|
BeastOfExmoor posted:On the other hand the pricing scales surprisingly linearly. It's less than the cost of 2 3800X's and clocks even higher out of the box. And that's why the 3800X exists.
|
# ? Jun 11, 2019 03:22 |
|
|
# ? Jun 5, 2024 18:32 |
|
Cao Ni Ma posted:The only reason I'd buy a 3800x is if it can overclock better, if it doesn't then I just cant justify the price increase Friends don't let friends buy 3800x. If you want some room to overclock, get a drat 3900x and put a huge cooler on.
|
# ? Jun 11, 2019 03:23 |
|
Twerk from Home posted:Friends don't let friends buy 3800x. If you want some room to overclock, get a drat 3900x and put a huge cooler on. Or just a normal cooler, these aren't Intel parts.
|
# ? Jun 11, 2019 03:27 |
|
$750 feels like a reasonable price, but I think a lot of people who will consider it are going to do so for future-proofing reasons when they could instead get a 3800X or 3900X for the same immediate benefit and trade it in for a future 4950X before the extra cores would really matter. I mean, it's a great chip for a home VM pod but are any games especially at 1440p+ going to care about the extra 8 cores in the next couple years? Impressive overall though, I'm pretty sure I'll be replacing my X5660 with one of these. Eletriarnation fucked around with this message at 03:44 on Jun 11, 2019 |
# ? Jun 11, 2019 03:41 |
|
Ian Cutress over at Anandtech wrote up a deep dive on the Zen 2 uarch improvements.
|
# ? Jun 11, 2019 03:47 |
|
For home users, the 3600X, 3700X, and 3900X are the interesting ones. The 3600X and 3700X will be the value buys, and the 3900X, while the extra cores will be useless, may be a real beast for applications that benefit from the extra cache. The 3950X is just silly and I can't really see a point for anything shy of serious workstation use. A big part of why they stomped Intel so hard in their benches is that they probably ran the Intel parts at 2666 since that's the "official" max ram speed, and given how they've been hyping up their memory controller I bet they went 4000+ on the Ryzen parts.
|
# ? Jun 11, 2019 03:54 |
|
DDR4-3733 will probably be the best for gaming due to latency.
|
# ? Jun 11, 2019 04:41 |
|
Lowen SoDium posted:A couple of PCIe questions: Can't answer question one, but for the second one, yes. Backwards compatibility is baked in to all PCIe devices since the 2.1 spec onwards, but you might run into some issues running on a mobo with PCIe 1.0/1.1 (and 1.1), but you should be able to run anything on a 2.0 board or later. You'll lose performance, however.
|
# ? Jun 11, 2019 04:53 |
|
Are there very many gaming situations where memory latency matters more than memory bandwidth? Those numbers look like at least 3200-3733 is comparable to 2nd-gen+3200CL14 latency.
|
# ? Jun 11, 2019 04:54 |
|
Lowen SoDium posted:A couple of PCIe questions: edit: apropos of nothing: drat, that's a loving rock. SwissArmyDruid fucked around with this message at 05:04 on Jun 11, 2019 |
# ? Jun 11, 2019 05:00 |
|
SwissArmyDruid posted:apropos of nothing: drat, that's a loving rock. Hell yeah, this is now a Lisas finger rock discussion thread.
|
# ? Jun 11, 2019 05:47 |
|
SwissArmyDruid posted:
I’m going to assume that’s stock vs stock and Zen 2 is using an advanced form of XFR, so the 9900KS will likely keep the lead when both CPUs are overclocked. There is little doubt that the Ryzen will be the better buy overall though. Nice!
|
# ? Jun 11, 2019 06:29 |
|
Unfortunately, that picture doesn't show if the PCIe lanes go to the CCX chiplets or to the memcontroller. I am curious how PCIe bandwidth works on the CPUs with more than one CCX die. Also, that image is misleading/wrong. It shows the 4x chipset downlink on both sides of the CPU. Lowen SoDium fucked around with this message at 06:47 on Jun 11, 2019 |
# ? Jun 11, 2019 06:44 |
|
incoherent posted:Hell yeah, this is now a Lisas finger rock discussion thread. She should really get it rebuilt on a newer node.
