AMD Threadripper 1950X Linux Benchmarks

[schmidtbag]

1. PSU --> EVGA SuperNOVA T2 80 Plus Titanium Netzteil 1000 Watt
2. Motherboard --> ASUS ROG ZENITH EXTREME
3. CPU --> AMD Ryzen Threadripper 1950X Socket TR4
4. Cpu cooler --> Noctua NH-U14S TR4-SP3
5. RAM --> G.Skill Trident Z 32GB DDR4 3600 CL16 (4x8GB) 32GTZSW
6. Graphic Card --> ZOTAC GeForce GTX 1080Ti AMP! Extreme 11GB
7. Hard Disk --> Samsung SSD 960 Pro M.2 512GB x2 ( RAID)

Any suggestions?

Yes:
Unless you intend to SLi and/or switch to liquid cooling (in order to overclock), the PSU is overkill. Also, RAIDing M.2 drives will very likely cause worse performance except in synthetic sequential read/write tests. Either get a single 1TB M.2 drive or have both drives serve different purposes (for example, have one for Linux and the other for Windows).

Otherwise, sounds good.

[Davidovitch]

Interesting! What would be helpful for me as a reader is to have a table included showing the (indicative) price of the different tested CPU's. You mention it for some of them in the beginning, but my rusty memory immediately forgets the price right after reading about it :-) A table (or label in the graphs maybe?) helps in getting the economic context back.

[jrch2k8]

Already Sold ... ordering one...

So here are the parts I have decided to order:

1. PSU --> EVGA SuperNOVA T2 80 Plus Titanium Netzteil 1000 Watt
2. Motherboard --> ASUS ROG ZENITH EXTREME
3. CPU --> AMD Ryzen Threadripper 1950X Socket TR4
4. Cpu cooler --> Noctua NH-U14S TR4-SP3
5. RAM --> G.Skill Trident Z 32GB DDR4 3600 CL16 (4x8GB) 32GTZSW
6. Graphic Card --> ZOTAC GeForce GTX 1080Ti AMP! Extreme 11GB
7. Hard Disk --> Samsung SSD 960 Pro M.2 512GB x2 ( RAID)

Any suggestions?

Note: As the other poster suggested if you wanna RAID will be problematic if you are not using ZFS.

So if you wanna use ZFS in RAID1 go for 2 M.2 drives it will work fine but if you don't wanna use ZFS then use 1 drive for Linux and the other for Passthrough for another OS

Note 2: Also note the RAID issues is software issue(outside ZFS) not hardware because unlike X299 the X399 offer each M.2 drive a fully dedicated X4 PCIe lane whereas X299 share the X4 PCIe chipset lanes for all devices in which case you won't ever reach more speed than 1 M.2 device speeds

[jrch2k8]

Already Sold ... ordering one...

So here are the parts I have decided to order:

1. PSU --> EVGA SuperNOVA T2 80 Plus Titanium Netzteil 1000 Watt
2. Motherboard --> ASUS ROG ZENITH EXTREME
3. CPU --> AMD Ryzen Threadripper 1950X Socket TR4
4. Cpu cooler --> Noctua NH-U14S TR4-SP3
5. RAM --> G.Skill Trident Z 32GB DDR4 3600 CL16 (4x8GB) 32GTZSW
6. Graphic Card --> ZOTAC GeForce GTX 1080Ti AMP! Extreme 11GB
7. Hard Disk --> Samsung SSD 960 Pro M.2 512GB x2 ( RAID)

Any suggestions?

Another suggestion if you plan to use ZFS and don't have an external NAS/SAN around, change those SSD to 256G(if you wanna save some cash)but get 3 instead and add at least 4(RAID10) to 8(RAIDZ2 aka RAID 60) HDD(NAS class maybe) for cold storage(sata doesn't affect SDDs on X399) and use the third SSD as a L2Arc/ZIL cache, M.2 SSD are not as trust worthy as HDD yet, so is always good to have a cold storage RAID around.

[nomadewolf]

Does it rip the threads?

