > 300W with all cores loaded and already gets 80 deg. C.

Memory speed is not very important, the difference in speed I measured between 2133 MT/s and 3200 MT/s memory is at max. 5%, buying very expensive G.Skill memory with low latency to gain maybe 1 or 2% is IMHO complete nonsense.

Joost,
Thanks for the info. So I could buy slower RAM than 3200 MT/s for the TR-3970X?
But if you are running at 80C with Noctua NH-U14S TR4-SP3 CPU cooler, that is supposed to be too hot. I believe that AMD's recommendation for long runs is 68C.
[/quote]

What I actually meant to say is that the performance of the TR-3970X doesn't depend much upon main-memory speed because it has very large caches, so I would buy standard 3200 MT/s memory with a latency of 10 ns. If you want to have utmost performance (and you have deep pockets) you can buy very expensive G.Skill memory with lower latency, but in practice you won't notice much difference.

Underneath the Noctua NH-U14S TR3-SP4 cooler there is not much room, with this cooler you need low-profile memory, that's why I use Corsair Vengeance LPX. DDR4 DRAM is very cheap at the moment, 128 GB Corsair Vengeance LPX 3200 MT/s costs < 600 euro.

I don't like the high temperature of 80 C. too, and I don't want to go for a noisy water-cooling solution, in practice I turn PB off and let the system run at 3700 MHz. max. which keeps the temperature below 65 C. at full load, the performance of the TR-3970X at 3700 MHz. is still amazing. For an occasional tournament or when I need max. performance I turn PB on.
[/quote]
Joost,
Thanks for the clarification on RAM. "Low-profile memory, Corsair Vengeance LPX. DDR4 DRAM is very cheap at the moment, 128 GB Corsair Vengeance LPX 3200 MT/s costs < 600 euro." That is exactly what I have been looking for.
"I turn PB off". I assume that PB is some kind of O/C? So one can run CPU at stock speed of 3.7 GHz and stay below 65C? That is fantastic and good enough to stay with air cooler, either Noctua NH-U14S TR3-SP4 or Dark Rock Pro TR4.
[/quote]

If you use SMT = DISABLED (I propose it) you can get lower temperature and a more stable system.

[Joost Buijs]

If you use SMT = DISABLED (I propose it) you can get lower temperature and a more stable system.

Of course the temperature will get somewhat lower with SMT disabled, but I don't understand what it has to do with stability, my system is stable as a rock, last time I had it run for 12 days in a row at full load and there was no sign of instability at all. The only thing getting unstable was me because I could not withstand the enormous amount of heat generated by the system.

Disabling SMT has the drawback that when you run the system at full load on all cores you leave very little room for other threads to run, like IO or other operating-system threads, I clearly notice Windows gets laggy when I do, maybe Linux is better in this respect.

[corres]

If you use SMT = DISABLED (I propose it) you can get lower temperature and a more stable system.

Of course the temperature will get somewhat lower with SMT disabled, but I don't understand what it has to do with stability, my system is stable as a rock, last time I had it run for 12 days in a row at full load and there was no sign of instability at all. The only thing getting unstable was me because I could not withstand the enormous amount of heat generated by the system.

Disabling SMT has the drawback that when you run the system at full load on all cores you leave very little room for other threads to run, like IO or other operating-system threads, I clearly notice Windows gets laggy when I do, maybe Linux is better in this respect.

If you run other program in the back the repeatability of your results (analysis, test) is decreased.
The weaker stability is caused by the more heat.
The stability is not mean only that your system did not freeze under 12 days.

[cma6]

But what is PB?

[corres]

But what is PB?

PB is Power Boost. It enhances CPU frequency, heating and decreases the stability and reproducibility.
It is useful mainly for video games and if your program use only 1-2 cores.

[cma6]

Corres, thanks for explanation.

[Zenmastur]

...
Memory speed is not very important, the difference in speed I measured between 2133 MT/s and 3200 MT/s memory is at max. 5%, buying very expensive G.Skill memory with low latency to gain maybe 1 or 2% is IMHO complete nonsense.

I see statements like this a lot.

