|
Cygni posted:Prolly aint gonna be cheap and you will really need specific workloads to flex them, but 64c/2TB DDR4-3200 with an eight channel memory in ATX form factor is p dang cool. 16 slots of memory ain't fitting into no ATX form factor
|
# ? Jul 11, 2020 04:43 |
|
|
# ? Jun 5, 2024 17:20 |
|
It's basically an overclockable (presumably?) Epyc. This is the fully enabled part, runs UDIMM/RDIMM/LRDIMM, unlocked memory controller clocks, 128 PCIe lanes, 8 memory channels, the full shebang. Single socket only but who cares. I'm sure the price is going to be eyewatering but nice.
|
# ? Jul 11, 2020 04:44 |
|
Klyith posted:16 slots of memory ain't fitting into no ATX form factor ah poo poo good call.... but what if like... you only have like 1 PCIe 16x slot or like somethin??
|
# ? Jul 11, 2020 04:45 |
|
Cygni posted:ah poo poo good call.... but what if like... you only have like 1 PCIe 16x slot or like somethin?? are you kidding, offer them a case of the good stuff and Asrock Rack's engineers will do at least 8 sticks on ATX, poo poo they'll do 4 channels and one slot on a mITX or some poo poo. It'll probably sell to some ridiculous edge/soho segment, RDIMMs on your microserver. 64 cores of VMs with tons of memory for zfs and poo poo. Medium office in a box. They have Skylake-SP ITX boards fo sho. The full LGA3647 part. Why? Why not, it's cool as gently caress. It's actually not even a particularly high-density component design for them. Asrock Rack are some crazy motherfuckers. Paul MaudDib fucked around with this message at 06:42 on Jul 11, 2020 |
# ? Jul 11, 2020 04:48 |
|
i believe in my heart of hearts that a 16 ram slot regular atx board is possible and in fact i think they are cowards if they don't release it personally JUST cut out them lovely PCIe slots for dorks and slide the fuckin thing over! how hard could it be! put some FOOD in the LIFERAFT, do i have to think of everything??
|
# ? Jul 11, 2020 05:00 |
|
Cygni posted:i believe in my heart of hearts that a 16 ram slot regular atx board is possible and in fact i think they are cowards if they don't release it personally one thing Asrock Rack has always done really well is making use of space on the back. They could totally do that, 16 slots, maybe put half the channels as laptop-style dual riser horizontal sockets on the back if needed (pretty sure they did that on some of their previous ITX stuff). Then whatever PCIe slots you can fit on the top, then fill every last mm2 of the top and the entire loving back with (dual risered) M.2 slots and 10GbE controllers or SAS controllers and 10gbe and poo poo. Use the thickest, most luxurious corinthian leather high-layer PCB design and just fill up as many of those lanes as possible with compact, efficient onboard controllers. Paul MaudDib fucked around with this message at 05:15 on Jul 11, 2020 |
# ? Jul 11, 2020 05:09 |
|
SwissArmyDruid posted:Thank god, AMD is finally realizing that the people that buy these kinds of chips need more than half a gig of RAM. That was just stupid and asinine. I need at least 2GB of ram.
|
# ? Jul 11, 2020 05:34 |
|
I'll be very curious how "Pro" they are. It sucks that they're perfect for my use case (CAD, engineering, simulation, VMs, lazy fuckery and never closing anything) but I'm really leery over their socket garbage for Threadripper.
|
# ? Jul 11, 2020 05:36 |
|
NewFatMike posted:I'll be very curious how "Pro" they are. It sucks that they're perfect for my use case (CAD, engineering, simulation, VMs, lazy fuckery and never closing anything) but I'm really leery over their socket garbage for Threadripper. I wouldn't bet on a socket upgrade, be nice if it happened but they haven't said anything official. Frankly part of the game here may be that we are most of the way to Zen3 and they are betting that the people who haven't bought in are regretting some niche aspect they're offering the full thing. So they are launching this now, Milan in say 6-12 weeks, and we get Milan Pro 4995X a year later, rinse and repeat. Some customers staying on Zen2 wouldn't be awful for AMD or the customers. I really think Zen3 is going to be more incremental - Zen2 had general IPC improvements, memory controller/infinity fabric improvements, the removal of NUMA penalties (nearly completely), the move to a much more efficient 7nm,PCIe 4, etc. Zen2 is a much much more reasonable general-purpose platform than Naples and will continue to rock if it's priced at the entry level. The big guys will pay extra for the latest thing that is 15% faster and 10% more efficient and custom-tailored to their exact use-case, others can stay on a fully unlocked Zen2 platform and have the weird particular combinations of capabilities they need and just dunk on Intel server in high-capability segments. The "$80 1600 vs $160 2600" strategy for server. If that's their strategy that's actually p. good, there are people who will buy that if the price is not completely insane. It still progressively turns the screws on Intel in other high-performance/high-capability segments but lets them also price higher for big clients who want the latest and/or the absolute most capable and won't be overclocking (multi socket basically). Paul MaudDib fucked around with this message at 06:52 on Jul 11, 2020 |
# ? Jul 11, 2020 05:44 |
|
ratbert90 posted:I need at least 2GB of ram. Whoops.
