|
If you get the chance, or can be bothered, if you felt like undervolting it I'd be real interested in how low you could go with stock clocks.
|
#
?
Mar 12, 2017 06:14
|
|
|
|
| # ? May 21, 2024 14:26 |
|
|
I really don't want to play with it anymore. I just spent an hour or two finding out that this RAM won't go over 2133 because reasons. Also apparently CPU-Z and stuff report voltages roughly around half of what is set in bios
|
|
#
?
Mar 12, 2017 06:41
|
|
|
The ram could be an early BIOS thing. Does it detect timings correctly? Someone on spcr will undervolt one eventually. It's just idle curiosity really.
|
|
#
?
Mar 12, 2017 07:53
|
|
|
Oc3d talked about ram on amd boards in their hero review. And has to specifically test the ram and send code to the board manufacturers to put in the bios. This is for the amd version of xmp to work. So check the qvl and only get ram that's on it.
|
|
#
?
Mar 12, 2017 14:19
|
|
|
That's kind of ew.
|
|
#
?
Mar 12, 2017 14:43
|
|
|
SwissArmyDruid posted:That's kind of ew. It's eventually going to work better, this is just something people who bought early on have to deal with.
|
|
#
?
Mar 12, 2017 15:05
|
|
|
Even entering the timings manually doesn't work? I must be misunderstanding how xmp works. I assumed it just held a list of all the timings the ram supported at diff frequencies. I guess it can't be that if amd can't just use that data? E- oh, i guess the AMD controller can't run at the same timings? GRINDCORE MEGGIDO fucked around with this message at 18:14 on Mar 12, 2017 |
|
#
?
Mar 12, 2017 17:27
|
|
|
Don Lapre posted:Oc3d talked about ram on amd boards in their hero review. GRINDCORE MEGGIDO posted:Even entering the timings manually doesn't work? I can edit the frequency, timing, and voltages, but I haven't gotten any setting to work besides auto though I didn't play with loosening the timings. I'm not going to bother tweaking it anymore at least for another couple bios updates. There are a lot of people having no luck with getting their ram over 2133.
|
|
#
?
Mar 12, 2017 21:43
|
|
|
Ah, so it can read the XMP profiles, OK. e- did you try 2666Mhz? GRINDCORE MEGGIDO fucked around with this message at 22:08 on Mar 12, 2017 |
|
#
?
Mar 12, 2017 21:53
|
|
|
Here is a good video explaining the weirdness with RAM in the Ryzen ecosystem. Basically the Ryzen memory controller is very finicky and tries to force some fairly silly timings on the RAM, this might be worked out in the future but for now YMMV.
|
|
|
#
?
Mar 12, 2017 21:57
|
|
|
GRINDCORE MEGGIDO posted:Ah, so it can read the XMP profiles, OK. I tried setting 2400, 2666, and 2933 at cas 15 1.35v, none of which worked AVeryLargeRadish posted:Here is a good video explaining the weirdness with RAM in the Ryzen ecosystem. Basically the Ryzen memory controller is very finicky and tries to force some fairly silly timings on the RAM, this might be worked out in the future but for now YMMV.
|
|
#
?
Mar 12, 2017 23:40
|
|
|
Got a question for folks. I'm in need of a cheap AMD processor for a toy machine to run some numerical problems (CFD/FEA) as proof of concept before moving to some serious computing hardware. I'm not concerned with Ryzen at this point, maybe at a later date. Most of the AMD chips on offer are from a generation I skipped - either the APU series (A6-7400K, A8-7600, A10-7700K) or FX series (FX4 - 4300, FX6-6300). Any recommendation on these?
|
|
#
?
Mar 13, 2017 01:19
|
|
|
Zero Gravitas posted:Got a question for folks. Is there any particular reason you're looking at amd?
|
|
#
?
Mar 13, 2017 01:24
|
|
|
Zero Gravitas posted:Got a question for folks. I know this is an amd thread and all but just buy a pentium dual core and an h110 motherboard.
|
|
#
?
Mar 13, 2017 01:28
|
|
|
AVeryLargeRadish posted:Here is a good video explaining the weirdness with RAM in the Ryzen ecosystem. Basically the Ryzen memory controller is very finicky and tries to force some fairly silly timings on the RAM, this might be worked out in the future but for now YMMV. the HAIR on this lil motherfucker
|
|
#
?
