|
Where do these +33% memory bandwidth come from?
|
#
?
Jun 26, 2018 20:01
|
|
|
|
| # ? Apr 29, 2024 11:07 |
|
|
Combat Pretzel posted:Where do these +33% memory bandwidth come from? Going from hexachannel to octochannel memory, probably.
|
|
#
?
Jun 26, 2018 20:09
|
|
|
A German webshop has the TR2 listed for 1509 Euros. No word if it's official.
|
|
#
?
Jun 27, 2018 17:03
|
|
|
Here's da link! https://www.cyberport.de/?DEEP=2003-006&APID=14 --edit: This could possibly put the 16C one at the current 1950X pricepoint, and maybe, just maybe, the 24C will come at the introductionary price of the 1950X (i.e. around a 1000€). vvv Maybe so, but that's the link those 1509€ come from. A lot of placeholder copypasta I figure. Combat Pretzel fucked around with this message at 17:08 on Jun 27, 2018 |
|
#
?
Jun 27, 2018 17:05
|
|
|
It has the wrong features listed: 16 cores & 32 threads. Edit: ^^ true, it just stood out is all. Mr Shiny Pants fucked around with this message at 17:12 on Jun 27, 2018 |
|
#
?
Jun 27, 2018 17:07
|
|
|
Paul MaudDib posted:(I consider a sample size of 1 to be quite weak evidence anyway, and that's all I've seen so far on kill-ryzen. Then why do you keep posting about it in multiple threads as if it's an actual problem anyone is likely to run into? You should maybe go back and edit your posts.
|
|
#
?
Jun 27, 2018 17:14
|
|
|
Those 12C Threadrippers were 3/3 + 3/3, and each die had the full L3 cache available? Just wondering, if the 24C TR2 is a good middle between lots of cores and OK memory bandwidth (losing double bandwidth between dies), specifically to the dies without memory lanes.
|
|
#
?
Jun 27, 2018 17:39
|
|
|
Yeah - from Anandtech's page, the 1920X is (3+3)+(3+3) and each die has the full cache available. Curiously it looks like the 1900X is (2+2)+(2+2) but has half the cache disabled. The 24-core may just double that and have 4 dies with 6 cores each, but as far as I know there's nothing stopping them from using 3 full dies and only having one without a direct memory controller connection. Seems like that design would have superior performance for most purposes and the only reason to go for 4 cut-down dies instead would be if AMD has so many dies coming out with 6/7 working cores that they need a way to use them up other than 2600[X]s and 12-core TR2s.
|
|
#
?
Jun 27, 2018 18:35
|
|
|
What material is the Threadripper heatspreader made of (or any IHS in general)? Considering going with liquid metal.
|
|
#
?
Jun 29, 2018 16:41
|
|
|
Combat Pretzel posted:What material is the Threadripper heatspreader made of (or any IHS in general)? Considering going with liquid metal. edit: oh, between the IHS and HSF, I get it now. edit2: Turns out the TR IHS is made of aluminum. Abort, abort, do not use LM. eames fucked around with this message at 16:57 on Jun 29, 2018 |
|
#
?
Jun 29, 2018 16:48
|
|
|
Someone on Reddit tried using liquid metal on top of the Threadripper IHS and it didn't end well https://www.reddit.com/r/buildapc/comments/8779mc/do_not_use_liquid_metal_or_conductonaut_on_a/
|
|
#
?
Jun 29, 2018 16:53
|
|
|
Ah, too bad. I read that it made a significant difference. Also, I just read that the AMD Ryzen IHS are supposedly nickel plated. I guess not.
|
|
#
?
Jun 29, 2018 17:04
|
|
|
eames posted:edit2: Turns out the TR IHS is made of aluminum. Abort, abort, do not use LM. I don't think it is, that looks like an oxidation layer. If it were aluminum, the gallium would have invaded the whole thing, it would have literally crumbled to bits within 24 hours. It's probably some tin in the alloy or something. Incidentally, that's pretty much what the inside of my 4770k's IHS looks like after 3 years of liquid metal contact and it hasn't lost any thermal performance in that time. It's not worth it either way for threadripper, you generally only get a couple degrees benefit when it's on top of the IHS, then you get to worry that some tiny glob will migrate its way somewhere important and murder your whole system.
|
|
#
?
