Threa Dripper RAM sanity check
 

I'm looking to build a 1950X system in a few months, and need advice on the type of RAM to get. The system will be used for more than just Prime95/Mlucas, but these will be the only programs that might benefit from ECC. Here are some options if I were to build it today, bold is the one I think is best:

[b]4x4GB 3200 CL16 non-ECC ~£170[/b]
4x4GB 3733 CL17 non-ECC ~£190
4x4GB 2133 CL15 ECC ~£200

I've seen conflicting things about the sensitivity of Threadripper to RAM speed, and can't find any reliable benchmarks that obviously apply to the use case. I'm pretty sure the 3733 CL17 should be ruled out (because of the CL17 timings, and it may be a pain to get stable), but it's there in case I'm wrong. 3200 CL16 seems to be the standard, and 2133 seems very low even with better timings, so I'm leaning towards the 3200.

a) Is the difference between the 2133 and 3200 as massive as it looks for Prime95/Mlucas?
b) Does ECC even matter that much for the current wavefront? Assuming 4 workers, 1 per CCX. I don't plan on doing 100M work.
c) Am I missing something? It's been years since my last build.

Thanks for any advice.

I'm only assuming it is not much different than desktop Ryzen, in which case 3200 is about the sweet spot. Look up exactly what results others have had with Ryzen and whatever kit you're looking at, as it can be a bit fussy.

On Prime95 performance, ram or not, Ryzen in general seems underwhelming and even when you expect it to ram limit like Intel, it still performs worse.

Zen loves fast memory. Going with 3200 over 2133 will make a massive performance difference.

I'd get the 3733 memory. It's not much more. You can probably run it at 3200 CL15 if it won't run at 3733.

3733/3200 = 1.17
17/1.17 = 14.5

But I'd still clock it as high as stable, since that's what makes the difference for Zen.

[QUOTE=mackerel;468978]I'm only assuming it is not much different than desktop Ryzen, in which case 3200 is about the sweet spot. Look up exactly what results others have had with Ryzen and whatever kit you're looking at, as it can be a bit fussy.

On Prime95 performance, ram or not, Ryzen in general seems underwhelming and even when you expect it to ram limit like Intel, it still performs worse.[/QUOTE]

3200 does look standard for Zen, it's definitely the baseline to expect to work. I will do proper compatibility checks when it comes time to buy, the specific parts I'm looking at now are probably not going to be the best then. Hopefully the mobo lineup matures a bit (and RAM doesn't go up too much).

Not to sound like an AMD fanboy, but I can't buy intel if AMD has a competitive product. I understand intel has the IPC, and wins in a few other ways, but AMD is competitive.

[QUOTE=Mark Rose;468979]Zen loves fast memory. Going with 3200 over 2133 will make a massive performance difference.

I'd get the 3733 memory. It's not much more. You can probably run it at 3200 CL15 if it won't run at 3733.

3733/3200 = 1.17
17/1.17 = 14.5

But I'd still clock it as high as stable, since that's what makes the difference for Zen.[/QUOTE]

I'm sold. I'll most likely do that, buy high speed (probably) lower timings and try tighter timings at 3200, see which one performs better if both work. The general recommendation for really high clocks seems to largely be because the infinity fabric is tied to RAM speed. That shouldn't matter for Prime95 with properly set affinities, it'll be interesting to find out.

Can we really do (17/1.17=14.5) < 15, even as an estimate? I thought the timings were more complicated than that.

Thank you both.

[QUOTE=M344587487;469118]
Not to sound like an AMD fanboy, but I can't buy intel if AMD has a competitive product. I understand intel has the IPC, and wins in a few other ways, but AMD is competitive.
[/QUOTE]
AMD has a very competitive product lineup with Ryzen/Threadripper/EPYC for most workloads. And thank god, it has taken them 5+ years to become competitive again (Bulldozer was a bust).

There are workloads in which Intel has an advantage, the ones that use AVX instructions (Prime95 and Mlucas both use those). Intel processors can do more AVX FMA instructions per clock than AMD. Intel also has AVX-512 capability in their server lineup. At some point you also hit memory bandwidth limitations and adding cores doesn't improve performance by much.