The first thing to note is the performance of the memory subsystem is, to a great extent, determined by the BIOS. Jedec/XMP profiles only store a few of the memory timings needed to initialize the memory subsystem. The rest are determined at boot time by a process called “learning” in which the BIOS tries different values in the many timing parameters visible in most BIOSs. Not all parameters are visible in the BIOS and therefore can't be changed or modified by “normal” means.

The other MAJOR issue is the quality of the MB memory trace layout. A poor design, will perform poorly and may not allow the RAM to reach it's full potential.

Since there are several different RAM manufacturers, many different memory chips with different strengths and weaknesses, many different retail memory vendors, many different motherboards manufactures, and many different BIOSs out there making a claim like “Memory speed is not very important,” is completely worthless and misleading.

If you know relatively little and just buy a bunch of “RANDOM” parts and then do NOTHING other than enable your XMP profile you can expect shit result most of the time. If you want better results, you have to put a little more time and effort in to choosing your motherboard and RAM. During setup you have to take the time to control most or ALL of the memory timings.

Most people aren't willing to do this because it takes time and can be frustrating if you don't have a clue what's going on.

On my system the difference between running the ram at 2133 and 3200Mhz was about 12% in NPS using stockfish. That was just for enabling the XMP profile. I got another 3-4% by tweaking the ram parameters even though I have crap ram sticks (crap meaning they don't overclock well). They're stable enough, just not much head room for overclocking. E.g. I couldn't get them stable at any speed faster than 3200Mhz but some of the timings at 3200 could be cut to fraction of the value the BIOS had them set at. Many of the parameters weren't even close to the minimum at which they would run stable.

Different BIOSs will perform the “learning” tasks with different results. Some are good at it and some aren't. Certain types of chips will perform better than others, not because they are “better” in general, but because the BIOS finds better parameters for those chips. The easiest way to get around this is to manually control ALL the available timing parameters. Like this:

Code: Select all

Memory Type=	Hynix CJR	128GB	8 X 8Gb	Dual Rank	DDR4-2133	C/18nm lithogrophy	Downbin	MB=B0 Rev. 8-layer	Page 1Kb	8 Dimms  Gear-down mode Disabled


								DDR4-2133	Time ns	DDR4-3200	Time ns	DDR4-3200	Time ns	DDR4-3200	Time ns
Parameter Description				Parameter	JDEC@1067		XMP@1600		Actual (tested)		BIOS set@1600
CAS latency					tCL		15		14.1	16		10.0	14		8.8	16		10.0
RAS to CAS Delay read				tRCDRD		15		14.1	18		11.3	15		9.4	18		11.3
RAS to CAS Delay write				tRCDWR		15		14.1	18		11.3	17		10.6	18		11.3
RAS Precharge time				tRP		15		14.1	18		11.3	17		10.6	19		11.9
Active to Precharge Delay Time 			tRAS		36		33.7	36		22.5	28		17.5	36		22.5
Act to Act/Refresh Delay Time 			tRC		50		46.9	54		33.8	42		26.3	56		35.0
Short Row Active to Row Active Delay 		tRRDS		4		3.7	6		3.8	4		2.5	6		3.8
Long Row Active to Row Active Delay		tRRDL		6		5.6	9		5.6	6		3.8	8		5.0
Long CAS to CAS Delay Time 			tCCDL		6		5.6	9		5.6	9		5.6	9		5.6
Four Active Windows Delay 			tFAW	      	23		21.6	36		22.5	16		10.0	34		21.3
Normal Refresh Recovery Delay Time 		tRFC1								312		195.0	560		350.0
2x mode Refresh Recovery Delay Time 		tRFC2								192		120.0	440		275.0
4x mode Refresh Recovery Delay Time 		tRFC4								132		82.5	255		159.4
							
Write to Read delay Short			tWTRS								3		1.9	4		2.5
Write to Read delay Long			tWTRL								9		5.6	12		7.5
Write recovery time (8 CLK min.)		tWR								12		7.5	12		7.5
Read to Read delay same bank group		tRDRD SCL							3		1.9	4		2.5
Write to Write delay same bank group		tWRWR SCL							3		1.9	4		2.5
CAS Write Latency				tCWL								14		8.8	16		10.0
Read to Pre-charge Time				tRTP								8		5.0	8		5.0
Read Write Command spacing			tRDWR								5		3.1	7		4.4
Write Read Command Spacing			tWRRD								1		0.6	3		1.9
Write to Write Time same chip			tWRWR SC							1		0.6	1		0.6
Write to Write Time same DIMM			tWRWR SD							3		1.9	7		4.4
Write to Write Time diff. DIMM			tWRWR DD							3		1.9	7		4.4
Read to Read time same chip			tRDRD SC							1		0.6	1		0.6
Read to Read time same DIMM			tRDRD SD							3		1.9	5		3.1
Read to Read time diff. DIMM			tRDRD DD							3		1.9	5		3.1
Clock enable time				tCKE								1		0.6	1		0.6
									

