AMD CPU and Platfrom Discussion Episode IV: A New Hope is Ryzen

The Something Awful Forums > Discussion > Serious Hardware/Software Crap > AMD CPU and Platfrom Discussion Episode IV: A New Hope is Ryzen

«‹›4 »

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

Mr.Radar posted:

Phoronix has released a test suite that can reliably reproduce the Ryzen GCC crash bug within minutes both on stock configurations and with both SMT turned off and the memory down-clocked to 2133 MHz. Unfortunately AMD isn't giving them Threadripper or Epyc samples so there's no word yet if they also suffer from this bug.

I've been quite close to hitting buy on the bits for Ryzen upgrade a few times as my current (AMD) windows rig is painful in Lightroom, but I'd stopped myself each time given how poorly Lightroom is written when it comes to making use of cores. This issue has kind of made up my mind to hold off.

I fear this might prove to be rather costly issue for AMD.

# ¿ Aug 5, 2017 14:45

Adbot: ADBOT LOVES YOU

# ¿ May 2, 2024 07:27

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

The Dell AIO Ryzen 7 looks nice...

https://www.youtube.com/watch?v=oGpAfauCrdU

..but I'm rather cynical about the cooling.

# ¿ Aug 6, 2017 13:34

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

I watched that video too. It started out so well, 'finally somebody who took the centre package out of the foam blocks before turning the handle and hence not dropping the CPU oon the floor" but it went downhill from there.

Pablo Bluth fucked around with this message at 14:30 on Aug 6, 2017

# ¿ Aug 6, 2017 13:52

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

ServeTheHome have started to post their EPYC benchmarks
https://www.servethehome.com/dual-amd-epyc-7601-processor-performance-and-review-part-1/

# ¿ Aug 8, 2017 23:43

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

I finally gave in to temptation and ordered the parts to replace my Athlon II X4 machine with a 1700....

# ¿ Aug 14, 2017 08:13

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

It should be double the single-thread performance, double the cores, un-bottleneck the SSD, give me GbE (it's a really crappy motherboard. I've got a spare GbE pcie card but the only PCIe slot is blocked by the graphics card) and give me USB3. Plus I'm replacing the 5,400 rpm hard-drive with a 7,200 one.

However it won't be quite the end of the road for the CPU. I've also got Linux box running on an AMD Athlon 64 X2. I plan to flash the bios to add AM2+ support and transplant the CPU over. Despite being an even older machine, it's a far nicer motherboard.

Pablo Bluth fucked around with this message at 08:43 on Aug 14, 2017

# ¿ Aug 14, 2017 08:36

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

All the stuff for my 1700 build arrived . Apart from the memory which ebuyer haven't even shipped yet...

# ¿ Aug 15, 2017 17:20

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

Devian666 posted:

Why not a 1700? I am prepared to pay for higher clockspeed and prefer not to overclock. Overclocking on a multithreaded job that could take up to a week to complete is a commercial risk if it's unstable.

The main issue with Ryzen as a dedicated CFD workstation is the lack of quad-channel memory, so even at the same-clock TR is likely to be superior.

# ¿ Aug 16, 2017 07:58

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

Munkeymon posted:

hmmmm

I had almost exactly the same issue last night when my memory arrived and I assembled the machine, except for 8/16 rather than 16/32. Turns out one of the dimms wasn't quite seated properly so it knew it was there but couldn't use it.

Pablo Bluth fucked around with this message at 07:25 on Aug 17, 2017

# ¿ Aug 17, 2017 07:08

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

I can now confirm that a Ryzen 1700 is just a wee bit faster than an Athlon II X4....

I managed to get it running with the memory at 2800MHZ with just two mouse clicks and no understandings of what I was doing. The rated 3000 failed to boot so I'm going to have to RTFM.