Yeah I get that, AVX-512 is a big deal for Prime95/Mlucas specifically, I meant that I can't go intel when AMD is competitive in general. Bulldozer was a bust, I went with a 2500K.

PS I don't know who edited the title, but I like it :P

edit: You've forced my hand <--- ;)

[QUOTE=M344587487;469118]I'm sold. I'll most likely do that, buy high speed (probably) lower timings and try tighter timings at 3200, see which one performs better if both work. The general recommendation for really high clocks seems to largely be because the infinity fabric is tied to RAM speed. That shouldn't matter for Prime95 with properly set affinities, it'll be interesting to find out.[/quote]

Prime95 isn't NUMA aware, so the memory allocation between dies will be suboptimal, and the infinity fabric bandwidth and latency will matter.

[quote]Can we really do (17/1.17=14.5) < 15, even as an estimate? I thought the timings were more complicated than that.[/QUOTE]

The CAS latency is set in clocks, but the physics of the amplification circuitry to get the data line ready are in nanoseconds.

So 17 clocks of 1866 MHz (DDR4 3733) is 1/1866000000 * 17 = 9.11 ns.

So how many clocks of 1600 MHz is 9.11 ns? 14.6.

Now there may be other latencies involved, so it's no guarantee.

[url]https://en.wikipedia.org/wiki/CAS_latency[/url]

[QUOTE=Mark Rose;469135]Prime95 isn't NUMA aware, so the memory allocation between dies will be suboptimal, and the infinity fabric bandwidth and latency will matter.[/QUOTE]

Prime95 isn't NUMA-aware, but does it need to be? When I say I want to set a worker per CCX, I mean run 4 instances of Prime95/Mlucas, with their affinities set to each CCX as required. It should then be down to the OS to allocate memory smartly. That's how I understood it to work, please let me know if this is wrong.

edit: I've found this and am researching elsewhere: [url]http://www.mersenneforum.org/showthread.php?t=22571[/url]

[QUOTE=Mark Rose;469135]The CAS latency is set in clocks, but the physics of the amplification circuitry to get the data line ready are in nanoseconds.

So 17 clocks of 1866 MHz (DDR4 3733) is 1/1866000000 * 17 = 9.11 ns.

So how many clocks of 1600 MHz is 9.11 ns? 14.6.

Now there may be other latencies involved, so it's no guarantee.

[url]https://en.wikipedia.org/wiki/CAS_latency[/url][/QUOTE]
It's the other memory timings that I thought may make calculations a pain, but I've never looked at them in detail as they seemed a little too "inside baseball" to worry about: [url]https://en.wikipedia.org/wiki/Memory_timings[/url]

[QUOTE=M344587487;468956][b]4x4GB 3200 CL16 non-ECC ~£170[/b]
4x4GB 3733 CL17 non-ECC ~£190
4x4GB 2133 CL15 ECC ~£200
[/QUOTE]

Can you really use ECC RAM with these AMD processors? With Intel processors you need Xeon models to use ECC.

[QUOTE=ATH;469143]Can you really use ECC RAM with these AMD processors? With Intel processors you need Xeon models to use ECC.[/QUOTE]

Yep! ECC across the full Threadripper lineup.

For Ryzen, it depends on motherboard support.

All Zen CPUs support ECC.

[QUOTE=ATH;469143]Can you really use ECC RAM with these AMD processors? With Intel processors you need Xeon models to use ECC.[/QUOTE]

Yes, definitely with Threadripper processors on AMDs HEDT platform (TR4 socket), and EPYC (SP3 socket). You can also supposedly use ECC RAM with Ryzen processors (AM4 socket), but that's more hit and miss as it's up to the mobo to support it, many don't as AMD haven't certified Ryzen as supporting it as Ryzen is their desktop platform.

[url]http://www.hardwarecanucks.com/forum/hardware-canucks-reviews/75030-ecc-memory-amds-ryzen-deep-dive.html[/url]

Honestly not a fanboy, just a fan :P

edit: I need to work on my response speeds ;)