|
I remember when Zen 1 came out they claimed that AM4 would last through 2020 so I'd assume yeah unless that's changed.
|
#
?
Nov 12, 2018 18:35
|
|
|
|
| # ? Apr 20, 2024 05:21 |
|
|
Doubt zen3 will have pcie4 then since it's really much harder to deal with that pcie3 unless they future proofed the pcb and sockets. Also it's pointless for home uses unless you desperately need massive bandwidth to your ssd or gpu
|
|
#
?
Nov 12, 2018 18:39
|
|
|
Devian666 posted:It's an old xeon but for the models I'm running I'm not hitting any memory bound issues. I'm running FDS and the fire calculations end up very intense. In the past I'm only seen memory become a problem on a very large model at the start of the run but once the fans start in the model the cpus would always sit at 100%. Even NIST who wrote the software suggest only running as many threads as cpu cores because memory latency isn't an issue.
|
|
#
?
Nov 12, 2018 18:43
|
|
|
Malcolm XML posted:Doubt zen3 will have pcie4 then since it's really much harder to deal with that pcie3 unless they future proofed the pcb and sockets. I could see AMD rolling out an "AM4+"/"TR4+" sockets that are just regular AM4/TR4 except the boards have to be validated to run at PCIe 4 speeds. Probably only high-end enthusiast boards would bother on the AM4 side, but I could see TR4 vendors adopting PCIe 4 as a standard feature for the next generation of boards. Mr.Radar fucked around with this message at 18:50 on Nov 12, 2018 |
|
#
?
Nov 12, 2018 18:46
|
|
|
Pablo Bluth posted:As I understand it, the usual CPU utilisation number shown doesn't tell if the CPU is memory bound. A CPU is 'busy' whether it is performing instructions or waiting for data from memory (because it's not in the CPU caches). Any CPU utilisation drop at the start of the run is usually a disk IO bottleneck when large runs can't be feed in to memory quickly enough. I'd eliminated disk bottleneck so that wasn't an issue as the model loads entirely into ram. In terms of cpu utilisation I've seen a talk by a coder for Netflix in relation to that. From testing I've done the halt while waiting for cache misses doesn't seem to be an issue either. e: AMD have suggested mixed integer and fp workload ipc is increased by29%. Devian666 fucked around with this message at 19:53 on Nov 12, 2018 |
|
#
?
Nov 12, 2018 19:46
|
|
|
Mr.Radar posted:I could see AMD rolling out an "AM4+"/"TR4+" sockets that are just regular AM4/TR4 except the boards have to be validated to run at PCIe 4 speeds. Probably only high-end enthusiast boards would bother on the AM4 side, but I could see TR4 vendors adopting PCIe 4 as a standard feature for the next generation of boards. PCIe 4 would need a new socket and pinout most likely, as the voodoo they do to get the signal to not poo poo out over 10 inches of busy and noisy motherboard isn't easy or cheap. They might do something wierd like first slot 4.0 rest are 3.0 to make it easier, but thats borderline RF engineering, and fuckkkkk that noise.
|
|
#
?
Nov 13, 2018 00:34
|
|
|
Craptacular! posted:Keep in mind this means selling some AM4 products until 2020. If, say, the halo products moved to a new socket while 4/6 core chips continued to use AM4, that technically counts as “support”.
|
|
#
?
Nov 13, 2018 01:25
|
|
|
I think again, it'll be an AM4/AM4+ thing, where the APUs trail behind on AM4 until 2020 and the CPUs move to AM4+, socket compatible but superior features and higher cost. Here is the AM4 socket pinout  There is a not insignificant number of reserved pins, so there still seems to be plenty of feature room for an AM4+ that differentiates itself from AM4. Obviously AM4+ CPUs will be backwards compatible with AM4 but lose out on features if you do so, such as dual PCIEx16 3.0 lanes, Quad Channel, eDRAM, dual 10Gbe or something - basically X570 will actually be worth it to upgrade. Around 2021 I can see everything getting moved to AM5, PCIE5 and DDR5 with the latest generation of APUs on 5nm to wow the gently caress out of everyone with last gen midrange performance for under 250$ or something. Then you'd get EPYC4 on a new socket followed by Ryzen 5000 series (DDR5, PCIE5, Gen5, one more 5 and AMD needs to source presentation music from Daft Punk) and TR 5000 series by late 2021. No I'm not going to toxx, just seems like a logical extension of AMD's strategy so far. Basically constantly confound Intel by forcing them to introduce features that've been at a premium tier for a lower cost and erode whatever high end advantages Intel purports to have.
|
|
#
?