|
# ? Jun 11, 2019 06:44 |
|
Lowen SoDium posted:Unfortunately, that picture doesn't show if the PCIe lanes go to the CCX chiplets or to the memcontroller. I am curious how PCIe bandwidth works on the CPUs with more than one CCX die. Based on slides that we've seen regarding Rome-based EPYC CPUS, the PCIe links hang off the IO die.
|
# ? Jun 11, 2019 06:52 |
|
Lowen SoDium posted:Unfortunately, that picture doesn't show if the PCIe lanes go to the CCX chiplets or to the memcontroller. I am curious how PCIe bandwidth works on the CPUs with more than one CCX die. Everything goes to the IO die. The CPU dies themselves have a single IF connection each to the IO die, and nothing else. Including no direct connection to each other.
|
# ? Jun 11, 2019 06:58 |
|
Subjunctive posted:She should really get it rebuilt on a newer node. I can currently only afford a 7nm diamond
|
# ? Jun 11, 2019 07:06 |
|
Is there any CPU conference scheduled in Fall yet? I needs me some new Threadrippers.
|
# ? Jun 11, 2019 07:17 |
|
Combat Pretzel posted:Is there any CPU conference scheduled in Fall yet? I needs me some new Threadrippers. https://www.amd.com/en/event-listing You might get some announcements at Supercomputing or SIGGRAPH but more likely about Rome and GPUs, respectively. I guess maybe they'll tack on Threadripper, seems unlikely though if they skipped mentioning it at Computex. I have no experience with the other two conferences but if someone announces something notable at the HP Enterprise one I'll eat my shorts.
|
# ? Jun 11, 2019 07:43 |
|
It's not listed there but they have a keynote at Hot Chips in August as well. http://www.hotchips.org/program/ quote:1:45 PM – 2:45 PM: Keynote 1: “Delivering the Future of High-Performance Computing with System, Software and Silicon Co-Optimization” by Dr. Lisa Su, CEO, AMD
|
# ? Jun 11, 2019 07:58 |
|
MaxxBot posted:DDR4-3733 will probably be the best for gaming due to latency. Hopefully someone will do a comprehensive gaming benchmark of Zen 2 including running different memory speeds and timings.
|
# ? Jun 11, 2019 09:01 |
|
that 70mb cache tho... the bump to 70mb may actually counter whatever impact cross-die latency has, for most practical purposes. Seems like the 3900X is a pretty solid value leader there. Still not sure about gaming when using 24+ threads (too bad nobody does AOTS anymore) but if it doesn't edge out in front it'll be close enough anyway. That would have been crazy back in the Monero days... 2mb per thread = 35 parallel monero threads. The days of Opteron barnburners are definitely over. Paul MaudDib fucked around with this message at 09:52 on Jun 11, 2019 |
# ? Jun 11, 2019 09:44 |
|
K8.0 posted:The 3950X is just silly and I can't really see a point for anything shy of serious workstation use. Room full of spark nodes and/or kubelets, for cheap
|
# ? Jun 11, 2019 09:54 |
|
Mr.Radar posted:Ian Cutress over at Anandtech wrote up a deep dive on the Zen 2 uarch improvements. Very interesting slide here: sweet spot for RAM will be ~3600 MHz
|
# ? Jun 11, 2019 12:15 |
|
Lube banjo posted:Very interesting slide here: sweet spot for RAM will be ~3600 MHz Will most previous gen motherboards support that kind of speed?
|
# ? Jun 11, 2019 12:23 |
|
Once again, I'm glad I bought nice ram (b die) that should clock up to 3733. Any chance of decent Matx boards this time around? Sininu posted:Will most previous gen motherboards support that kind of speed? Doubt it.
|
# ? Jun 11, 2019 12:36 |
|
Why 3600C16+, when 3200C14 will do fine according to that slide?
|
# ? Jun 11, 2019 14:22 |
|
Combat Pretzel posted:Why 3600C16+, when 3200C14 will do fine according to that slide? I'm assuming this is about more bandwidth, and recent pricing meaning that you don't start paying a huge price premium until 3733 or higher. This is so many cores to keep fed. Go back a couple years and you've got quad channel RAM on 8C parts, and now dual channel has to keep up with 16C.
|
# ? Jun 11, 2019 14:23 |
|
Because bandwidth is really important.
|
# ? Jun 11, 2019 14:23 |
|
GRINDCORE MEGGIDO posted:Once again, I'm glad I bought nice ram (b die) that should clock up to 3733.
