|
PC LOAD LETTER posted:True but the x86 "front end" is what will dictate what those "back end" RISC-y execution units get to work on and when so that is what you're essentially forced to judge things by. No, not really. ARM extensions and changes that are "good" for parallelization would be similarly new as the x86 changes.
|
# ? Oct 18, 2018 00:33 |
|
|
# ? Jun 8, 2024 07:31 |
|
Here's the answer to a question that nobody asked: Something might be wrong with Picasso. I have heard loving nothing about its existence from the people I usually talk about this with in the industry, and it kind of makes it more conspicuous by its absence. Which is a shame, because I really want to see how AMD targets the ULV market in time for the next iteration of the Surface Go.
|
# ? Oct 18, 2018 21:02 |
|
SwissArmyDruid posted:Here's the answer to a question that nobody asked: It being on the roadmaps is what's wrong with Picasso. Raven Ridge already got the misc tweaks Pinnacle Ridge had, 12nm isn't amazing and there's no important new GPU tech either.
|
# ? Oct 18, 2018 22:59 |
|
I've lost track. Was that the custom console SoC?
|
# ? Oct 18, 2018 23:34 |
|
No, it's the APU what's supposed to come after Raven Ridge early next year, with either Zen+ or Zen2 silicon, but like Arzachel said, maybe AMD decided it wasn't worth it.
|
# ? Oct 19, 2018 04:32 |
|
I'd have to agree. 7nm production capacity is going to be in massive demand right out the gate (hahaha, gate) and it would be better spend on filling Ryzen, TR and Epyc orders.
|
# ? Oct 19, 2018 05:06 |
|
The thing is, there is evidence for a Picasso and. Raven Ridge2, which is even more confusing. Maybe Picasso was a workstation or EPYC part? Up to 16C32T, and 2816SP for heterogeneous workloads? Also always the possibility a 7nm design was just that much better and available soon enough at low enough cost that a 12nm respond just made no sense, not even to target ULV.
|
# ? Oct 19, 2018 15:14 |
|
I don't really know how to compare intel and amd processors, but what does this 7mm +10-15% performance bump mean? Am I to look forward to something i9-9900k level performance from the zen 2 flagship? Is it just wishful thinking? I really want my refusal to build a new gaming PC despite my desperate wanting to rewarded finally.
|
# ? Oct 19, 2018 15:43 |
|
Broose posted:I don't really know how to compare intel and amd processors, but what does this 7mm +10-15% performance bump mean? Am I to look forward to something i9-9900k level performance from the zen 2 flagship? Is it just wishful thinking? I really want my refusal to build a new gaming PC despite my desperate wanting to rewarded finally. To me, absolute best case is probably a 9900k with better power usage. Some people may say the bar is higher, but i'm being more conservative. The normal caveat is that if you are gaming at 1440p or 4k, none of it will make too much of a difference and the graphics card is the much better investment.
|
# ? Oct 19, 2018 17:16 |
|
It's likely the 3700X and 9900k are within +/-5% of each other (single thread and 1080p), but the 3700X is going to be cheaper and more power efficient (also, likely have shittier thermals due to smaller die and denser transistors). The 3700X is likely to be noticeably faster in multitasking since it doesn't choke when using virtual threads.
