|
Gamers Nexus has Threadripper drawing less power than 7900X at stock http://www.gamersnexus.net/hwreviews/3015-amd-threadripper-1950x-1920x-review/page-4 Hardware Unboxed also has TR drawing less. I've noticed some other discrepancies between reviews, but I guess we'll get plenty of follow-up testing in the coming months to get to the bottom of it
|
#
?
Aug 10, 2017 21:11
|
|
|
|
| # ? Apr 20, 2024 01:34 |
|
|
The AMD review slides show that TR dies are the very top 2-5% of all zen dies which means 1800X units were binned with a 4.0-4.1 Ghz ceiling all along. I mean the 1800X might have been worth it if you had a shot at 4.2-4.3Ghz at 1.45V once in a blue moon but we thought those chips don't exist; turns out they do but they all went into TR.
|
|
#
?
Aug 10, 2017 21:15
|
|
|
TR reviews have the least consistent results measured by different review outlets that I have ever seen in about 15 years of following CPU reviews. Once the smoke clears I'm really curious what's going on.
|
|
#
?
Aug 10, 2017 21:16
|
|
|
The power stuff seems to vary pretty greatly depending on where/how you are measuring. AT seemed to use the self reported numbers, GN measured at the rail, OC3D did at the wall. For me, at the wall makes the most sense, mainly because im interested in how the whole platform is performing cause its new.
|
|
#
?
Aug 10, 2017 21:22
|
|
|
Cygni posted:The power stuff seems to vary pretty greatly depending on where/how you are measuring. AT seemed to use the self reported numbers, GN measured at the rail, OC3D did at the wall. For me, at the wall makes the most sense, mainly because im interested in how the whole platform is performing cause its new. So did OC3D forget they had like 3x 1080 Ti's in it or...
|
|
#
?
Aug 10, 2017 21:28
|
|
|
The boards themselves, at least the ASUS Zenith, seem to eat a lot of power too. Gamers Nexus 1950x is drawing 10-30w less power than their 7900x on the EPS12v. The image is power over time while streaming and recording CSGO through two different applications.
|
|
#
?
Aug 10, 2017 21:42
|
|
|
You know what, that OC3D power measurement just has to be an error. Everyone else has OC'd TR at ~300W package power. Even if AMD is cheating the onboard sensor a little bit, even if their voltage is higher than others (almost 1.42v), there's got to be something else going on there to explain ~250W of missing TDP. Or else they weren't kidding about 3V to the core 
|
|
#
?
Aug 10, 2017 23:57
|
|
|
Eh, looking all over the benchmarks, the jack of all trades thing seems to be more the 7820X. I'd love ECC and I don't remotely render as much as needed to justify the TR, so I'm sitting it out to TR2 and an hopefully async and souped up IF.
|
|
#
?
Aug 11, 2017 00:20
|
|
|
My stopgap until Zen2 is probably going to be hoping for a flood of super cheap second hand 7700Ks and Z270 boards. Gotta get off this i7-3820.
|
|
#
?
Aug 11, 2017 00:21
|
|
|
SlayVus posted:According to AMD's information, game mode turns on legacy mode. When in game mode a 16c/32t will run at 8c/16t on one die. A 12c/24t also becomes 6c/12t. So you're saying this is wrong? Anandtech posted:The two switches are:
|
|
#
?
Aug 11, 2017 00:28
|
|
|
tijag posted:So you're saying this is wrong? The issue was the other poster was saying that game mode turned off SMT. In game mode, with Ryzen Masters software, a whole physical die is disabled giving you 8 cores and 16 threads. If you were to disable SMT under game mode, you would only have 8 cores and 8 threads. Game Mode, by default enables legacy mode and sets the CPU to NUMA. This mode has SMT enabled and one physical die of the CPU is disabled. Making it run exactly like a Ryzen 7 1900x. SlayVus fucked around with this message at 01:06 on Aug 11, 2017 |
|
#
?