Nov 13, 2018 03:18
|
|
|
I read somewhere that some version of AM4 will support ddr5 with Zen3, or that there may be two versions of Zen3 for ddr4 and ddr5 seeing as they can move IO off the core die. The latter seems unlikely. What I'm trying to say is that AM4 or a revision thereof is going to be with us for a long time. I for, one, miss the pentium 3 Sega cartridge.
|
|
#
?
Nov 13, 2018 06:16
|
|
|
Methylethylaldehyde posted:PCIe 4 would need a new socket and pinout most likely, as the voodoo they do to get the signal to not poo poo out over 10 inches of busy and noisy motherboard isn't easy or cheap. They won't need a new socket/pinout for PCIe 4 (or 5); the layout is exactly the same as 3 (and 2 and 1). They just keep increasing the bandwidth each gen while keeping everything else the same. This way they can keep all the hardware backwards/forwards compatible. The AM4 socket supports 24 lanes of PCIe; 4 of these lanes are reserved for the chipset (X470/B450/etc) and is what's used for all the I/O the chipsest handles (SATA,NIC,Extra PCIe lanes/etc). It's up to the board manufacturer how the other 20 lanes are used; generally they use 4 for an NVME connector and a pair of PCIex16 connectors that can do x16 or x8x8. When an add-in-board is plugged into one of these PCIe sockets the CPU and AIB auto-negotiate the fastest PCIe link speed they can both handle. Generally they both support PCIe3 so that's what's used but if you put in an older PCIe2 only card then the link speed will auto-negotiate down to that level and everything will 'just work'. It's no different with PCIe4; the CPUs will start supporting 4 as their highest link speed and will try to use that initially but auto-negotiate down to 3 if that's the fastest that the AIB supports. Theoretically if the PCIe traces between an x16 slot and the CPU on a current AM4 board are overbuilt enough to handle the increased bandwidth that PCIe4 brings then that board would start supporting PCIe4 when you dropped in a new CPU and AIB that both supported it. Now there's tons of business/engineering reasons why this won't happen; but it's theoretically possible.
|
|
#
?
Nov 13, 2018 12:30
|
|
|
Devian666 posted:I'm running FDS and the fire calculations end up very intense. Protip: intense fire coming from your workstation is not part of the calculation!
|
|
#
?
Nov 13, 2018 19:04
|
|
|
Krailor posted:They won't need a new socket/pinout for PCIe 4 (or 5); the layout is exactly the same as 3 (and 2 and 1). They just keep increasing the bandwidth each gen while keeping everything else the same. This way they can keep all the hardware backwards/forwards compatible. All this to say it's theoretically possible? It ain't happening.
|
|
#
?
Nov 13, 2018 19:59
|
|
|
Krailor posted:Theoretically if the PCIe traces between an x16 slot and the CPU on a current AM4 board are overbuilt enough to handle the increased bandwidth that PCIe4 brings then that board would start supporting PCIe4 when you dropped in a new CPU and AIB that both supported it. Now there's tons of business/engineering reasons why this won't happen; but it's theoretically possible. The difference between PCIe-3 and 4 is pretty substantial, going from 8Gbit/sec/pin to 16Gbit/sec/pin isn't trivial. A lot of really wonky poo poo starts to happen when your signaling frequency gets that high. The signal integrity might require additional ground pins connected directly to the cpu in order to get the lanes to work at the high speeds it needs. Or they might use repeaters or conditioners and eat the very tiny latency increase. No idea how the poor engineers are gonna make it work, just that it's probably not gonna be super duper fun to do it.
|
|
#
?
Nov 13, 2018 20:14
|
|
|
https://slickdeals.net/f/12249727-amd-ryzen-7-1700x-8-core-3-4ghz-desktop-processor-150-free-shipping?src=frontpage this is a good deal
|
|
#
?
Nov 13, 2018 20:32
|
|
|
Heres a partial answer to PCIe4 chat. First Rome specific board spotted in the wild.AnandTech posted:However, one of the new features of Rome is the use of PCIe 4.0. PCIe 4.0 has different standards for on-board signalling in order to get the required speed, so even though the processors are backwards compatible with PCIe 3.0, we expect new motherboards and new systems to be developed with PCIe 4.0 specifications in mind. https://www.anandtech.com/show/13596/first-amd-epyc-rome-motherboard-spotted
|
|
#
?
Nov 13, 2018 21:19
|
|
|
I guess that means PCIe extenders are dead now.
|
|
#
?
Nov 13, 2018 23:05
|
|
|
It might be an issue to use them at 4.0 speeds, but presumably they'll still work fine at 3.0 which will still be fast enough for many purposes for a long time.
|
|
#
?
Nov 13, 2018 23:25
|
|
|
Perplx posted:I guess that means PCIe extenders are dead now. your video card is perfectly happy without pcie 4.0
|
|
#
?