|
# ? Oct 19, 2018 19:24 |
|
EmpyreanFlux posted:It's likely the 3700X and 9900k are within +/-5% of each other (single thread and 1080p), yourself
|
# ? Oct 19, 2018 21:16 |
|
Broose posted:but what does this 7mm +10-15% performance bump mean? It would mean that Zen2 on a core to core basis is 10-15% faster at the same clock speed vs Zen+. Generally its accepted that Zen+ is ~5% slower vs Skylake (which architecturally the latest Intel cores are nearly the same as, biggest difference for Coffeelake-refresh over Skylake is probably the various bug fixes for all the exploits that popped up in the last year or so) at the same clock speed. So if AMD pulls that IPC increase off then Zen2 will be 5-10% faster than a 9xxx series Coffeelake-refresh CPU at the same clock speed. Broose posted:Am I to look forward to something i9-9900k level performance from the zen 2 flagship? Right now (and probably at Zen2 launch date) Intel can get their top end chips to 5Ghz out of the box with stock clocks and the latest rumors (probably with early engineering samples which can have various bugs and/or lower clock speeds vs final shipping product) on Zen2 suggest either little or no clock speed improvements over Zen+ so potentially Zen2 would be operating at a base clock of 3.7Ghz still with a auto OC of ~4.3Ghz or so if you've got the cooling for it. IPC matters but so does clock speed and a clock speed deficit of ~700Mhz to 1Ghz+ is significant enough to matter. There were some drat interesting and solid sounding rumors suggesting AMD would get Zen2 to ~5Ghz but that was from a GF foundry guy and GF abandoned their 7nm efforts so that is worthless now. There have been some others who are knowledgeable about TSMC's "7nm" process who seem to think AMD can probably get to ~4.5Ghz base clocks. If we get ~4.5Ghz base clocks + 10-15% IPC increase over Zen+ than yea Zen2 will probably be noticeably faster than the Coffeelake-refresh 9xxx series chips. No word about power usage for Zen2 yet but this all assumes they're willing to keep power consumption about the same.
|
# ? Oct 19, 2018 22:19 |
|
SPOOKCORE MEGGIDO posted:yourself Why the gently caress would I toxx on this?
|
# ? Oct 19, 2018 23:05 |
|
prolly cause it aint happening and youll get banned
|
# ? Oct 20, 2018 00:27 |
|
if you're as confident as you sound in those oddly specific predictions you have nothing to lose, no?
|
# ? Oct 20, 2018 00:28 |
|
You can be confident enough to say you think something will happen without wanting to gamble on it with a toxx, especially considering there's no upside to toxxing if you're right. It's not like he's saying that it's absolutely certain or asking you to gamble on it.
|
# ? Oct 20, 2018 03:46 |
|
yeah but thats nowhere near as fun as peer pressuring someone into toxxing, cmon bro
|
# ? Oct 20, 2018 03:58 |
|
AMD CPU and Platfrom Discussion: Why the gently caress would I toxx on this? E- platfrom GRINDCORE MEGGIDO fucked around with this message at 05:55 on Oct 20, 2018 |
# ? Oct 20, 2018 05:29 |
|
What's so outlandish about my statement? The 2700X is on average 19% slower @ 1080p, and seems to fall behind in synthetics equal to the core clock difference when using CL14 3200. Pushing clocks to 4.5-4.7Ghz and getting ~8-10% increase in IPC should put a 3700X on par with the 9900K. https://www.youtube.com/watch?v=_I--zROoRws Like, I'm not even going "Lol5Ghz", but a fairly minor increase in clockspeed and an equivalent jump as was between Ivy and Haswell.
|
# ? Oct 20, 2018 17:12 |
|
I was originally looking at an Intel Xeon E5-2650 v4 as the CPU for a deep learning rig, but have found a mobo that fits well with the specs I need (128gb ram, lots of PCIe slots for 10+ GPU's) -- the X399 Designare-EX -- but it only supports Threadrippers. What would an equivalent AMD processor to the Xeon e5-2650 be?
