Hacker News Reader

Comment by bhouston | Hacker News Reader

Fun fact: a lot of user benchmarks of Ryzen Threadripper have lower averages than standard performance. For example passmark and userbench. This is because many are running TR with just 2 dimms rather than 4 dimms. This basically reduces performance by +20%. I almost think that AMD should have just not made TR work in this case because the lower than accurate benchmarks hurt then.

I do not know exactly why it has this performance characteristic but I've witnessed it first hand. It is very ease to reproduce with Passmark on Windows.

View on HN

Who coughs up boku bucks for a Threadripper and an X399 board and then cheaps out on RAM?

It does feel weird paying more for RAM than you do for your CPU.

Well, let's see. A TR4 mobo can host 8 dimms. Say you want to max out in RAM that's a minimum of $1100. Add at least 10% if you go for ECC modules. Eventually you end up paying for RAM more than what you pay for the CPU. That's why a lot of people are conservative with RAM when they build a new rig. Somehow it doesn't feel right to pay that much money for memory. So you buy the bare minimum and then wait for prices to normalize, which seems like taking forever in this freaking industry.

RGB > ECC

Prove me wrong.

ECC has a technical benefit whereas RGB has no benefit at all unless you prefer your DIMMs to look like Christmas tree ornaments on acid.

Isn't every blinking light a gigaflop?

Yes, plus an additional 100 FPS in games.

> you prefer your DIMMs to look like Christmas tree ornaments on acid

This is a technical benefit.

The trick is not to max out in RAM, but to use all the channels. So for a processor with 4 RAM channels, instead of a single 32GiB DIMM, you should use four identical 8GiB DIMMs.

That's solid advice if you're aiming for performance. If you're aiming for performance per buck at the expense of today's performance then it's far better to spend the bare minimum on a fewer but large-ish DIMMs and max out on RAM when your wallet recovers from the initial hit.

Sometimes you don't notice the impact on performance, but you do notice doubling or quadrupling your RAM with a fraction of your initial budget.

s/boku/beaucoup/

Interesting eggcorn; where'd you learn it that way?

Noted. I have no idea. I appreciate the corn removal.

I'm totally going to say it your way from now on. But only for currency: "boku bucks", "watashi won", and "ore reals".

Is that "bokuno bucks", or did I miss the idea?

It's a reference to the (grand?)parent typo.

Belter creole?

It's listed by Wikipedia (not a native speaker and not aware of this word in English), see https://en.m.wiktionary.org/wiki/beaucoup.

You might be surprised. A lot of us early adopters had a lot of issues with getting RAM to run at its advertised clock speed. Buying expensive RAM didn't really have benefits over cheap RAM.

It took something like 4-6 months for a BIOS update to finally bring my TR's RAM to advertised clock speeds (and CPU RAID....). Though, it still had stability problems and overclocking still isn't worth the hassle. Getting the machine to a stable state was very reminiscent of 2008.

Allegedly the second gen threadrippers are going to be much more stable. heh.

I had lots of stability issues with mine on Linux kernel 4.15 which I eventually tracked down to the X399 motherboard M2 controller, not the CPU or ram as I had feared. TRIM on my M2 drive had somehow gotten enabled and was causing intermittent crashes everytime it ran as a cron job.

Which SSD was that?

It sounds like your issue was a SSD firmware bug that would have been a problem on any CPU+motherboard platform. It also sounds like your software had really unusual and probably stupid defaults.

Western Digital 1TB M2 nvme

haha they wont get a second chance with me

I cant get my 3200 mhz Ram run with my x399 Designare it runs only at 2133 mhz AMD has a big problem there my Z170 Board runs all my ram on 3200 xmp without any trouble Both kits I have are certified for the X399 board but non is working above 2133mhz Very dissapointing

It could be the overall memory bandwidth. Being able to write to various sticks at the same time could do a lot, especially in a CPU with this many cores

It's definitely a memory bandwidth issue, since ThreadRipper 1 has two dies, each of which have direct access to two DIMMs. So if you only have two DIMMs total, one die has to go through InfinityFabric to the other for all memory accesses.

Is this a Threadripper or general Ryzen thing? Got this Asus RoG GL702ZC here, didn't upgrade and it comes like that out of the box (16G, just one socket filled) afaik.

Desktop platforms (Intel 11xx, AMn) generally have two memory channels. So if you're only running with one module, you're probably leaving performance on the table, regardless of platform.

4 dimm requirement seems kind of asinine. Normal systems can boot and work fine with a single stick of ram.

Well for high performance systems requiring 4 dimms isn't bad as the tr series has 4 memory controllers.

You don't have to run with 4 DIMMs but you'll run slower if you do; it's kinda like having a one lane road vs a 4 lane road, you can put all your traffic through one lane, but it's a whole lot slower than using four lanes.

That was not the point I was arguing with. Let me qoute.

>AMD should have just not made TR work in this case

Where "this case" is <4 dimms.

Does the equivalent Intel platform suffer to the same degree of performance degradation when not using all 4 memory channels?

For the most part yes, the only sizeable difference is the caching/NUMA strategies and implementation between the two platforms (L3 <=> DRAM controller(s), and interconnects if necessary).

Reply on news.ycombinator.com