Aug 11, 2017 01:03
|
|
|
tijag posted:So you're saying this is wrong? SlayVus posted:The issue was the other poster was saying that game mode turned off SMT. There are two settings that might be getting mixed up here. The other interesting toggle is Local vs Distributed memory mode. When you're in legacy mode, half of your memory is mapped to a die you're not using, and this toggles how the memory controller handles that. Distributed will maximize bandwidth (which should help for newer titles) and Local will minimize latency. Depending on how they implement this there's a couple different possible performance scenarios. Is Local Mode like hooking all 4 DIMMs into a single die in dual-slot mode? So basically you're going down to 2 channel, just like literally a Ryzen 7? Or are they still hooked into both dies and there's now a "fast segment" and a "slow segment" (near vs far memory) that have different latency? Or, do you only use the memory controller on the current die and lose half of your RAM capacity? Same questions for PCIe, does this cut your lanes or cause some of them to run over a slow NUMA link? And with Distributed Mode, is there a fast segment/slow segment with "best effort" delivery, or do they just levellize everything to the longer of the two to simplify instruction scheduling/etc? How much higher is bandwidth in Distributed Mode, and how much lower is the latency in Local Mode? When you figure that half of a Threadripper is a fully-functional Ryzen 7 it raises some interesting possibilities for implementation. I actually legitimately want to read the programmers guide on this, and I can't wait for Agner Fog's take either both on full mode and legacy mode. The main goal of all this is to get older non-NUMA-aware applications to behave - don't believe the crap about "oh one game crashes with >20 threads", they didn't do two microcode toggles over a single game. NUMA is weird and depending on the exact threading/memory access patterns of the title it may not play nicely - but these give you the ability to lock it into basically being a slightly slower Ryzen 7 with some kind of reasonable behavior for the application. The binning should help with that some too, although I agree that there does seem to be some surprising variation in clocking given that this is supposed to be the top 5% bin. Paul MaudDib fucked around with this message at 02:08 on Aug 11, 2017 |
|
#
?
Aug 11, 2017 01:03
|
|
|
Paul MaudDib posted:Legacy mode locks you to one die with SMT disabled is what I've seen most places. You can manually disable SMT in full mode but SMT is always disabled in Legacy Mode regardless. SlayVus posted:https://community.amd.com/community/gaming/blog/2017/08/10/amd-ryzen-threadripper-for-gaming 
|
|
#
?
Aug 11, 2017 01:12
|
|
|
This looks like a last minute change from somebody, honestly. We know they've been scrambling to get working BIOS out for both Epyc and TR so maybe somethin changed right at the end. Disabling SMT is much more attractive than just disabling a whole friggen die to me, i dunno. Also if you disabled a die, why bother having NUMA on? Unless you only disabled the cores and not the memory controller?
|
|
#
?
Aug 11, 2017 01:24
|
|
|
Cygni posted:Disabling SMT is much more attractive than just disabling a whole friggen die to me, i dunno. Also if you disabled a die, why bother having NUMA on? Unless you only disabled the cores and not the memory controller? The problem isn't "a game sees too many cores and freaks out". AMD doesn't actually give a poo poo about that, that's marketing spin. One title does not justify a whole huge microcode mode. It's about games that just poo poo themselves because they thrash the cache on a NUMA architecture (when you see speedups, that's where the gains are coming from). You see the same thing on Intel systems, games just don't like NUMA. Frankly it's also probably about applications too - you just know that someone's workstation utility or legacy app isn't gonna like suddenly running on NUMA, but this way they put the focus on the gamers rather than the enterprise market. I'm not too worried about the server market, that stuff will be NUMA optimized, but workstation applications may not be, a LGA2011 or 1366 was the standard and most workstation stuff is single-die. There may also be weird legacy apps (custom, etc) that hit pathological thrashing behavior, and those could potentially drop to a fraction of even Ryzen 7 performance, like 25% or less of expected performance. So, they've given you a way to lock it to what more or less behaves like a Ryzen 7 with some (selectable) reasonable memory behavior. Can't thrash over the interconnect if the other die is turned off. /taps forehead Because Threadripper is the same die as Ryzen 7 I think they have a lot of implementation options, and I'm not sure we know the details on how they did it. Can they shunt all four traces to one die and run clamshell mode? If so they could literally be just one Ryzen 7 for all intents and purposes. If not you now have to figure out what happens when you try to get to memory that's on the other die. That's the usual problem in a NUMA system, part of your memory is farther away than others, it's non-uniform. And not everything likes to run on a system like that. edit: I guess it does not necessarily disable SMT, that's good then Paul MaudDib fucked around with this message at 02:02 on Aug 11, 2017 |
|
#
?