Jun 29, 2018 17:11
|
|
|
Is "Dual Intelligent Processors 5" from ASUS' AI Suite something that runs in *tandem* with the BIOS settings, or does it go and do its own thing? I don't want to uninstall it if it's needed, but my Overwatch keeps crashing and I'm trying to look for suspects. Should I also set my AI Overclocking Tuner setting to "Auto" instead of "Manual" in BIOS? I assume it's already automatic since it's 2700X, but I have no idea. ufarn fucked around with this message at 17:28 on Jun 29, 2018 |
|
#
?
Jun 29, 2018 17:21
|
|
|
I forgot those threadrippers have so many memory channels (though it makes sense if they are going to be handling lots of memory, I'm just not familiar with normal workstation/enthusuast processors). Is it possible that quad channel memory might one day become a more common/cheaper feature? It seems like it would really help apus, if the biggest barrier to performance for them is bandwidth. For that matter, will the coming of DDR 5 RAM matter to those much, or will the increased transfer rates not be enough to matter (due to there being other limitations like actual latency)?
|
|
#
?
Jun 29, 2018 19:26
|
|
|
More channels is more difficult to lay out on the motherboard, and by nature requires more modules, so there will certainly be a premium going forward.
|
|
#
?
Jun 29, 2018 19:30
|
|
|
The real solution is an L4 Cache, be it eDRAM, HBM or AMD's Near Memory.
|
|
#
?
Jun 29, 2018 21:18
|
|
|
ufarn posted:Is "Dual Intelligent Processors 5" from ASUS' AI Suite something that runs in *tandem* with the BIOS settings, or does it go and do its own thing? I don't want to uninstall it if it's needed, but my Overwatch keeps crashing and I'm trying to look for suspects. Try to run some memory tests, Quake Champions kept crashing for me and it turned out my ran could not run at 3200 in a stable manner and it just so happened that QC would hit my ram in a way that made it fall over.
|
|
#
?
Jun 30, 2018 12:33
|
|
|
TheCoach posted:Try to run some memory tests, Quake Champions kept crashing for me and it turned out my ran could not run at 3200 in a stable manner and it just so happened that QC would hit my ram in a way that made it fall over. For now, I've disabled all the Xbox/GFW/Game Mode crap which sounded like a cause of similar issues for other people. Also deleted the ASUS software since it was confusing anyway. I'll remember to look at RAM speed for debugging if it keeps happening after this.
|
|
#
?
Jun 30, 2018 13:11
|
|
|
One thing to note about memory bandwidth is that a single core is unable to saturate the total available bandwidth on its own; the real benefit of quad (or more) channel memory only applies when a lot of cores are involved. See this for a relatively in-depth explanation.
|
|
#
?
Jun 30, 2018 17:36
|
|
|
FaustianQ posted:The real solution is an L4 Cache, be it eDRAM, HBM or AMD's Near Memory. It annoys me that kaby lake g cannot use the 4gb of hbm as llc
|
|
#
?
Jun 30, 2018 17:42
|
|
|
ufarn posted:Seems crazy that it wouldn't be able to handle Trident Z's at 3200, though, especially on a decent motherboard like the Crosshair VII Hero. Maybe it can but at worse timings. I just know that relaxing the timings and dropping ram to 3000 helped my B350 1700 system to become stable.
|
|
#
?
Jun 30, 2018 17:54
|
|
|
If BIOS knows best, would it be safe to set the memory frequency to AUTO and the AI overclocking profile to DOCP/DEFAULT rather than 3200 and DEFAULT? I'm less concerned about 3000 vs 3200 than the low frequencies I seem to get with AUTO mode, but maybe it's scaling fine during high loads without me being able to notice?
|
|
#
?
