Zen 2 for Distributed Computing: Any interest?

[biodoc]

For those of us who run our computers 24/7 to support science, the Zen 2 processors are intriguing based on power savings along with a potential for a 2x increase in performance on projects where Floating Point calculations are important. Up to a 15% increase in IPC, along with substantial increases in level 3 cache sizes, may have some impact on many DC projects. I've attached thumbnails of specs from AMD on the 2700X (comparison purpose) and the 3700X, 3800X and 3900X. I'm considering replacing one of my ivy bridge xeons with a 3900X. Anyone else thinking about Zen 2 for DC?

 

[crashtech]

Lots of interest, can't wait to see how Zen2 performs. Although, I've been going to high core count boxen these days, so I might have to content myself with Haswell/Broadwell-EP Xeons until higher core count Zen2 with quad channel becomes available.

 

[biodoc]

Yes, the dual-channel RAM in the 3900X could very well limit performance on some projects. I'm assuming AMD attempted to compensate for loss of memory throughput for a 24 thread processor by increasing the level 2 cache from 4->6 MB and level 3 cache from 16 MB in the 2700X or 32 MB in the 3700X/3800X to 64 MB in the 3900X. Waiting for the next gen threadrippers is certainly worth considering.

 

[crashtech]

Well I need to start making the expected lifetime electricity cost part of the calculation, since these rigs spend a great deal of their lives at full throttle. My recent purchases of retired E5-2678v3s for $170 each might not look so economical if power was figured in. They are hungry beasts. I'm sure Zen2 will be the efficiency king, so it may actually be less expensive over time to deploy 2 12 core Zen2 rigs than 1 24 core (2P) Xeon rig. Guess there's going to be a lot of searching for data to get the math right once they are available.

 

[Howdy]

Well, I had looked at them and have been following along . I do have a AM4 socket I could swap it in to with a BIOS update. It's a matter of this particular board is only 7 mos old (I think) and I just purchased another CPU recently. I should have waited on either to see what came out performance wise. My question is, on PG projects or where the AMD CPUs do not perform well- will it perform at or close to Xeon/ i7-9? Or is it too early to say? You guys know more about the finer details than I do.

 

[crashtech]

The expectation is that AVX2 will be at least on par with the latest Intel offerings. That's the only area that I've seen Zen struggle (PG uses AVX2 I think). I have a 2P Ivy and a 2P Sandy system that might just get retired sooner in favor of Zen2 if the math works out.

 

[Howdy]

The expectation is that AVX2 will be at least on par with the latest Intel offerings. That's the only area that I've seen Zen struggle (PG uses AVX2 I think). I have a 2P Ivy and a 2P Sandy system that might just get retired sooner in favor of Zen2 if the math works out.

So what your saying is:
I should wait until you have purchased one and given it thorough testing on AVX2 projects before I kick my 7mo old to the street!!

 

[crashtech]

Ha, don't hold your breath too long! My money situation is like a roller coaster, I'm probably doing some wishful thinking.

 

[Howdy]

LOL, I have to get the crunching slush fund rebuilt anyway. I have time!!!

 

[Orange Kid]

3900x.....wwwwwwweeeeeeeeeeeeee!!!!!!!!!!!!!!

I'll be stepping up my upgrade map sooner than thought, just to avoid possible higher costs down the road. (politics)

 

[Assimilator1]

For those of us who run our computers 24/7 to support science, the Zen 2 processors are intriguing based on power savings along with a potential for a 2x increase in performance on projects where Floating Point calculations are important. Up to a 15% increase in IPC, along with substantial increases in level 3 cache sizes, may have some impact on many DC projects. I've attached thumbnails of specs from AMD on the 2700X (comparison purpose) and the 3700X, 3800X and 3900X. I'm considering replacing one of my ivy bridge xeons with a 3900X. Anyone else thinking about Zen 2 for DC?

View attachment 7162 View attachment 7163

Yes!
Which projects would be limited by dual channel RAM?
How do the current Ryzens 2x00X compare to current Intel offerings, and how would a 2600X compare to my i7 4930k for DC? (which draws 240-250w @ 4.1 GHz). (I used 'bench' for non DC progs).
Zen2 is going to have double the FPU of Zen+s??