Nov 14, 2018 00:17
|
|
|
Perhaps more budget oriented boards deliver 8x pci 4 lanes to graphics cards and 2x to NVME. This decreases pin counts from the CPU, and allows for more storage lanes. There's a million configurations one could imagine. I suspect aside from multi card compute, you won't need more than the bandwidth of 8x pcie4 or 16x pcie3 for years.
|
|
#
?
Nov 14, 2018 00:27
|
|
|
Klyith posted:your video card is perfectly happy without pcie 4.0 This brings me back to the days of AGP 4x and 8x I recall that GPUs still didn't need 8x by the time pcie rolled around
|
|
#
?
Nov 14, 2018 05:20
|
|
|
A quick googling and it seems the difference between 3.0 8x and 3.0 16x on a GTX 1080 was less than 1%. PCI-E 4.0 will benefit superfast storage more than anything. I recall an article back in the day where high end at the time PCI-E 2.0 video cards not being significantly affected until they were dropped down from 4x to 1x. I guess the reason they were all x16 is because of better power delivery and some meaningless spec number boosts.
|
|
#
?
Nov 14, 2018 08:14
|
|
|
The benefit of pcie4 isn't any one specific thing, like storage or gpu, but rather that they can use fewer lanes to achieve the same amount of bandwidth. That means more devices, better links to pcie switches, chipsets and so on.
|
|
#
?
Nov 14, 2018 09:33
|
|
|
Using fewer lanes (for a given amount of bandwidth) really doesn't matter unless you can slice it into smaller increments, which is what's coming in PCIe 5. Whoopee, my 3.0x4 NVMe drive can have 4 lanes with twice the bandwidth it can't use! Same for my 2.0x8 network card, quadruple the bandwidth it can't use! And even if we get newer cards, CPUs don't support dissecting it finer than x8x4x4. So, same as always. Until PCIe 5. edit: or do I have this the other way around? Paul MaudDib fucked around with this message at 09:40 on Nov 14, 2018 |
|
#
?
Nov 14, 2018 09:38
|
|
|
I was watching some dude's video about installing a Ryzen 1600 and he recommended downloading the chipset drivers from AMD's own website to get the Ryzen-specific "balanced" power setting for Win10. He even said that it gave a performance advantage over simply setting the CPU to work in performance mode. Is that the conventional wisdom, and does that still hold true for the 2700/x series?
|
|
#
?
Nov 14, 2018 10:09
|
|
|
Paul MaudDib posted:Using fewer lanes (for a given amount of bandwidth) really doesn't matter unless you can slice it into smaller increments, which is what's coming in PCIe 5. The CPU can negotiate with the device to use fewer lanes, how else would a GPU work in a x8 slot. exquisite tea posted:I was watching some dude's video about installing a Ryzen 1600 and he recommended downloading the chipset drivers from AMD's own website to get the Ryzen-specific "balanced" power setting for Win10. He even said that it gave a performance advantage over simply setting the CPU to work in performance mode. Is that the conventional wisdom, and does that still hold true for the 2700/x series? No, it doesn't even hold true for the 1X00 series. Win10 power plans were updated a while ago and Ryzen Balanced can cause weird issues.
|
|
#
?
Nov 14, 2018 10:22
|
|
|
Arzachel posted:The CPU can negotiate with the device to use fewer lanes, how else would a GPU work in a x8 slot. sure but the CPU can only split its lanes so many ways regardless of what the device negotiates - in a consumer platform that's usually either x16x0x0, x8x8x0, or x8x4x4. The world where we're all doing NVMe and fast networking is going to have to have a more intelligent way to split that down, and it's coming in either the PCIe 4 or 5 standards, can't remember which.
|
|
#
?
Nov 14, 2018 10:29
|
|
|
exquisite tea posted:I was watching some dude's video about installing a Ryzen 1600 and he recommended downloading the chipset drivers from AMD's own website to get the Ryzen-specific "balanced" power setting for Win10. He even said that it gave a performance advantage over simply setting the CPU to work in performance mode. Is that the conventional wisdom, and does that still hold true for the 2700/x series? Ryzen CPUs had lower performance on Windows 10 originally because the default power plans were parking cores (and saving power) too much, so changing the power plan (or installing the AMD-modified balanced one) fixed this. However, Microsoft eventually changed this in a Windows 10 (feature?) update so it shouldn't be necessary just for the power plan/core parking issue anymore.
|
|
#
?
Nov 14, 2018 10:29
|
|
|
Thanks for the advice everyone! I did a dumb thing and bought a Ryzen 2700x / 2080 for maximum  and I'm glad to see that AMD is competitive again. The very first system I ever put together was on a good old Athlon back in 2002 when dual core still was a novel concept, and the trepidation + fear of royally loving something up on a new installation has been with me ever since. 
|
|
#
?