# ¿ Aug 18, 2017 01:49

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

Anime Schoolgirl posted:

it might be 2T at 2933, since ryzen by default demands 1T memory timings unless you update the bios to one with a newer AGESA version (not all of the mobos out now have it although new ones are shipping with it)

though 1T 2800 might be faster than 2T 2933 and if it's stable there after some stress runs i'd say go with it

The MB is an Asus Prime B350 Plus, and I upgraded to BIOS 806 they released not too long ago includes AGESA 1.0.0.6. The memory is sold as 3000MHz 15-17-17-35. It's set to 2800MHz 15-17-17-35 but seems to end up as 16-17-17-35 when it's actually booted.

# ¿ Aug 18, 2017 21:07

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

Risky Bisquick posted:

AMD Zen is different from other platforms in which it tries a multitude of ram timings prior to boot (you can configure the number of attempts) and then uses whatever works. It might take 3+ tries to get 2933+ running. You can set 15-17-17-35 and get 16-17-17-35 like you saw.

It does seem to do a cycle before it starts to boot proper, so I thought for now I'd set the timing to 16 and leave further investigation for later. However it turns out it can't boot if I set it to 16-17-17-35 and it eventually gives up and reverts to stock.

# ¿ Aug 19, 2017 22:30

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

So strangely my 3000MHZ QVL RAM that I could only get to run at 2800, decided this morning after a week or so of being happy, to suddenly stop booting and require a drop to 2600. That finally got me having a look at timings and voltages . Bumping the memory voltage from 1.35 to 1.37 has finally got it booted at 3000 and so far not giving any errors in Prime95-Blend.

# ¿ Aug 26, 2017 14:08

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

Paul MaudDib posted:

Zen2 has always had the potential for greatness. Zen1 is already not half bad, if they can bump the IPC and clocks up a bit, fix the memory controller's touchiness, and get some of the errata fixed then Zen2 could easily be going toe-to-toe with Coffee Lake.

They should probably introduce AVX512 too. Even if it's not that relevant to most people, it gives a poor impression to lose badly in subset of benchmarks.

# ¿ Sep 5, 2017 20:13

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

The devs who do the performance work on Photoshop/Lightroom/Premiere are the ones who failed to cut it in the Flash security team...

# ¿ Sep 6, 2017 08:02

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

After a decade, I've finally retired my Athlon 64 X2 Dual Core Processor 5000+. However score one for AMD's long term socket compatibility; after a bios update I was able to transplant my AMD Athlon II X4 635 from it's rubbish motherboard in to the older-but-superior board vacated by the 64 X2. All it cost me was a tube of thermal paste and a TDP bump...

# ¿ Sep 10, 2017 21:45

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

There's a new Ryzen bios out that seems to be aimed at improving memory compatibility.

# ¿ Sep 17, 2017 23:06

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

On the subject of memory bandwidth; I have a background in Computational Fluid Dyanmics, so I've been keen to find out how Epyc does against Xeon in CFD. CFD tends to be very memory bandwidth limited. Found somebody on a forum who's posted some comparisons, and the Epyc comes out very nicely.
https://www.cfd-online.com/Forums/hardware/196400-amd-epyc-cfd-benchmarks-ansys-fluent.html

# ¿ Jan 28, 2018 02:03

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

Cygni posted:

idk i kinda like that methodology from a product review standpoint, although I would also like to see 1-1 testing with identical ram settings too.

How often do the big brand's exceed official memory speeds in their pre-built systems?

# ¿ Apr 19, 2018 17:00

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

If 'every bit of electricty' that went in to a CPU came out as heat it'd wouldn't be able to do any other work and wouldn't be very useful. Just most of it.

# ¿ Apr 20, 2018 22:08

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

Yet another bios update for my board seems to have doubled the number of DOCP profiles in the bios, and I can now finally run the memory at the rated 3000 instead of having to settle for 2933...

# ¿ Apr 23, 2018 08:58

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

Deuce posted:

I want 16 cores that all run at 5.0ghz though.

No. A hundred cores.

Man I want more TR. My next build is almost definitely going to be a threadripper line.

I could make use of ten thousand cores at 5Ghz. Scientific computing could really do with some high core count server chips that also blow through 3Ghz on all cores....