Mar 13, 2017 01:29
|
|
|
Because CFD and FEA is sped up by distributing the job to cores via MPI. I have options of dual core intel chips at sixty quid, or some AMD quad core chips (assuming theyre actual physical cores and not just virtual ones).
|
|
#
?
Mar 13, 2017 01:37
|
|
|
Zero Gravitas posted:Because CFD and FEA is sped up by distributing the job to cores via MPI. I have options of dual core intel chips at sixty quid, or some AMD quad core chips (assuming theyre actual physical cores and not just virtual ones). What about old cheap server cpus? http://www.ebay.com/itm/HP-XW8600-F...BsAAOSwawpXtNie http://www.ebay.com/itm/Intel-S5000...QIAAOSwOtBXTTSB wargames fucked around with this message at 02:04 on Mar 13, 2017 |
|
#
?
Mar 13, 2017 02:00
|
|
|
Didn't someone in this very thread say that CFD is more about memory bandwidth, in which case you could probably find some used tri/quad channel intel machine on the cheap. Back in the day I bought a 980x machine for like $400 because the guy I bought it from was upgrading to the latest and greatest, and it was only a year old then
|
|
#
?
Mar 13, 2017 02:12
|
|
|
Zero Gravitas posted:Because CFD and FEA is sped up by distributing the job to cores via MPI. I have options of dual core intel chips at sixty quid, or some AMD quad core chips (assuming theyre actual physical cores and not just virtual ones). Unless you're actually doing MPI stuff that relies on having actual cores (i.e.things that are using low-level thread barriers and inter-thread synchronization) then having more cores actually doesn't buy you anything if they're slower. Facetiously, you can make the same argument about putting a Core2Quad up against a Skylake i5 and seeing what happens - but the thing is Bulldozer actually isn't much faster than Core2Quad (which came out years earlier). That's the problem with the AMD Bulldozer-derived products - they have lots of cores but the actual per-core performance is so terrible that the Intel chips outperform them significantly at any given price point. In fact it was effectively a step backward even for a circa-2012 architecture, as AMD's previous product (Phenom II) actually outperformed it too, in many cases. Also - the other thing with Bulldozer is that despite the marketing, CMT is not an actual independent core. Each pair of Bulldozer cores still share execution resources and can bottleneck each other just like a hyperthread - especially in floating-point heavy tasks which it sounds like this is. A 4-core Bulldozer is basically the same thing as a 2C4T i3 processor. If you want to build a really high-end workstation on the cheap, the best approach is to buy an LGA2011v3 or LGA2011 and buy some engineering sample Xeons. Or as mentioned, older workstations built with Sandybridge or Bloomfield Xeon chips are a dime a dozen these days, you can buy the whole machine for less than the CPU/motherboard for a nice modern build just because nobody wants to pay the power bill for them. If you're memory-bandwidth-bound, then look for something using an E5 Xeon chip that has quad-channel memory. Paul MaudDib fucked around with this message at 03:06 on Mar 13, 2017 |
|
#
?
Mar 13, 2017 02:52
|
|
|
I've seen the new tests that show that the Windows 10 scheduler doesn't have a problem knowing what to do with Ryzen. They've got the latency tests for each core, numbers on CCX switching, etc. It all seems pretty conclusive. ... so why do most games perform better when SMT is switched off?
|
|
#
?
Mar 13, 2017 18:36
|
|
|
games are held together by lovely glue and duct tape, why are you surprised lol
|
|
#
?
Mar 13, 2017 18:39
|
|
|
Maxwell Adams posted:I've seen the new tests that show that the Windows 10 scheduler doesn't have a problem knowing what to do with Ryzen. They've got the latency tests for each core, numbers on CCX switching, etc. It all seems pretty conclusive. But they don't have that problem on Windows 7, or at least that's what I remember reading..
|
|
#
?
Mar 13, 2017 18:42
|
|
|
Maxwell Adams posted:I've seen the new tests that show that the Windows 10 scheduler doesn't have a problem knowing what to do with Ryzen. They've got the latency tests for each core, numbers on CCX switching, etc. It all seems pretty conclusive. 
kirtar fucked around with this message at 19:35 on Mar 13, 2017 |
|
#
?