Nov 14, 2018 10:52
|
|
|
exquisite tea posted:Thanks for the advice everyone! I did a dumb thing and bought a Ryzen 2700x / 2080 for maximum Why this is a great thing, not dumb 
|
|
#
?
Nov 14, 2018 11:09
|
|
|
Paul MaudDib posted:sure but the CPU can only split its lanes so many ways regardless of what the device negotiates - in a consumer platform that's usually either x16x0x0, x8x8x0, or x8x4x4. The world where we're all doing NVMe and fast networking is going to have to have a more intelligent way to split that down, and it's coming in either the PCIe 4 or 5 standards, can't remember which.
|
|
#
?
Nov 14, 2018 11:13
|
|
|
Arzachel posted:That's just the most common physical configuration on Intel's consumer platforms. There are Z390 boards that let you do x4x4x4x4 or x8x4x2x2 etc. only with PLX switches also, are there actually any Z390 boards with PLX switches? the boards are pretty bad this time around
|
|
#
?
Nov 14, 2018 11:22
|
|
|
Paul MaudDib posted:only with PLX switches Yeah, you're right. I think Z390 Taichi Ultimate supports x8x4 + 4 x1 links but other than that pcie bifurcation is more limited than I thought.
|
|
#
?
Nov 14, 2018 11:24
|
|
|
Paul MaudDib posted:only with PLX switches I know of exactly one: the Asus WS Z390 Pro. There are rumors about one more board (from Supermicro, I think) but that's all.
|
|
#
?
Nov 14, 2018 11:33
|
|
|
HalloKitty posted:The benefit of pcie4 isn't any one specific thing, like storage or gpu, but rather that they can use fewer lanes to achieve the same amount of bandwidth. That means more devices, better links to pcie switches, chipsets and so on. Yeah, so intel can continue to choke their PCHes and only give them TWO lanes of PCIE 4.0. 
|
|
#
?
Nov 14, 2018 13:56
|
|
|
SwissArmyDruid posted:Yeah, so intel can continue to choke their PCHes and only give them TWO lanes of PCIE 4.0. are we really arguing about hypothetical lane expansions that don't really exist? in practice Ryzen has the same lane expansions as Coffee Lake.. except for the 4 lanes that AMD deigned to actually activate for a NVMe drive (in contrast to Intel's generally superior PCH performance). AMD could have done another 8 PEG lanes right off the CPU, for a total of 24. They didn't want to, Ryzen is the result. Will they activate them on Zen2? Probably not. it's all about the product segmentation, even with AMD chips. Paul MaudDib fucked around with this message at 14:40 on Nov 14, 2018 |
|
#
?
Nov 14, 2018 14:32
|
|
|
Sure, as long as Intel wants to keep fostering situations where their Thunderbolt controllers get hung off the PCH, it's a valid criticism. Here's hoping they use some of those newly freed-up lanes to actually give Thunderbolt controllers their own dedicated lanes straight to the CPU for improved performance.... but I'm not holding my breath. SwissArmyDruid fucked around with this message at 14:47 on Nov 14, 2018 |
|
#
?
Nov 14, 2018 14:42
|
|
|
SwissArmyDruid posted:Sure, as long as Intel wants to keep fostering situations where their Thunderbolt controllers get hung off the PCH, it's a valid criticism. so... did AMD hang their thunderbolt controllers off the PCH? (or, are there valid design reasons to do otherwise... or what?)
|
|
#
?
Nov 14, 2018 14:46
|
|
|
Well, seeing as how the only Thunderbolt-capable AMD systems are desktops with add-in boards, the answer to that would be "depends on what PCIe slot you plug the Thunderbolt card into", now, wouldn't it? And yes, there is a valid design reason to do it, running Thunderbolt off the PCH in notebooks bottlenecks eGPU performance. Unironic LTT video link here: https://www.youtube.com/watch?v=FHwXOwTgYB0 SwissArmyDruid fucked around with this message at 15:51 on Nov 14, 2018 |
|
#
?
Nov 14, 2018 14:48
|
|
|
Those new 64 core EPYCs are clocked in at 2.35 GHz inside a new supercomputer. That seems good, I guess? At Least comparing them all on wikipedia, and seeing how higher core count usually means less GHz. I'm not good at computers, but I'm thirsty for Zen 2 news. https://www.anandtech.com/show/13598/amd-64-core-rome-deployment-hlrs-hawk-at-235-ghz
|
|
#
?
Nov 14, 2018 15:11
|
|
|
|
| # ? Apr 20, 2024 05:21 |
|
|
2.35 is very good for supercomputer, usually those run at 1.8-2ghz still. Probably means they have decent memory bandwidth/good latency, since that's usually the bottleneck in HPC.
|
|
#
?
Nov 14, 2018 15:14
|
|