Aug 11, 2017 01:35
|
|
|
Cygni posted:Disabling SMT is much more attractive than just disabling a whole friggen die to me, i dunno. Also if you disabled a die, why bother having NUMA on? Unless you only disabled the cores and not the memory controller? The reason for the for disabling could just be the simple fact that they feel the die to die latency is too high. Even though it's only 250ns on 2400Mhz RAM and 200ns on 3000Mhz ram. If you're running in Legacy mode your threadripper is essentially just a Ryzen 7 1900x with quad channel memory and 64 PCI-E lanes. Games/apps that are already optimized for Ryzen will than have no issue on Legacy mode. https://youtu.be/Hn96ty9D0aQ?t=278s The graph does the core to core latency numbers. SlayVus fucked around with this message at 02:03 on Aug 11, 2017 |
|
#
?
Aug 11, 2017 01:51
|
|
|
 https://twitter.com/NVIDIAGeForce/status/895746289589039104
|
|
#
?
Aug 11, 2017 02:02
|
|
|
AMD replied with a gif of man in a green shirt high fiving a man in a red shirt.
|
|
#
?
Aug 11, 2017 02:05
|
|
|
The 12C/24T 1920X IMO is shmuck bait. Who in the right mind looking for an already expensive super-MT CPU would want to save 20% for 25% less threads?
|
|
#
?
Aug 11, 2017 02:13
|
|
|
Palladium posted:The 12C/24T 1920X IMO is shmuck bait. Who in the right mind looking for an already expensive super-MT CPU would want to save 20% for 25% less threads? Put that 20% towards faster memory or an SSD.
|
|
#
?
Aug 11, 2017 02:16
|
|
|
Someone has gotta buy those defective cores.
|
|
#
?
Aug 11, 2017 02:18
|
|
|
NVIDIA sells (sold?) prebuilt "battleboxes" with Ryzens. NVIDIA doesn't give a gently caress, they're not really competitors in this market. In fact they benefit from people impulse-splurging on a super-rig, and I'm sure AMD is seen as a valuable counterweight to the actual market threats, like Intel, Samsung, Google, perhaps Qualcomm... (Arguably they're hardly even competitors in the graphics market either. How long till Volta launches?)
|
|
#
?
Aug 11, 2017 02:18
|
|
|
They aren't wrong.
|
|
#
?
Aug 11, 2017 02:19
|
|
|
Palladium posted:The 12C/24T 1920X IMO is shmuck bait. Who in the right mind looking for an already expensive super-MT CPU would want to save 20% for 25% less threads? "threadripper: starting at only $799" --every ad for the next month
|
|
#
?
Aug 11, 2017 02:23
|
|
|
Threadripper starts at $599
|
|
#
?
Aug 11, 2017 03:21
|
|
|
wargames posted:They aren't wrong. In some parallel universe, Raja tells RTG to release the real Vega drivers.
|
|
#
?