Mar 13, 2017 18:43
|
|
|
Paul MaudDib posted:Unless you're actually doing MPI stuff that relies on having actual cores (i.e.things that are using low-level thread barriers and inter-thread synchronization) then having more cores actually doesn't buy you anything if they're slower. Facetiously, you can make the same argument about putting a Core2Quad up against a Skylake i5 and seeing what happens - but the thing is Bulldozer actually isn't much faster than Core2Quad (which came out years earlier). That's the problem with the AMD Bulldozer-derived products - they have lots of cores but the actual per-core performance is so terrible that the Intel chips outperform them significantly at any given price point. In fact it was effectively a step backward even for a circa-2012 architecture, as AMD's previous product (Phenom II) actually outperformed it too, in many cases. LIke I said, this machine is intended as a toy machine to test out some workflow processes and whether its possible to do this idea i have (however slowly) on a shitbox frankenstein machine thrown together from spare parts I have lying around and some from the bargain basement bin instead of dropping serious coin on it. I know from work that OpenFOAM doesnt play particularly nice with hyperthreading so I was trying to avoid that. I appreciate the nod towards ex-server hardware, but I'm really not looking for that kind of stuff right now. Is there any kind of recommendation on the hardware I've previously listed?
|
|
#
?
Mar 13, 2017 19:09
|
|
|
PSA: Crosshair Hero VI and EK Supremacy Evo (and possibly other AM3-mounted coolers!) The Crosshair Hero VI motherboard has holes for both AM3 and AM4 coolers. However, certain types of mounts using the AM3 holes can have issues with mounting pressure as the heights are slightly different. Specifically the EK Supremacy EVO CPU block wont boot when the block is screwed down. The board gets power, the LEDs and green "CPU ready" light will turn on, but nothing happens when you hit the start button. On mine, I unscrewed the mounts and the board would power on. Many people are solving this problem by removing the center portion of the rubber gasket despite the instructions specifically telling you to use it for both AM3 and AM4 backplates. WTF, Asus.
|
|
#
?
Mar 13, 2017 20:01
|
|
|
Zero Gravitas posted:LIke I said, this machine is intended as a toy machine to test out some workflow processes and whether its possible to do this idea i have (however slowly) on a shitbox frankenstein machine thrown together from spare parts I have lying around and some from the bargain basement bin instead of dropping serious coin on it. I know from work that OpenFOAM doesnt play particularly nice with hyperthreading so I was trying to avoid that. I appreciate the nod towards ex-server hardware, but I'm really not looking for that kind of stuff right now. Given the choice of APUs or FX processors, I would go with the FX processors for sure if you are remotely performance-sensitive.
|
|
#
?
Mar 13, 2017 20:08
|
|
|
If Zen has such severe penalties for crossing clusters, maybe AMD should have set it up as a dual CPU instead of blabbing about infinity fabrics.
|
|
#
?
Mar 13, 2017 20:15
|
|
|
Maxwell Adams posted:I've seen the new tests that show that the Windows 10 scheduler doesn't have a problem knowing what to do with Ryzen. They've got the latency tests for each core, numbers on CCX switching, etc. It all seems pretty conclusive. The short answer? We still don't know. Articles might be coming out which declaratively state one thing or another, but the reality of it is, what we are going through right now, with reviews and testing is the same kind of thing that happens any time someone writes a scientific paper in a journal. People are posting data, posting their experimental setup, and their conclusions, and everyone is trying to reconcile why one person's results are the way they are, which spawns more questions, more experiments, and more data, and their conclusions. So even though PCPer might have an article saying, "we don't think the Windows scheduler is the problem", it's probably more correct to wait for corroboration from other sources before taking their word for it. PCPer isn't a respected source because they're right, PCPer is a respected source because they diligently report their results and justify the conclusions that they draw from their data without making many assumptions. I mean when you get down to it, hardware reviews are basically the scientific method. SwissArmyDruid fucked around with this message at 20:24 on Mar 13, 2017 |
|
#
?
Mar 13, 2017 20:20
|
|
|
|
|
#
?
Mar 13, 2017 20:26
|
|
|
sauer kraut posted:If Zen has such severe penalties for crossing clusters, maybe AMD should have set it up as a dual CPU instead of blabbing about infinity fabrics. Numa isn't the answer because numa looks at two different memory controllers for two differant cpus. Zen only has one memory controller. So what might need to happen is a new scheduler just for this new type of architecture.
|
|
#
?