# ¿ Apr 26, 2018 20:32

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

Eyes Only posted:

I'd imagine most of the problems that benefit from more than 16 full x86-64 cores and can't be trivially done better on a GPU are not going to be memory bound anyway.

Computational Fluid Dynamics. You always want more cores than you've got but it's also massively memory bandwidth bound.

# ¿ Jun 5, 2018 21:55

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

Methylethylaldehyde posted:

CFD is still strongly single thread limited. Not nearly as badly as most things, but it's often better to have 10 cores at 5ghz vs. 20 cores at 3ghz. Some code sets scale very nicely, where each extra proc gives you 0.97x the previous performance, some code is poo poo and gives you only 0.71x. In some specific cases, it's better to run the code on some hilariously OC'd HEDT proc vs. putting it on the shiny new 90 core quad socket servers.

The very high core count CPUs have tender to be poor choices because they both trade too much mhz and memory bandwidth per core. Everything I've seen points to Eypc performing brilliantly for CFD because of its octo channel memory.

We use Ansys Fluent which claims pretty good linear scaling until quite a low cells per core number.

# ¿ Jun 5, 2018 23:41

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

PC LOAD LETTER posted:

Except nothing really sucks at all about current Epyc parts except for maybe the name. By all accounts they're selling well and the main limiting factor is how fast they can make them. They're not as good at Intel for HPC stuff but AMD acknowledged that last year and said they were targeting the remaining 80% or so (according to them I believe) of the server market that isn't HPC so to say there is a "ton" of stuff where AMD isn't sensible/dominant might be technically true in a narrow but is also mostly eye rolling hyperbole in a practical sense.

Everything I've seen says Eypc is actually really good for HPC. A lot of HPC work is memory bound and it's eight channel memory it has an advantage over Intel.

# ¿ Jul 21, 2018 08:17

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

I had a look to see if anyone had actually announced an Epyc decision for HPC. Last month the Italian National Institute for Nuclear Physics did so although without scale or timeline.

https://insidehpc.com/2018/06/amd-epyc-powers-hpc-national-institute-nuclear-physics-italy/

They've gone for the AMD EPYC 7351 which is the two-socket 16 core variant. I reckon that'll be most common choice, as the best balance between the number of cores, all-core boost and memory bandwidth.

# ¿ Jul 30, 2018 10:24

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

Devian666 posted:

Where I'm finding Zen 2 interesting is in the doubling of the floating point instruction pipelines. When I run fluid dynamics simulations it's all cpu bound and primarily floating point instructions. The difference between my old xeon workstation and my 1700 at home is a 60% increase in speed by doubling cores (even though ryzen is a bit slower in IPC and clock speed). I'm prepared to wait for something Zen 2 with at least 8 cores for my new workstation.

Whats the Xeon configuration? With CFD you pretty quickly run in to being memory bound. So quad channel Threadripper should beat any Ryzen based solution, and Epyc will again be faster due to having eight memory channels. Given eight memory channels is ahead of Intel's offering of six, there's been a lot of interest in Eypc for HPC uses, unfortunately like-for-like benchmarks have been thin on the ground.

# ¿ Nov 11, 2018 12:02

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

Devian666 posted:

It's an old xeon but for the models I'm running I'm not hitting any memory bound issues. I'm running FDS and the fire calculations end up very intense. In the past I'm only seen memory become a problem on a very large model at the start of the run but once the fans start in the model the cpus would always sit at 100%. Even NIST who wrote the software suggest only running as many threads as cpu cores because memory latency isn't an issue.

As I understand it, the usual CPU utilisation number shown doesn't tell if the CPU is memory bound. A CPU is 'busy' whether it is performing instructions or waiting for data from memory (because it's not in the CPU caches). Any CPU utilisation drop at the start of the run is usually a disk IO bottleneck when large runs can't be feed in to memory quickly enough.

# ¿ Nov 12, 2018 18:43

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

How much is driven by the market's needs being forced by per-socket software licensing?