 

[crashtech]

From wikichip:

The floating-point unit underwent major modifications in Zen 2. In Zen, AVX2 256 bit single and double precision vector floating-point data types were supported through the use of two 128 bit micro-ops per instruction. Likewise, the floating-point load and store operations were 128 bits wide. In Zen 2, the datapath and the execution unitswere widened to 256 bits, doubling the vector throughput of the core.

With two 256-bit FMAs, Zen 2 is capable of 16 FLOPs/cycle.

 

[Howdy]

From wikichip:

The floating-point unit underwent major modifications in Zen 2. In Zen, AVX2 256 bit single and double precision vector floating-point data types were supported through the use of two 128 bit micro-ops per instruction. Likewise, the floating-point load and store operations were 128 bits wide. In Zen 2, the datapath and the execution unitswere widened to 256 bits, doubling the vector throughput of the core.

With two 256-bit FMAs, Zen 2 is capable of 16 FLOPs/cycle.

I'd have better luck trying to read a law book.

 

[crashtech]

Oh, essentially it says that Zen 2's AVX2 performance ought to be double Zen's which puts it very close to being even with current Intel offerings.

 

[Howdy]

Glad I selected the X470 MOBO for my CPU, Zen 2 will be a good fit. I found this too about the X570 boards.

 

[lane42]

Found this on M/B forums,

Okay, here's what everyone needs to keep in mind.

Set aside the 3900x for a moment.

3800x, 3700x, and below will have basically the same VRM requirements as older Zen products. What I mean to say is . . . core count has a lot to do with how the total power demand for the chip breaks down between voltage and current. More cores = higher current draw, and drawing more current puts more stress on the VRMs. So if you are on 8c today, then the stress to the VRMs within a given power envelope remains the same as it was two years ago. If you are trying to push a 105w 8c Matisse, the current demands are about the same as a 105w Summit Ridge. At least until we bring in AVX2 but let's not get too far out there.

More current draw tends to heat up VRMs. VRMs have a per-phase current limit in their specs. Sometimes mobo manufactuerers are nice enough to tell you that limit. You don't want your VRMs getting close to their current limits, because that makes them run hot, and if they get over 90-100C or higher they can cook. That shortens the life of the board. See overclocked R7 1700x on cheap B350 boards, etc.

The best VRM layout on X370 was the X370 Taichi, with a . . .12+4 setup (6+2 with doubler). They're 60-amp rated. You won't find many VRM configs better with X570 except maybe on one of the MSI boards, I don't know. They might go crazy with more than 16 phases. It looks like the top X570 Asus boards have a similar config to the Taichi (which is far better than what the Crosshair VI Hero had in 2017). X370 Taichi can push over 300W out of the socket. I would imagine that the X570 boards with good 12+4 configs will do about the same . . . maybe more.

So to summarize, if you have an 8c Matisse, you want about the same VRM config that you used on Summit Ridge or Pinnacle Ridge. And looking at X570, it looks to me like anything 8+2 will suffice for regular operation or maybe even mild overclocks. 12+2 will be better because 4 more phases for the CPU. 12+4 gives you an extra two phases for SoC functions/RAM/memory controller. Er, I think.

I would not go with anything less than 8+2 config for an 8c. Fortunately, everything X570 seems to have 8+2 as a minimum (at least in Asus' lineup; you'll have to examine other OEMs to see if they do the same).

3900x is going to tilt its current draw by up to +50% compared to the 3800x within a given power envelope. That's gonna heat up those 8+2 configs quickly. You might be able to get away with it if you stay within the 105w TDP strictly, but any kind of PBO/XFR/overclocking and watch out. You need to spread that current draw out among as many phases as possible. 12+2 or 12+4 configs are preferred for this chip. Eventual 16c Matisse will be even more hardcore.