The first four columns show what the JEDEC/XMP profiles control and the latencies at the given clock speed. The second four columns show what I manually set them to and what the BIOS set them to along with latencies. In some cases the difference between what the BIOS set them at and what they will reliably run at is VASTLY different. Two likely causes are the Hynix CJR chips were relatively new and the bios is probably tuned for Samsung and Crucial chips or earlier Hynix chips.

Some of the "Tested" timing are VERY pedestrian while others are VERY fast compared to other types/brands of memory chips. I got about an extra FEW IN PERFORMANCE OVER AND ABOVE THE PERFORMANCE THE XMP PROFILE PROVIDED. THE REASON THIS IS SO LOW IS BECAUSE THESE CHIPS WON'T RUN RELIABLY FASTER THAN 3200MHZ EVEN WITH HIGH VOLTAGES. But at the time I bought them I didn't really have another choice and I was tired of waiting for better memory to become available. Samsung B-die chips are known to run very fast. A lot faster than the ones I bought.

On Ryzen chips faster memory means faster infinity clocks (up to the fabrics clock limits) which means lower memory latency. Most AB engine are latency sensitive ( not sure about NN and NNUE) so lower latency means more NPS.

So, picking random components and expecting stellar performance isn't likely to happen. But people expect it to perform better because they spent more money on the RAM. Few ever think about learning what makes a memory subsystem fast. They just spend money and hope that's good enough. This happens all the time and then you see claims that faster memory doesn't make a difference.

If you want stellar memory performance you need to know a lot and specify every component, even the BIOS revision your going to use and then spend the time to actually set the parameters and do the testing to VERIFY that the settings you choose are actually faster. If you're not going to do that then it “IS” a waste of money.

Last note: The memory times given in this post aren't likely to run on any other system unless they have the exact same components as I do, and even then they will likely need to be tweaked to match your CPU/MB/Memory Dimms due to silicon lottery effects.


Regards,

Zenmastur

[Milos]

On my system the difference between running the ram at 2133 and 3200Mhz was about 12% in NPS using stockfish. That was just for enabling the XMP profile. I got another 3-4% by tweaking the ram parameters even though I have crap ram sticks (crap meaning they don't overclock well). They're stable enough, just not much head room for overclocking. E.g. I couldn't get them stable at any speed faster than 3200Mhz but some of the timings at 3200 could be cut to fraction of the value the BIOS had them set at. Many of the parameters weren't even close to the minimum at which they would run stable.

Lol at 12+3-4%. That is more than a difference between running with minimal Hash size (e.g. 32MB) and optimal hash size (e.g. 16GB).
I bet you have some totally crappy test conditions that have nothing to do with realistic game conditions (like testing speed of TB searches/caching or testing totally ridiculous hash sizes, etc or your system has some other serious bottlenecks).

[Zenmastur]

On my system the difference between running the ram at 2133 and 3200Mhz was about 12% in NPS using stockfish. That was just for enabling the XMP profile. I got another 3-4% by tweaking the ram parameters even though I have crap ram sticks (crap meaning they don't overclock well). They're stable enough, just not much head room for overclocking. E.g. I couldn't get them stable at any speed faster than 3200Mhz but some of the timings at 3200 could be cut to fraction of the value the BIOS had them set at. Many of the parameters weren't even close to the minimum at which they would run stable.

Lol at 12+3-4%. That is more than a difference between running with minimal Hash size (e.g. 32MB) and optimal hash size (e.g. 16GB).
I bet you have some totally crappy test conditions that have nothing to do with realistic game conditions (like testing speed of TB searches/caching or testing totally ridiculous hash sizes, etc or your system has some other serious bottlenecks).

Put that crack pipe down and go outside and get a little air, you need it!