# ¿ Apr 9, 2019 19:01

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

How long am I going to have to wait to for an eight core CPU with integrated graphics so I can finally upgrade my linux desktop (where I don't need or want the a graphics card)? I assume the next round of APUs will be a no-go as it's still based on Zen+ and won't have the die space?

# ¿ Jun 9, 2019 20:15

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

If they're not busy working on 3700X with a GPU as the second chiplet, they're surely being crazy? There's got to be a reasonable market for people who want the high core count options for productivity reasons without the need for a dGPU. I know I'd love one to finally upgrade my linux box where a dGPU would just be a waste.

# ¿ Jun 21, 2019 12:15

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

mdxi posted:

If anybody else has any fairly self-contained things that they'd like to see run for comparisons, I'm happy to consider adding them in. No purely synthetic benchmarks, please; I'm interested in actual tasks.

See if you can get OpenFOAM CFD to work and run the benchmark case discussed here. CFD is great for giving memory access a workout.

# ¿ Jun 29, 2019 09:05

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

mdxi posted:

The benchmark template tarball on the post you linked to is inaccessible unless you're logged into the forum

Follow the link to the wiki at the top then there's a download link there that works.

Edit: beaten by Doolittle's edit

# ¿ Jun 30, 2019 07:46

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

Phoronix's memory benchmarking under Linux
https://www.phoronix.com/scan.php?page=article&item=amd-3900x-ram&num=1

Interestingly the top rating memory is good on everything except a specific Apache test when it shits the bed.

Pablo Bluth fucked around with this message at 19:56 on Jul 10, 2019

# ¿ Jul 10, 2019 19:47

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

HalloKitty posted:

I can't actually remember how much my first PC had, but not much (IBM AT).. first one I built from random parts had 4MB, that much I remember well

My first PC was also an IBM AT, my dad brought it home after work closed his office and made him a roaming worker. It had a beast of a hard drive at 20 MB...

# ¿ Nov 11, 2019 15:21

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

AMD haven't helped themselves by lagging behind on their APU line. The world of non-workstation business desktops seem to have decided that separate GPUs are an unnecessary expense; here AMD still only offer four cores which plays against their advantage. A 3700X APU would stand out a lot more and give reason to build lines around AMD.

# ¿ Jan 4, 2020 20:01

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

NUCs...

# ¿ Jan 7, 2020 00:08

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

Shaocaholica posted:

Zen2 ryzen doesn't support multiple sockets right? Have to go Epyc?

Only non-P variants of Eypc will do dual-socket, Eypc Ps, Threadripper and Ryzen are strictly one-socket.

# ¿ Jan 7, 2020 22:57

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

I believe Crysis was a resource hog whose highest possible graphics setting was beyond all available hardware at the time of release and gave rise to a "But can it run Crysis?" meme.

Pablo Bluth fucked around with this message at 11:29 on Feb 11, 2020

# ¿ Feb 11, 2020 11:27

Adbot: ADBOT LOVES YOU

# ¿ May 2, 2024 07:27

Pablo Bluth: Sep 7, 2007; I've made a huge mistake.

CaptainSarcastic posted:

I have a socket 939 rockin' an Athlon 64 X2 4800+ and 4GB of DDR sitting in my garage. I haven't booted it in a while, but it was running as a secondary PC when I moved a couple years ago. I would need to throw a GPU in it, but I have a GeForce 9500 GT I could throw into it (the 9800 GTX I had it in it died, I think).

My Linux box is still an Athlon II X4 635. However that was a drop-in replacement for an Athlon 64 X2 5000+ so I've got 15 years out of the motherboard.
When the Ryzen 4 are out I think I'll build a new windows box and the Athlon will finally be retired for the windows box hand-me-down (a 1700).

Pablo Bluth fucked around with this message at 10:28 on May 17, 2020

# ¿ May 17, 2020 10:26

The Something Awful Forums > Discussion > Serious Hardware/Software Crap > AMD CPU and Platfrom Discussion Episode IV: A New Hope is Ryzen

«‹›4 »