It Looks Like AMD's Linux Developers Have Begun Work On Zen 2 / EPYC 2 "Rome" Support

[kgardas]

I do have some extra DDR4 ECC RDIMMs arriving today, before putting them in POWER9 server, will see if any of them work in any of my Threadripper boxes (haven't checked on compatibility from that angle in a while; though I guess it probably only works for ECC UDIMMs).

Sorry for hijacking AMD thread, but you have already received Talos? If so, what's your timeframe for first benchmark run? Also please do not forget LLVM 6 compilation please. Thanks! Karel

[Michael]

Sorry for hijacking AMD thread, but you have already received Talos? If so, what's your timeframe for first benchmark run? Also please do not forget LLVM 6 compilation please. Thanks! Karel

I heard this morning that there was a delay in the chassis, but that they hope to have the Talos II server sent out to me on Monday... so hopefully by end of next week I'll finally have it. My DDR4 server memory for it arrived yesterday, so as soon as I have the server will have initial tests out within a day or two likely.

[torsionbar28]

so - there are plenty of articles that show performance scaling with DDR frequency. You can get > 3200mhz RAM with fairly tight timings (c14 for a 16gb 3200 kit is something I saw today).

The fastest ECC RAM I have seen goes at 2666mhz.

from memory there is a ~5-10% performance difference between these speeds.

The tightest timing for the 2666mhz ECC ram was C17.

So yes, there will be a performance difference, it won't be huge unless you start throwing massive amounts of money at >4000mhz RAM.

Comparing different speeds is apples and oranges. Of course 2666 is going to be faster than 2400. And 2400 is faster than 2133. ECC vs. non-ECC doesn't matter when talking different speeds.

As for the current market conditions, there is a DDR4 shortage right now, which is completely unrelated to ECC vs non-ECC performance. The new Threadripper supports DDR4-2933. The fact that DDR4-2933 ECC modules are not available at retail today has nothing to do with the benchmark performance of the technology.

The fact remains, there is no measurable performance difference between ECC and non-ECC memory modules. Don't take my word for it, it has been benchmarked many times already.

From the article's conclusion: "In the real-world benchmarks (Audio encoding, Photoshop CC and Premiere Pro CC), there was little if any drop in performance with either ECC or REG ECC memory. In fact, they actually benchmarked better than standard memory in many cases."

[boxie]

Comparing different speeds is apples and oranges. Of course 2666 is going to be faster than 2400. And 2400 is faster than 2133. ECC vs. non-ECC doesn't matter when talking different speeds.

As for the current market conditions, there is a DDR4 shortage right now, which is completely unrelated to ECC vs non-ECC performance. The new Threadripper supports DDR4-2933. The fact that DDR4-2933 ECC modules are not available at retail today has nothing to do with the benchmark performance of the technology.

The fact remains, there is no measurable performance difference between ECC and non-ECC memory modules. Don't take my word for it, it has been benchmarked many times already.

From the article's conclusion: "In the real-world benchmarks (Audio encoding, Photoshop CC and Premiere Pro CC), there was little if any drop in performance with either ECC or REG ECC memory. In fact, they actually benchmarked better than standard memory in many cases."

Except you are ignoring the biggest performance impact - ECC memory just isn't available is speeds that are as fast as non ECC memory.

Sure if you are comparing just thing "ECC vs NonECC at the same timings" you see 0 impact, and if you compare MHz vs default timings on memory you see very minimal performance differences. sure. I see what you are getting at.

my point - you can get a speed boost my using higher frequency ram. (it's sort of a no brainer obvious point).