|
# ? Jun 11, 2019 14:34 |
|
According to Paul's Hardware, the benchmarks AMD showed off at E3 were done with all the Spectre/Meltdown/etc. mitigations disabled and on Windows 10 1809 which doesn't have the Ryzen scheduler improvements from 1903: https://www.youtube.com/watch?v=vzV24ZqrJ4s&t=394s
|
# ? Jun 11, 2019 14:35 |
|
Anime Schoolgirl posted:B-die is going out of production, sadly. samsung is replacing it with something, so i expect the next memory they will also be super good.
|
# ? Jun 11, 2019 14:43 |
|
GRINDCORE MEGGIDO posted:Doubt it. People have been doing 3600 CL18 with Zen+, I don't think it's crazy that Zen2 might be able to do 3600 CL16 with the new faster memory kits. But probably only with 2 sticks populated. Combat Pretzel posted:Why 3600C16+, when 3200C14 will do fine according to that slide? That slide shows latency, while bandwidth would be a straight line going up.
|
# ? Jun 11, 2019 15:01 |
|
.
sincx fucked around with this message at 05:50 on Mar 23, 2021 |
# ? Jun 11, 2019 15:17 |
|
So if I have 2x8GB 3200Mhz CL16 (on my Ryzen 2600X), would that be a problem or something if I ever wanted to go up to a 3700/3800/3900X ? Or is one of those "yeah you'll lose 1% of actual real-life performance, don't sweat it" situations?
|
# ? Jun 11, 2019 15:23 |
|
K8.0 posted:Because bandwidth is really important.
|
# ? Jun 11, 2019 16:09 |
|
sincx posted:The only reasonably priced 16GBx2 3600 kits I can find are CL19. We can probably extrapolate off Epyc and the new Rome stuff. Epyc was 1 channel for every 4 cores @2666 and did not seem to be bandwidth limited. Rome is going to double that to 1 channel per 8 cores @3200 so only a moderate 20% frequency bump while doubling the addressed cores so there must be a fair amount of memory bandwidth available on the vast majority of workloads. They'd be adding in more memory channels if it wasn't. I'd venture a guess that even with a 16 core dual-chiplet Zen2 chip you'd be better off with 3733 and the higher IF frequency. You're dealing with higher core clocks on the desktop processors but the ram is also clocked faster and I suspect those will about balance each other out.
|
# ? Jun 11, 2019 16:16 |
|
Klyith posted:People have been doing 3600 CL18 with Zen+, I don't think it's crazy that Zen2 might be able to do 3600 CL16 with the new faster memory kits. But probably only with 2 sticks populated. Mine won't do it, hoping zen2 does with a 570 board.
|
# ? Jun 11, 2019 16:27 |
|
I'm finally getting into that AnandTech article, and this is really interesting especially for VM nerds (I'm still new to virtualization, so I'm not sure how much this will impress other folks):AnandTech posted:As with Intel’s implementation, when a series of VMs is allocated onto a system on top of a hypervisor, the hypervisor can control how much memory bandwidth and cache that each VM has access to. If a mission critical 8-core VM requires access to 64 MB of L3 and at least 30 GB/s of memory bandwidth, the hypervisor can control that the priority VM will always have access to that amount, and either eliminate it entirely from the pool for other VMs, or intelligently restrict the requirements as the mission critical VM bursts into full access. This is super cool, and it makes even higher-end Ryzen processors pretty great for virtualization, since that's an awesome use for all those cores. I'm extra excited for Threadripper for workstation builds & SFF NAS builds. Not needing to find the space for a 1U cabinet or anything is super helpful.
|
# ? Jun 11, 2019 17:11 |
|
|
# ? Jun 5, 2024 18:32 |
|
Mr.Radar posted:According to Paul's Hardware, the benchmarks AMD showed off at E3 were done with all the Spectre/Meltdown/etc. mitigations disabled and on Windows 10 1809 which doesn't have the Ryzen scheduler improvements from 1903: Oh wow, more specifically, BOTH systems were done without hyperthreading mitigations. So on a fully-updated bench rig, Intel perf is only going to go down, and Ryzen perf should only go up, making the gaps even wider. Talk about a kicking a man when he's down.
|
# ? Jun 11, 2019 20:17 |