Also, not all VRMs are made the same. AM4 boards of the past have not only had some low-phase configs like 4+3 and 4+2, but they've also had VRMs with current limits lower than 60 amps. That's bad juju for high-core-count CPUs. Look to someone like Buildzoid (or similar) to give you the skinny on what kind of VRMs are in use on any given board. If you are seeing current limit per phase below 60 amps, you might want to steer clear of that board.

 

[TennesseeTony]

The 3700X 8/16 core at 65watts is the only chip that looks interesting to me for 24/7 operation, but not on a 570 chipset (power efficiency is the goal). The 3900X is only appealing due to the doubling of the cache.

I will have to wait. To equal ONE 32/64 ThreadRipper, you would need 4 3700X systems, or 3 3900X's. Waiting for the newest ThreadRipper would be cheaper in system cost, and in power usage.

But that is for the nutz among us. For those with more sense than me, with one or maybe 2-3 systems max, and a firm resolve not to go nutz with DC, the 3900X is hard to resist!

 

[biodoc]

I see your point @lane42 but my philosophy as a DC guy is to keep the vcore voltage low enough to maintain stability at or near the base clock of the processor. With the Zen processors, there's very little benefit to be gained by over volting the chip for a very modest gain in clock speed. It's a waste of electricity. I suspect the 3700X will be the best choice for an 8c Zen processor for DC. At 65 Watt TDP, I would predict a full load clock speed to be at the base clock of 3.6 GHz. The 3800X is the same chip but probably severely over volted to generate a clock speed somewhere north of 4 GHz. The latter is perfect for gamers and enthusiasts into benchmarking.

 

[biodoc]

I've been tracking total power draw at the wall on one of my 2700X computers @3.9 GHz (1.21 vcore) at full load for various DC projects. If I could drop in a 3700X and reduce power consumption by 40%, that would be great. Computational capability could be better too (FP and IPC).

SRBase (AVX/AVX2): 172 watts
amicable numbers (AVX?): 165 watts
DHEP: 155 watts
Rosetta: 125 watts

 

[TennesseeTony]

With a TDP of 105 watts, those figures are surprising! Are you at least using a 80+ Bronze PSU? edit: I forgot to add an idle GPU into my calculations, so the numbers don't look so bad now.

 

[biodoc]

My power supply is a corsair RM750X which is 80+ Gold. My idle GTX 980 draws 14 watts according to nvidia-smi. Any app with FP math and power usage goes up dramatically. Rosetta is good science and draws less power.

 

[Howdy]

Thoughts after the E3 reveal? $749 for the 3950x
I might be back to the days of old- Running AMD CPUs and nVidia GPUs in my system.

 

[biodoc]

The 3950X looks interesting. When compared to the 3900X, 4 cores are added, the base clock is reduced by 300 MHz, the level 2 cache is increase to 8 MB vs. 6 MB, the level 3 cache on both is 64 MB and both have a TDP of 105 watts. Before I invest too heavily into Zen 2 hardware, think I'll replace one of my 2700X's with a 3700X and do some careful comparisons of my other 2700X with the 3700X on some DC projects. We'd should think about which projects to use to compare the 2 processors. I guess an llr primegrid app would be one and maybe TNGrid too. PPD and PPD/watt would be the parameters to follow I think.

Specs comparison of 3900X and 3950X from AMD.

 

[crashtech]

How low can you go on a motherboard for the 3950X, do you think? Probably not wise to go cheap, one with robust power delivery is needed. I'm trying to wrap my head around how much electricity would have to be saved by a new Zen2 setup compared to my 2P Haswell rigs. I'm guessing the throughput of those would be roughly similar to a single 3950X. If so, it would have to save quite a bit of electricity over its expected 2-3 year career vs the Haswells. I'm sure it will be vastly more power efficient, but where I am, electricity is fairly cheap and favors less efficient, less costly rigs when looking at total costs.

 

[biodoc]

I don't see any reason for switching from your 2P Haswell rigs to 3950X builds if you have cheap electricity. My electricity is expensive and I've got a a single socket Ivy bridge (10 core/20 threads) with an x79 MB so that one is worth upgrading to Zen 2 or the next gen TR I think.