|
# ? Jul 11, 2020 05:50 |
|
That does track - I do wonder if the gains are going to be a little more significant. Maybe 15% over these latest releases, not necessarily launch ones. The drop rate of EPYC SKUs from a more mature process have been really interesting, too. Every few months, a new one drops and even the entry level gains have been pretty significant. Milan is gonna be p cool at any rate, and maybe if it's not a great deal I'll find a good entree with Zen 2.
|
# ? Jul 11, 2020 06:00 |
|
NewFatMike posted:That does track - I do wonder if the gains are going to be a little more significant. Maybe 15% over these latest releases, not necessarily launch ones. That's pretty much quibbling over like 1-2% so whatever. Yeah, if they hit a solid 15% IPC gain average and they clock 100-200 MHz higher 17-18% overall is possible, at iso power, no efficiency gains. Another few percent plur or minus is basically within the margin of error about the design unknowns, AMD CPU dept has pretty well clamped down on at least public leaks, so I don't think all that much is really publicly known apart from the CCX reconfiguration and general expectations of node gains. I would assume some misc gains here and there in clock and stuff but... Presumably this is some ongoing shift in their product segmentation strategy, otherwise they wouldn't launch a whole new chipset for it. I don't see a compelling differentiation in the Zen3 featureset as far as home users are concerned. If the price is right then Zen2 24C would totally make a great workstation or gaming rig or whatever. If AMD starts to drop Zen2 TRX40 parts to TR 1000 prices then uh, shut up and take my money. And yeah AMD has done cool poo poo and continues to do cool poo poo. I've suggested 3700X and 3900X (over our 9700 build) for developer desktops but it didn't really get any traction. Paul MaudDib fucked around with this message at 06:28 on Jul 12, 2020 |
# ? Jul 11, 2020 07:06 |
|
Super cheap Matisse is more than about pricing tiers, I think it's about taking up valuable foundry space. If their talks with TSMC have gone amazingly well, they have found themselves in a poition where TSMC will keep rpriotizing them as long as they can pay. If they can keep the 7nm wafers flowing, it's not merely about making server market demand but actively strangling competitors if AMD can just keep munching on more and more of TSMC's wafer supply and no one else can get high enough volume, while also forcing Intel to keep dropping prices on their SKU's to keep up. Extremely Nth Dimensional Chess Fantasy, but like we've heard rumors about Intel moving to TSMC for dGPU, and we know Nvidia needs TSMC to keep manufacturing GPUs. Now imagine AMD can leverage the success of Zen2 and Zen3 to cripple Nvidia's supply of GPUs while also slowing down their ability to adopt new nodes because AMD slots in after Apple on new node priority. Your GPU rival can't beat your gpu stack if they can't produce or sell their own designs in volume, can they?
|
# ? Jul 11, 2020 09:21 |
|
Well, there was a rumour that Ampere would be produced on Samsung's 8nm process instead of TSMC's 7nm.