Jun 30, 2018 21:13
|
|
|
I'd look at the latencies/subtimings and go with whichever results in lower.
|
|
#
?
Jun 30, 2018 21:18
|
|
|
ufarn posted:If BIOS knows best, would it be safe to set the memory frequency to AUTO and the AI overclocking profile to DOCP/DEFAULT rather than 3200 and DEFAULT? if AUTO is setting memory frequency to under 3000, you may need to also enable XMP or something like that to use the technically-out-of-JDEC-spec higher frequencies. this is kinda hard to help with because every mobo maker labels their poo poo differently. also I'm not sure whether you're trying to OC things or just run the best defaults?
|
|
#
?
Jul 1, 2018 01:19
|
|
|
Klyith posted:if AUTO is setting memory frequency to under 3000, you may need to also enable XMP or something like that to use the technically-out-of-JDEC-spec higher frequencies. DOCP is AMD for XMP as far as I understand, although it's only there for "AI Overclock Tuning", a general profile rather than for memory specifically. This is what the BIOS looks like, except I don't have multiple DOCP profiles to choose from: https://www.youtube.com/watch?v=RZRjoeyz4Z0&t=357s
|
|
#
?
Jul 1, 2018 01:34
|
|
|
Have you enable the XMP Profile?
|
|
#
?
Jul 1, 2018 14:52
|
|
|
I just bought a Fury X to gently caress about with because it looks pimp, and I need a small card with Freesync for an ITX case. I know it's a hungry beast but I can undervolt / clock it. I'm cool with that. Has anyone here done this and checked with a Kill a Watt?
|
|
#
?
Jul 1, 2018 15:39
|
|
|
GRINDCORE MEGGIDO posted:I just bought a Fury X to gently caress about with because it looks pimp, and I need a small card with Freesync for an ITX case. IIRC the Fury cards undervolt well - they were pushed pretty hard at stock to try to keep up with the 980 and that made them super hungry.
|
|
#
?
Jul 1, 2018 15:42
|
|
|
GRINDCORE MEGGIDO posted:I just bought a Fury X to gently caress about with because it looks pimp, and I need a small card with Freesync for an ITX case. Here’s a reddit post I made with my nitro fury undervolting and underclocking it. Measured wattage at Wall with kill a watt. https://www.reddit.com/r/Amd/comments/56ncbi/sapphire_fury_nitro_undervolting_and_underclocking/
|
|
#
?
Jul 1, 2018 17:55
|
|
|
B-Mac posted:Here’s a reddit post I made with my nitro fury undervolting and underclocking it. Measured wattage at Wall with kill a watt. That's interesting, this test: 900/500 -72mV Temp ~ 65C Total score: 11958 Graphics score: 13782 Firestrike Wattage:200W Furmark Wattage: 290W That's a massive drop for not much drop in performance, isn't it? How does it compare to a 580 at that setting? E - oh you mentioned 390x, that's not far off. This is good, I'll post a bit more about it when the card arrives. I'm stoked TBH. It's such a strange folly of a card. GRINDCORE MEGGIDO fucked around with this message at 18:43 on Jul 1, 2018 |
|
#
?
Jul 1, 2018 18:40
|
|
|
GRINDCORE MEGGIDO posted:I just bought a Fury X to gently caress about with because it looks pimp, and I need a small card with Freesync for an ITX case. How does the AIO cooler work in an ITX case?
|
|
#
?
Jul 2, 2018 02:05
|
|
|
Depends on the case. ITX is nebulous enough that it spans everything from "just barely can't fit an micro-ATX board" to "needs a riser cable for the GPU."
SwissArmyDruid fucked around with this message at 02:37 on Jul 2, 2018 |
|
#
?
Jul 2, 2018 02:32
|
|
|
If the Fury X actually fits into your ITX case, it'll probably be pretty nice. You can either dump heat right out, or suck cool air right into the case, either way it's going to run pretty cool.
|
|
#
?
