![]() |
|
|
#111 |
|
Apr 2019
CD16 Posts |
Speaking of system tuning, I saw this video recently (from the "Actually Hardcore Overclocking" channel) which I found interesting regarding getting the most out of memory clocks on Zen 2.
https://www.youtube.com/watch?v=xSAFs-t0DCg Basically he recommends running FCLK at its max capable rate, (unbound from UCLK) if the difference would be greater than 166MHz: FCLK - UCLK > 166MHz. Because you get better bandwidth across CCXs at the expense of some potential latency to RAM. So with 3000 MT/s RAM that user's UCLK should be at 1500MHz, and according to the video he should be capable of at least 1766 FCLK, so that would be 266MHz difference therefore worth separating them (theoretically). I wonder how well it works for prime95 though. I don't have a Zen 2 to play with though, only "Zen+" 2000 series. Last fiddled with by hansl on 2019-09-01 at 22:40 |
|
|
|
|
|
#112 |
|
Feb 2016
UK
1C016 Posts |
Haven't watched that particular video, but I had also seen elsewhere suggesting if you have slow(er) ram, running the max IF speed possible can still lead to gains for gaming. I'm not sure it is only inter-CCX communication benefiting from it. When AMD revealed that each CCD only had half the write bandwidth compared to read, that would affect ram also. Unless you run single channel, you're not likely to ever saturate ram with writes on a single CCD CPU as the IF will choke it long before then. Any increase to IF speed will help with that, assuming bandwidth is more limiting than latency.
I haven't tested it myself yet. For now 1800 on IF seems a safe bet, but people aren't getting much further than that. I forget what I saw on one of mine, but I didn't reach 1900. |
|
|
|
|
|
#113 |
|
"Composite as Heck"
Oct 2017
95010 Posts |
Threadripper 3 speculation on splitting into consumer (sTRX4 HEDT) and workstation (sWRX8 workstation) lines:
https://youtu.be/1Xrys7sOr4A?t=55 tl;dr:
Threadripper hardware being for consumers means it's easier for us to obtain so that's nice, but I'm not as optimistic as the guy in the video that these workstation SKUs will be a "poor mans Epyc". Core for core the performance should beat the P line of Epyc 2 because of the TDP (unless the bins are much worse which I doubt). It wouldn't surprise me if the price for a workstation TR3 part is between the core equivalent P and non-P Epyc 2 parts. Still it's an interesting development, a 64 core workstation TR3 part if it exists would be top CPU in terms of throughput. |
|
|
|
|
|
#114 |
|
"Composite as Heck"
Oct 2017
2×52×19 Posts |
A new top-end Epyc part has been announced ( https://www.phoronix.com/scan.php?pa...=AMD-EPYC-7H12 ), it's like the 7742 with 64 cores but comes with a higher TDP of 280W and a higher base clock. The 280W TDP is interesting as it matches what is expected of the Threadripper parts, however it makes me even less optimistic about how price-competitive the TR parts will be. Maybe this move indicates TR3 will top out at 48 cores (supply is constrained, it'd be low volume anyway as it'll be hideously expensive for a workstation part and this way zen3 can have a natural progression to 64 cores next year).
|
|
|
|
|
|
#115 |
|
"Composite as Heck"
Oct 2017
2·52·19 Posts |
https://www.tomshardware.com/news/am...3750x-105w-cpu
tl;dr 3750X is a potential Zen2 part, many unknowns including if it'll even be for general consumption, would slot between the 3700X and 3800X with a 105W TDP like the 3800X. I'm only bothering to mention it because one of the possible configurations is two quad core chiplets which would have a tasty 64MB of cache. |
|
|
|
|
|
#116 |
|
Feb 2016
UK
26·7 Posts |
As interesting a 2 CCD part would be, feels more intuitive for a 3750X to literally sit between 3700X and 3800X. Not that there is much room to squeeze another part in there, but might be another OEM special.
If they had gone 2 CCD, I think that would be more likely a 3850X. |
|
|
|
|
|
#117 |
|
"Composite as Heck"
Oct 2017
2·52·19 Posts |
You're probably right about a 1 CCD OEM special. Still, a 2 CCD part would not clock as high as a 3800X all other things equal (and quad core bins are likely much worse than normal reducing clocks further) to the point where the 3750X name might make sense. It may perform more like a poor mans Epyc than a flagship part.
An interesting wrinkle is that the PS5 and next Xbox will have 8 core zen2 CPUs. These may be in a 2 CCD low clock configuration to soak up the low-binned parts just like the PS4 did for the Jaguar cores before. If the console stockpile demand is already being satisfied they may be looking for another outlet. |
|
|
|
|
|
#118 | |
|
1976 Toyota Corona years forever!
"Wayne"
Nov 2006
Saskatchewan, Canada
123158 Posts |
Quote:
|
|
|
|
|
|
|
#119 |
|
"Sam Laur"
Dec 2018
Turku, Finland
1001111012 Posts |
From message 103 in this thread (that later graph was just an update of that first benchmark, with 3900X results added) :
As a baseline comparison, the lowest graph curve is the now retired Ryzen 3 2200G, four cores, one worker. Speeds normalized by multiplying FFT length (in K) with throughput (iters/sec), then dividing that value by the slowest such result, which happened to be 8064K FFT on the 2200G processor. |
|
|
|
|
|
#120 |
|
"Composite as Heck"
Oct 2017
2·52·19 Posts |
The 3950X, 3960X and 3970X are released November 25th (16, 24, 32 cores respectively for $750, $1400, $2000 respectively on platforms AM4, TR40, TR40 respectively). 3960X and 3970X are within a few hundred dollars of comparably specced (wrt cores/cache) Epyc parts, with 128MB of L3 and a 280W TDP that should probably be liquid cooled or underclocked. The TR40 platform is for HEDT enthusiasts and will likely be expensive and high quality across the board, not backwards compatible with TR4.
https://www.anandtech.com/show/15062...19-fall-update https://videocardz.com/82654/amd-ann...-ryzen-9-3950x https://www.techpowerup.com/260974/m...s-motherboards |
|
|
|
|
|
#121 |
|
"Composite as Heck"
Oct 2017
16668 Posts |
"Eco mode" is an interesting feature that comes with the AM4 1004 bios update. It makes any zen2 105W TDP part stick to a 65W TDP, and any 65W TDP part stick to a 45W TDP. It may be the best and simplest way to underclock for some of our needs, any traditional overclock doesn't do a great deal for zen2 and the same is probably true for an underclock.
I have a feeling it'll help 105W parts fit the efficiency curve better than it helps 65W parts. Worth investigating at any rate. |
|
|
|
![]() |
Similar Threads
|
||||
| Thread | Thread Starter | Forum | Replies | Last Post |
| RX470 and RX460 announced | VictordeHolland | GPU Computing | 0 | 2016-07-30 13:05 |
| Intel Xeon D announced | VictordeHolland | Hardware | 7 | 2015-03-11 23:26 |
| Factoring details | mturpin | Information & Answers | 4 | 2013-02-08 02:43 |
| Euler (6,2,5) details. | Death | Math | 10 | 2011-08-03 13:49 |
| Larrabee instruction set announced | fivemack | Hardware | 0 | 2009-03-25 12:09 |