|
# ? Jul 11, 2020 14:53 |
|
Paul MaudDib posted:Words! Good ones, too! Thank you, my dude! Workstation chat is always a good time. Release will be fun at any rate. EmpyreanFlux posted:Extremely Nth Dimensional Chess Fantasy, I'm not sure they even if the new TSMC facility here in the States opened tomorrow and AMD had all the capacity in it that it would be enough to have these knock on effects or if it would even make sense for TSMC to do that. I think everyone's waiting for the other shoe to drop with AMD. It might not happen, but if Nvidia and Apple are also big customers of yours, you know they can move product and it makes more sense to keep yourself diversified. E: I finally had some coffee and realized I took your internet comment way too seriously NewFatMike fucked around with this message at 17:52 on Jul 11, 2020 |
# ? Jul 11, 2020 16:40 |
|
If we can call 500MB half a gig I am going to start calling 500GB half a teb teb
|
# ? Jul 11, 2020 18:50 |
|
FuturePastNow posted:If we can call 500MB half a gig I am going to start calling 500GB half a teb Careful, it won't be long before Linus says this in one of his videos and it's suddenly everywhere.
|
# ? Jul 11, 2020 18:59 |
|
FuturePastNow posted:If we can call 500MB half a gig I am going to start calling 500GB half a teb nice
|
# ? Jul 11, 2020 19:02 |
|
Shouldna it be ter? They don't get called gibs.
|
# ? Jul 11, 2020 19:16 |
|
Some Goon posted:Shouldna it be ter? They don't get called gibs. otherwise I propose to just call em kgigs (pronounced kay-gig)
|
# ? Jul 11, 2020 19:29 |
|
Half a rabbit imo
|
# ? Jul 11, 2020 19:29 |
|
Paul MaudDib posted:It's basically an overclockable (presumably?) Epyc. This is the fully enabled part, runs UDIMM/RDIMM/LRDIMM, unlocked memory controller clocks, 128 PCIe lanes, 8 memory channels, the full shebang. Single socket only but who cares. I feel dirty saying it but since I basically amortized the cost of my 2600K over ten years...well I'm ready to jump ship after getting a taste of 3960X at work. I just hope everything will fit into my Meshify C. Fuckload of money but if paradigms continue... should just be a GPU upgrade every 3 years still.
|
# ? Jul 11, 2020 19:45 |
|
Some Goon posted:Shouldna it be ter? They don't get called gibs. ok, terb then, terb is a funny word half a terb 20 terbs Urban Dictionary tells me a terb is a man with a huge penis, which seems appropriate.
|
# ? Jul 11, 2020 20:58 |
|
One time I heard a guy say GIBIBYTE and I nearly stabbed him after I went rage blackout.
|
# ? Jul 11, 2020 22:33 |
|
gibibyte is an actual thing though, it’s the 1024-based version of a gigabyte, abbreviation is GiB. You technically have 16 gibibytes of memory or whatever.
|
# ? Jul 11, 2020 22:41 |
|
Is the Noctua NHD15 overkill for AMD CPUs? I picked one up and I’m going to have some problems fitting the ram or the case. I can still return it for something else like a U14S. My goal is to keep things running as cool as possible while generating a minimum of noise. For reference I’m probably gonna run a 3700 until the 4000 series comes out.
|
# ? Jul 11, 2020 22:52 |
|
redeyes posted:One time I heard a guy say GIBIBYTE and I nearly stabbed him after I went rage blackout. I don’t post a lot here but this reminded me when I used to work as a manager for AT&T I was helping this old dude with getting his google voice assist working and he could never trigger it because he couldn’t actually pronounce “ok google” and kept saying “OK googer” and it was all I could do not to laugh
|
# ? Jul 11, 2020 23:04 |
|
Kraftwerk posted:Is the Noctua NHD15 overkill for AMD CPUs? I picked one up and I’m going to have some problems fitting the ram or the case. I can still return it for something else like a U14S. My goal is to keep things running as cool as possible while generating a minimum of noise. For running a 65W TDP part (more like ~85W under stress) at AMD reference settings, it's a bit much yes. Be careful that your mainboard doesn't do anything stupid with auto-OC or voltages, but that's just general AMD Zen advice.
|
# ? Jul 11, 2020 23:13 |
|
Paul MaudDib posted:gibibyte is an actual thing though, it’s the 1024-based version of a gigabyte, abbreviation is GiB. Gigabytes can be defined as 1024-based too per JEDEC; it's just stupid because it can be both 10-based and 1024-based; they really should have revised kilo, mega, and giga prefixes to their kibi, mebi, and gibi equivalents
|
# ? Jul 12, 2020 00:19 |
|
Kraftwerk posted:Is the Noctua NHD15 overkill for AMD CPUs? I picked one up and I’m going to have some problems fitting the ram or the case. I can still return it for something else like a U14S. My goal is to keep things running as cool as possible while generating a minimum of noise. It's overkill for anything below a 3900X. With a 3600 & 3700 the U14S has enough surface area that it can achieve the same results. On the CPUs with more than 8 cores the D15 has a significant a lead. OTOH there's no such thing as overkill for heatsinks -- reducing fan speed is always an option. But the D15 really does need some consideration for its size. You pick the D15 first, then select other parts based on what it needs.