Jul 2, 2018 06:59
|
|
|
Case is a nzxt h200, should be no problem. I'm stoked for das blinkenlights. Think I might use it as an exhaust.
|
|
#
?
Jul 2, 2018 07:29
|
|
|
GRINDCORE MEGGIDO posted:
Almost all GPUs and many CPUs work like this these days. Their power/performance efficiency peak is close to or below the base clock but manufacturers then spec higher turbo clockspeeds for more performance at the cost of additional heat. GPU board manufacturers often add a huge heatsink/fan and "factory overclock" on top of that which trades another 5% performance for 15% power consumption to look good in benchmarks. That and more sophisticated boost algorithms are the reasons why overclocking is eventually going away unless temperatures/power consumption are no concern at all. Speaking of power efficiency and all that, Buildzoid mentioned that 32C Threadripper will have a stock Vcore of 1.05V.
|
|
#
?
Jul 2, 2018 08:27
|
|
|
eames posted:Speaking of power efficiency and all that, Buildzoid mentioned that 32C Threadripper will have a stock Vcore of 1.05V. we don't need to imagine what it would look like running at the normal ryzen 1.35v, because james mickens has already imagined it for us: quote:John began to attend The Church of the Impending Power Catastrophe. He sat in the pew and he heard the cautionary tales, and he was afraid. John learned about the new hyper-threaded processor from AMD that ran so hot that it burned a hole to the center of the earth, yelled “I’ve come to rejoin my people!”, discovered that magma people are extremely bigoted against processor people, and then created the Processor Liberation Front to wage a decades-long, hilariously futile War to Burn the Intrinsically OK-With-Being-Burnt Magma People. The future was bleak, and John knew that he had to fight it. So, John repented his addiction to scaling, and he rededicated his life to reducing the power consumption of CPUs. It was a hard path, and a lonely path, but John could find no other way. Formerly the life of the party, John now resembled the scraggly, one-eyed wizard in a fantasy novel who constantly warns the protagonist about the variety of things that can lead to monocular bescragglement. At team meetings, whenever someone proposed a new hardware feature, John would yell “THE MAGMA PEOPLE ARE WAITING FOR OUR MISTAKES.”
|
|
#
?
Jul 2, 2018 08:50
|
|
|
eames posted:Almost all GPUs and many CPUs work like this these days. Ya. I've been noticing the wall where voltage / heat starts increasing rapidly since clocking a k62. I'm really surprised AMD chased that last 100 MHz so hard. I'm quite taken by how bizarre this card is.
|
|
#
?
Jul 2, 2018 12:45
|
|
|
Being honest, the problem is that the Fury wasn't quite enough to beat the 980 on equal footing... nor the 980 Ti on equal footing. It can't quite keep up with the 980 on efficiency, and the 980 Ti blows it out of the water on performance (when overclocked). Repeat with the 480-vs-1060, and the V64-vs-1080 matchups. Like, it's not bizarre, it's just that AMD couldn't bear to suffer through the review slugline "5-10% below the equivalent NVIDIA card", just repeat that times three different cards. Or at least that was RTG under Raja. If you want to get down to it... Vega, in its sweet-spot, is not really that much faster than a Fury in its high-end performance. It's just that the Fury, when clocked down into its sweet-spot, is actually more like a 980/1060. Or, a 390X in its high-end performance state, which if you clock it down... Paul MaudDib fucked around with this message at 13:03 on Jul 2, 2018 |
|
#
?
Jul 2, 2018 12:57
|
|
|
|
| # ? Apr 29, 2024 11:07 |
|
|
The whole thing seems so premium for nothing. What a folly. Hbm, really nice vrms, semi custom watercooling with a Nidec fan, led performance indicators..... I bought it for less then a 1060, freesync is a bonus. It looks totally pimp, I do have to say. GRINDCORE MEGGIDO fucked around with this message at 13:08 on Jul 2, 2018 |
|
#
?
Jul 2, 2018 13:03
|
|