|
# ? Oct 21, 2018 00:48 |
|
MarksMan posted:I was originally looking at an Intel Xeon E5-2650 v4 as the CPU for a deep learning rig, but have found a mobo that fits well with the specs I need (128gb ram, lots of PCIe slots for 10+ GPU's) -- the X399 Designare-EX -- but it only supports Threadrippers. What would an equivalent AMD processor to the Xeon e5-2650 be? The 1920X has 12 cores/24 threads like the E5-2650, but it's got almost double the TDP and over 50% faster clock speed, so it should perform significantly better. There's also the Zen+ refresh, the 2920X, which is the same thing but a bit better (better memory compatibility, a bit more power efficient, slightly faster), but it's not released quite yet - it launches October 29th. The 1920X is currently around $400 which is a pretty amazing deal for a 12-core CPU. If you were budgeting for a Xeon that costs (as far as I can tell) around $1200, there's really no reason not to go for a 16-core 1950X or 2950X instead. e: do note, I'm not sure what motherboard support for Threadripper 2 (2920X/2950X) is like on that particular motherboard. It might not work at all, or it might work with a BIOS update which might or might not be available and which might or might not require a Threadripper 1 to install. TheFluff fucked around with this message at 01:07 on Oct 21, 2018 |
# ? Oct 21, 2018 00:58 |
|
TheFluff posted:The 1920X has 12 cores/24 threads like the E5-2650, but it's got almost double the TDP and over 50% faster clock speed, so it should perform significantly better. There's also the Zen+ refresh, the 2920X, which is the same thing but a bit better (better memory compatibility, a bit more power efficient, slightly faster), but it's not released quite yet - it launches October 29th. Wow, well that would work perfect if so, because I just traded one of my 1080's for a new in box 1920x to a goon in SA-Mart a week or two ago. I've never built an AMD rig before, but I had thought of doing so in the future and had a 1080 laying around, so I agreed to the trade. That would be great if it worked as well as a E5-2650 v4
|
# ? Oct 21, 2018 01:05 |
|
TheFluff posted:e: do note, I'm not sure what motherboard support for Threadripper 2 (2920X/2950X) is like on that particular motherboard. It might not work at all, or it might work with a BIOS update which might or might not be available and which might or might not require a Threadripper 1 to install.
|
# ? Oct 21, 2018 01:13 |
|
EmpyreanFlux posted:What's so outlandish about my statement? The 2700X is on average 19% slower @ 1080p, and seems to fall behind in synthetics equal to the core clock difference when using CL14 3200. Pushing clocks to 4.5-4.7Ghz and getting ~8-10% increase in IPC should put a 3700X on par with the 9900K. https://www.techspot.com/article/1616-4ghz-ryzen-2nd-gen-vs-core-8th-gen/page3.html (a slightly outdated test now, but I'm sure someone will do one with 2700X vs 9900K and a 2080 Ti) At equal clock speeds the 8700K is ~10% ahead of the 2600X in games (and sometimes more) so it'll take than just a 10% increase in IPC to make back whatever frequency deficit they end up with. Llamadeus fucked around with this message at 01:37 on Oct 21, 2018 |
# ? Oct 21, 2018 01:19 |
|
MarksMan posted:Wow, well that would work perfect if so, because I just traded one of my 1080's for a new in box 1920x to a goon in SA-Mart a week or two ago. I've never built an AMD rig before, but I had thought of doing so in the future and had a 1080 laying around, so I agreed to the trade. That would be great if it worked as well as a E5-2650 v4 It's a much better chip in nearly every way except power consumption. Also, GPU passthrough for VMs tends to be trickier than intel, but it's tricky no matter what, frankly.
|
# ? Oct 21, 2018 18:35 |
|
I'm not entirely sure why everyone is that skeptical about the supposed IPC gains on Zen 2. Sure, it sounds a little good to be true... But given that this is a new architecture, there's probably quite a bunch of low hanging fruit that didn't make it in time, and maybe having the architecture out in the open and in use also uncovered a bunch of things that looked better in theory than they do practically, and need fixing. Plus, implementing PCIe 4.0 will speed up the Infinity Fabric, which should speed up inter-CCX comms.
|
# ? Oct 21, 2018 18:44 |
|
Why will implementing pcie 4 speed up IF?
|
# ? Oct 21, 2018 18:51 |
|
Everyone keeps telling me that IF is based on PCIe (sharing PHY or whatever). PCIe 4 has double bandwidth of PCIe 3.
|
# ? Oct 21, 2018 18:56 |
|
Infinity Fabric is described everywhere in technical articles as "a superset of HyperTransport" so I have never been clear exactly what the relationship is between that and PCIe. I think it is running the IF/HT message protocol over PCIe physical links? In that case, the gen 4 controllers for the exterior of the chip are of no benefit. IF is already running a protocol that's much faster than pcie gen 4 on the inside of the chip. You can support a far more demanding transaction & data spec when you only have to travel a few 10s of mm between chips.
|
# ? Oct 21, 2018 19:46 |
|
If it's only on the chip exterior, it'll still do something for the Threadrippers. If packet size stays the same, double bandwidth will clear the link faster and reduce latencies a bit.