Mar 13, 2017 20:26
|
|
sauer kraut posted:If Zen has such severe penalties for crossing clusters, maybe AMD should have set it up as a dual CPU instead of blabbing about infinity fabrics. Lol, no. It's not like applications would just use both CPUs in a setup like that, also you have per-core licensing to deal with, making the CPU identify as two cores is an absolutely terrible idea. All that really needs to be done is to make applications and OSes aware of the extra latency when sending things between clusters so that behaviour is minimized, that should be something that MS can sort out in a couple months. Also it's not a really severe penalty, remember that within clusters the switching time is 40ns, it goes up to 140ns when switching between clusters, but on Intel CPUs the normal switching time is 80ns, so it's not nearly as severe as it seems. Lastly while the switching time is probably causing problems I doubt it is all there is to this.
|
|
|
#
?
Mar 13, 2017 20:26
|
|
|
wargames posted:Numa isn't the answer because numa looks at two different memory controllers for two differant cpus. Zen only has one memory controller. So what might need to happen is a new scheduler just for this new type of architecture. Do we know if this is true for Naples too, or is Naples achieving it's "8 channel memory" with 4 separate dual channel controllers?
|
|
#
?
Mar 13, 2017 20:29
|
|
|
Twerk from Home posted:Do we know if this is true for Naples too, or is Naples achieving it's "8 channel memory" with 4 separate dual channel controllers? Since napples isn't out yet we don't know yet?
|
|
#
?
Mar 13, 2017 20:30
|
|
|
hardwarereviews.jpg
|
|
#
?
Mar 13, 2017 20:36
|
|
|
perchance was it you who bought anandtech
|
|
#
?
Mar 13, 2017 21:22
|
|
|
Zero Gravitas posted:LIke I said, this machine is intended as a toy machine to test out some workflow processes and whether its possible to do this idea i have (however slowly) on a shitbox frankenstein machine thrown together from spare parts I have lying around and some from the bargain basement bin instead of dropping serious coin on it. I know from work that OpenFOAM doesnt play particularly nice with hyperthreading so I was trying to avoid that. I appreciate the nod towards ex-server hardware, but I'm really not looking for that kind of stuff right now. I'd just get an FX-8350 or 8370. They're 125 Watt CPUs so you'll need a decent cooler and a motherboard that can handle them. If you want 95 Watts instead there's an 8320e and 8370e which are low power. I bought an OEM 8300 for a VM box a while back that's been fine and seems to still be available for $100: https://smile.amazon.com/dp/b00tr8yl4w/ . I'd try to pay as little as possible but definitely get an 8 core if you have to get some old AMD garbage. If you were more interested in old server hardware I've put together a couple of Westemere dual socket Xeon boxes for around $300 for motherboard, CPUs, 48GB of RAM. With dual 6 core Xeons they're pretty good for the money and the supermicro boards I used are ATX form factor so I was able to use normal cases and power supplies.
|
|
#
?
Mar 13, 2017 21:39
|
|
|
I know this is the AMD thread but I would not honestly recommend anyone buy an FX at this point. If Ryzen is too pricey and you can't wait for lower-end Zen models, get the cheapest i5.
|
|
#
?
Mar 13, 2017 21:48
|
|
|
FuturePastNow posted:I know this is the AMD thread but I would not honestly recommend anyone buy an FX at this point. If Ryzen is too pricey and you can't wait for lower-end Zen models, get the cheapest i5. Oh, I agree, but he specifically wants one for some reason. Despite all of the warnings about how horrible they are and that they just became obsolete, he still wants one. So, I'm giving the advice he asked for with those caveats.
|
|
#
?
Mar 13, 2017 21:52
|
|
|
FuturePastNow posted:I know this is the AMD thread but I would not honestly recommend anyone buy an FX at this point. If Ryzen is too pricey and you can't wait for lower-end Zen models, get the cheapest i5. Or, a 4-thread G4560 for $50, the best value running in CPUs right now. It's half an i7 for 1/6 the price.
|
|
#
?
Mar 13, 2017 22:07
|
|
|
|
| # ? May 21, 2024 14:26 |
|
|
Rexxed posted:Oh, I agree, but he specifically wants one for some reason. Despite all of the warnings about how horrible they are and that they just became obsolete, he still wants one. So, I'm giving the advice he asked for with those caveats. I think he has the machine sans cpu sitting around in parts and wants to use them or something, not that it makes a huge difference as far as what is most sensible.
|
|
#
?
Mar 13, 2017 22:33
|
|