|
# ? Jul 12, 2020 00:28 |
|
|
# ? Jul 12, 2020 00:42 |
|
EmpyreanFlux posted:Extremely Nth Dimensional Chess Fantasy, but like we've heard rumors about Intel moving to TSMC for dGPU, and we know Nvidia needs TSMC to keep manufacturing GPUs. Now imagine AMD can leverage the success of Zen2 and Zen3 to cripple Nvidia's supply of GPUs while also slowing down their ability to adopt new nodes because AMD slots in after Apple on new node priority. Your GPU rival can't beat your gpu stack if they can't produce or sell their own designs in volume, can they? Nvidia is using Samsung for 7nm, for their low-end (aka high-volume) parts. They're only using TSMC for the stuff that sells for big bucks, which means they can afford as much "priority" as they want. The samsung move was to try to get TSMC to reduce prices... But they're hardballing TSMC just because they can -- they're in a position where they still have enough lead over AMD in GPUs that they can accept a worse process and still win.
|
# ? Jul 12, 2020 01:18 |
|
Klyith posted:Nvidia is using Samsung for 7nm, for their low-end (aka high-volume) parts. They're only using TSMC for the stuff that sells for big bucks, which means they can afford as much "priority" as they want. Honestly now that Apple is buying up as much of TSMC's 5nm production as they can, and probably will increase it quite a bit more for ARM Macs, I wouldn't be surprised if TSMC just told Nvidia to pound sand at some point.
|
# ? Jul 12, 2020 02:20 |
|
A couple more Milan leaks today. I think we're in the run-up to the official launch. Guessing 6-8 weeks to availability for companies that aren't a part of the FAANG club. https://www.igorslab.de/amd-epyc-milan-leak-drei-fruehe-genesis-samples-im-a0-stepping-samt-kernkonfiguration-und-taktraten-aufgelistet/
|
# ? Jul 12, 2020 06:36 |
|
Klyith posted:Nvidia is using Samsung for 7nm, for their low-end (aka high-volume) parts. They're only using TSMC for the stuff that sells for big bucks, which means they can afford as much "priority" as they want. I was actually kind of making a joke post, I don't think TSMC is going to drop Nvidia, Nvidia is a guaranteed seller so TSMC gets theirs regardless of the temperature of the relationship. But Apple and AMD are also known safe competitors, and I do think Nvidia has permanently lost bargaining room with TSMC especially over the latest stunt. TSMC is firmly in control now. We're also in Hellworld, the multi-universes most loving awful and derived sitcom and Hauwei has already lost it's access to 7nm wafers and died so anything is possible.
|
# ? Jul 12, 2020 08:58 |
|
Is there a good overclocking guide for Ryzen? I picked up a 3700x and a X570 Phantom Gaming 4S and have a Noctua D15, so I should be good for thermals.
|
# ? Jul 12, 2020 20:28 |
|
Wrar posted:Is there a good overclocking guide for Ryzen? I picked up a 3700x and a X570 Phantom Gaming 4S and have a Noctua D15, so I should be good for thermals. Turn on PBO and forget about it. Ryzen already operates near its limits.
|
# ? Jul 12, 2020 20:43 |
|
Wrar posted:Is there a good overclocking guide for Ryzen? I picked up a 3700x and a X570 Phantom Gaming 4S and have a Noctua D15, so I should be good for thermals. What’s your case and RAM?
|
# ? Jul 12, 2020 21:18 |
|
Cooler Master TD500 Mesh and Corsair Vengeance LPX 32GB (2x16GB) 3200MHz C16
|
# ? Jul 12, 2020 21:35 |
|
|
# ? Jun 5, 2024 17:20 |
|
Some Goon posted:Turn on PBO and forget about it. Ryzen already operates near its limits. Does PBO even do anything?
|
# ? Jul 12, 2020 22:28 |