|
# ? Oct 21, 2018 19:49 |
|
Apparently Intel is axing 10nm, according to Charlie over at Semiaccurate https://semiaccurate.com/2018/10/22/intel-kills-off-the-10nm-process/. So, it's Charlie and requires the appropriate salt, but that means Intel is moving to a true 7nm node but that's also in like, late 2020 optimistically as well. That's Zen2 and Zen3 competing against 14nm Infiplus products for 18 months, and if TSMC makes early production for 5nm in 2021 then Intel's only ever going to get the crown back for like 3-6 months before Zen3+ drops.
|
# ? Oct 22, 2018 14:39 |
|
Yeah if true that is a gigantic windfall for AMD. Its so contradictory vs what Intel has been saying lately publicly + some of the comments from SA and others that Intel was going to gimp 10nm to at least get something out the door to pay for the fabs and increase supply that I got a hard time believing it right now. Still he might be right about it being good for Intel in the long run if it he is also right about even the "fixed" gimped 10nm being financially non-viable. quote:Intel is moving to a true 7nm node but that's also in like, late 2020 optimistically as well
|
# ? Oct 22, 2018 15:02 |
|
I'm pessimistic about it, if 7nm is 2021-2022 realistically then it's basically conceding the performance and server market for 24-30 months. Look I know AMD survived that, but that's a hell of a tumble for Intel and it's not even comparable to the old P4/Athlon days. Potential mass layoffs, maybe shuttering their GPU venture [again, lol], and worst case not even managing to get to 7nm before the fabs need to be spun off. I'd think they'd be trying to backport as much as they can into the 14nm process, expand 14nm fabs again, and just try and get 7nm running instead of struggling with 10nm at all. "12nm", but starting from 14nm not 10nm, is what I'm seeing. Or it's bullshit because it's Charlie and Intel proceeds to 10nm in late 2019.
|
# ? Oct 22, 2018 15:18 |
|
Intel has a huge marketshare and mindshare they will weather the storm but hopefully in the end of it all we'll have 2 strong X86 competitors going head to head in price, performance and marketshare. There really is nothing to mourn as a consumer. Intel meltdown is good for all of us really.
|
# ? Oct 22, 2018 15:25 |
|
You're right about it being a gigantic hit to Intel and all the effects it'd have, and it is a rumor after all, but its not too much of a stretch to believe the same leadership that messed up Intel's 10nm process might've repeated some mistakes on their 7nm process. Looks to me like they're already trying to expand their 14nm process quite a bit so you might be right too about what they will do.
|
# ? Oct 22, 2018 15:27 |
|
Intel have been doing AI/ML-specific chips for a while, I don't know how separate the so-called GPU division is; Intel will have to do something to keep Nvidia from being in every car a few years from now or they'll be super boned in compute and CV, supposed threats from AMD aside. At some point, MS will probably also get impatient and start putting ARM chips in their Surface devices, Intel need to do something when the paradigm changes again like it did with mobile.
|
# ? Oct 22, 2018 15:32 |
|
So hang on to my 2500k for another 3 years if the RAM and mobo hold up, got it. Gyah, the Intel 9 series release has me itchin for Zen 2 news.
|
# ? Oct 22, 2018 15:39 |
|
TheCoach posted:
Not sure if that pun was intended but I disagree from a x86 consumer perspective. If AMD takes the clear lead they will exploit that position exactly like Intel did. Apple isn't going to wait for Intel 7nm "to maybe turn out ok in a few years", same with Microsoft. I think this news will be a massive catalyst for ARM in the laptop (and possibly soon desktop) market, with the architecture switch giving software vendors an excellent opportunity to further close down their walled gardens.
|
# ? Oct 22, 2018 16:23 |
so exactly how many of their fabs -don't- have 14nm capability?
|
|
# ? Oct 22, 2018 17:48 |
|
|
# ? Jun 8, 2024 07:31 |
|
.
sincx fucked around with this message at 05:50 on Mar 23, 2021 |
# ? Oct 22, 2018 18:05 |