Aug 11, 2017 03:24
|
|
|
Don Lapre posted:Threadripper starts at $599 Oh, interesting, what's the launch date? Same as the 1920X and 1950X? With 8 cores it's only useful in a "just enough to light up the platform" build, unless it has a massive amount of cache or some other redeeming quality. Otherwise you'd be better off with Ryzen 7, which is cheaper, has the same number of threads, and no NUMA weirdness. It seems like a pretty narrowly tailored SKU. Also in general you're already buying a stepping we know to be older. That's one thing on a $200 consumer processor but when you're talking about dropping $800-1000 it gives me a little bit of a heartburn. How long until B2 trickles down to Threadripper?
|
|
#
?
Aug 11, 2017 03:27
|
|
|
Threadripper 8 core does make way more sense then x299 kabylake-x only because pci-e lanes.
|
|
#
?
Aug 11, 2017 03:33
|
|
|
Considering Threadripper already got all the binned dies, what big improvement are you expecting from improved stepping? Surely any major process improvement would just go to a new SKU. And with 2 years of socket support, if you need the IO more than anything else, you'll get those new SKUs/stepping if the base model will do for the time being.
|
|
#
?
Aug 11, 2017 03:36
|
|
|
The appeal of the octocore TR part is the idea of being able to upgrade in the next generations, along with the issues that Ryzen has with 16GB DIMM clock speeds and wanting 64GB for VMs.
|
|
#
?
Aug 11, 2017 03:38
|
|
|
Yeah im really torn now as if the 1900x or what ever they call it is just like a 1800x but with more pcie lanes etc i would buy that But if its a 4c/4c split over the interdie then rip so much for a threadripper build that would set me up to transition to zen2 later on
|
|
#
?
Aug 11, 2017 03:46
|
|
|
wargames posted:Threadripper 8 core does make way more sense then x299 kabylake-x only because pci-e lanes. Also quad channel RAM.
|
|
#
?
Aug 11, 2017 03:49
|
|
Scarecow posted:Yeah im really torn now as if the 1900x or what ever they call it is just like a 1800x but with more pcie lanes etc i would buy that It is 4c/4c split, AMD have said this, just like how the 1920x is 6c/6c instead of 8c/4c.
|
|
|
#
?
Aug 11, 2017 03:56
|
|
|
Well bummer that sucks but i can understand why tho
|
|
#
?
Aug 11, 2017 03:57
|
|
|
AVeryLargeRadish posted:It is 4c/4c split, AMD have said this, just like how the 1920x is 6c/6c instead of 8c/4c. God, I'd kill for a 4c/8c part.
|
|
#
?
Aug 11, 2017 04:00
|
|
|
Scarecow posted:Yeah im really torn now as if the 1900x or what ever they call it is just like a 1800x but with more pcie lanes etc i would buy that How could it possibly be a single die and have quad channel memory?
|
|
#
?
Aug 11, 2017 04:04
|
|
|
PerrineClostermann posted:How could it possibly be a single die and have quad channel memory? There are two memory controller chips per die.
|
|
#
?
Aug 11, 2017 04:07
|
|
|
SlayVus posted:There are two memory controller chips per die. Wait, how is that set up? I thought they were literally Ryzen chips linked up. Double dual channel gives you quad. Is that not correct?
|
|
#
?
Aug 11, 2017 04:08
|
|
|
PerrineClostermann posted:Wait, how is that set up? I thought they were literally Ryzen chips linked up. Double dual channel gives you quad. Is that not correct? Ryzen isn't fully enabled, there's also four PCIe lanes they're clipping off too.
|
|
#
?
Aug 11, 2017 04:14
|
|
|
|
| # ? Apr 20, 2024 01:34 |
|
|
PerrineClostermann posted:Wait, how is that set up? I thought they were literally Ryzen chips linked up. Double dual channel gives you quad. Is that not correct? http://images.anandtech.com/doci/11697/threadripper_architecture-final_16.jpg I think the only way we'll see a 8c TR is with a 2+2 CCX + 2+2 CCX scenario. This would make it still capable of quad channel memory across all 8 dimm slots.
|
|
#
?
Aug 11, 2017 04:15
|
|