[M@yeulC]

It sounds pretty compelling, but what is the reason for using a different socket from EPYC? The form factors look the same, as is the TDP, IIRC.

[aicom]

Instead of compiling the linux kernel, could you try compiling chromium? There was some odd results on Anandtech where it was slower than the 4c/8t Intel chips, but that makes very little sense, so it would be interesting to see how it performs whening compiling that big a project on Linux.

I'm wondering whether the NUMA vs. UMA setting for ThreadRipper makes a (big) difference there. AnandTech tested AMD's recommended modes (both dies+UMA=Creator mode and one die+NUMA=Gaming mode) but neither is the "optimal" mode like how EPYC would run (NUMA+all dies). I've yet to see any site bench ThreadRipper in this way.

In Gaming mode, you're starved for cores since an entire die is disabled. In Creator mode, the memory latency is significantly higher because the OS is unaware of the 2 separate memory controllers. Both modes have significant performance problems with compilation (since compilation is embarassingly parallel at the process level and prefetching compilation memory fetches is really hard).

Michael, any chance you could run at least a Linux compilation benchmark with your "Memory Access Mode" set to "Local"?

[bisby]

Thanks for the numbers, Michael.
But I'm sometimes astonished that some benchmarks... well, don't they scale well on multicore? I'd say we can compare RyZen 7 and Threadripper like apple to apple in a benchmark. They share the same architecture and have similar clocking. But the big Threadripper has twice the cores, still in some benchmarks Ryzen 7 1800X is on top.
Or is there somwhere else an obstacle in Threadripper (die internal communication between blocks? memory latencies?)?
Looks strange to me.

It's been shown that a 7700k beats a 1800x in most games, but a 1800x crushes a 7700k in gaming and streaming at the same time. I imagine the same concept holds true for going beyond 8 cores. Trying to play a game, stream, and render a video all at the same time would benefit from more cores. It's a pretty excessive use case, but it's also the kind of target use case right now. And who knows, in the future, we may wind up with better threaded tools that scale up better with more cores. But as it stands, I would say that the "use case" is simply doing multiple heavy workloads at the same time.

[tg--]

Already Sold ... ordering one...

So here are the parts I have decided to order:

1. PSU --> EVGA SuperNOVA T2 80 Plus Titanium Netzteil 1000 Watt
2. Motherboard --> ASUS ROG ZENITH EXTREME
3. CPU --> AMD Ryzen Threadripper 1950X Socket TR4
4. Cpu cooler --> Noctua NH-U14S TR4-SP3
5. RAM --> G.Skill Trident Z 32GB DDR4 3600 CL16 (4x8GB) 32GTZSW
6. Graphic Card --> ZOTAC GeForce GTX 1080Ti AMP! Extreme 11GB
7. Hard Disk --> Samsung SSD 960 Pro M.2 512GB x2 ( RAID)

Any suggestions?

Your chosen PSU is _way_ over-dimensioned.
I know, all the forums are full of threads that huge PSUs are needed for any modern hardware, but that's just plain and simply a myth.
A high-quality 750W unit is plenty, will likely cost you less, and most importantly, will be much more efficient in idle and in typical use scenarios.
A rule of thumb for PSUs is 20% more power than expected maximum draw (personally I will usually chose less, because high-quality units handle short term peaks over their rating just fine).

Personally I'd also chose ECC ram. I know, there are no 3600 MT ECC DIMMs available, but I'd prefer the safety ECC provides. Since the high-speed DIMMs don't ever have ECC, nobody even knows how many non-corrected errors this ram produces.
I've overclocked my ECC ram, and with very high clock the system will run fine, but ECC errors suddenly become very common (while they're otherwise very rare). In fact, when people run their ram at the limits (which for Ryzen often is 3200 MT), the system will often become unstable, and people expect that when they move down a bit and it seems stable, it is error free.
In my experience it is not error free, there are just not enough bitflips to crash the system, but very likely enough to corrupt data. But again, nobody tested it, because the high-spec DIMMs never have ECC, so nobody actually knows.

[Anty]

Could you for sake of sanity test kill_ryzen script (all threads ofc)?