|
Biostar and dvi
|
# ¿ Apr 20, 2017 05:04 |
|
|
# ¿ May 8, 2024 00:49 |
|
eames posted:We'll find out soon enough I guess. Uh of course it is Zen is built using 8c zeppelin modules connected by infinity fabric The bigass 32c thing is effectively a 4 chip module And yes NUMA will be an issue but just wait until poo poo gets patched and it'll be fine
|
# ¿ May 30, 2017 19:52 |
|
wargames posted:If the interposer is pretty expandable it shouldn't be an issue, and amount to huge savings. Bingo this allows for lots of cost savings on construction since big dies are crazy expensive Intel is likely gonna do the same thing with EMIB
|
# ¿ May 30, 2017 19:53 |
|
Tr4 mini itx when
|
# ¿ Jun 2, 2017 20:05 |
|
Kazinsal posted:They also need to get deals in with vendors for the server kit. Having super inexpensive 32C/64T chips that clock just as high as their Intel counterparts is worth next to nothing if you don't have HP/Dell/Cisco/IBM/etc. prepared to sell systems with them. Nah just create an ocp box and then the clouds will eat that poo poo up cloud providers are the new drivers here
|
# ¿ Jun 4, 2017 12:44 |
|
Rastor posted:Cloud providers buy in the biggest quantities but at the same time for that reason they want to be sure of the performance and stability before signing the purchase order. Some design wins with HP/Dell/Cisco/IBM/etc. will no doubt be high priority for AMD right now. Lol cloud providers are like 95% driven by cost Failure is built into the design dude. They just migrate loads around. Azure used to use poo poo opterons since they were cheaper than xeons Low margins ultra high volumes > than ultra high margin ultra low volumes for amd here. Sure the hospital datacenter will buy E7 but aws will buy 100x the epyc 32 core proc Malcolm XML fucked around with this message at 20:42 on Jun 4, 2017 |
# ¿ Jun 4, 2017 20:40 |
|
Palladium posted:On the other hand, these software corps aren't exactly the thrifty sort either. Uhh if you can save 1% of power in a million core dc than that saves millions The perks are peanuts in operating costs
|
# ¿ Jun 4, 2017 20:44 |
|
Rastor posted:Yeah, they're driven by cost, and if you've got nodes failing left and right your costs shoot up. If the nodes are cheap enough then it doesn't matter Its a complex roi calc Long story short I guarantee amd spoke with all of the major clouds when designing EPYC THREADRIPPER and built it to easily slot into their requirements and that there are test machines running ocp epyc platforms
|
# ¿ Jun 4, 2017 20:51 |
|
Rastor posted:Oh no doubt there are test machines in labs, I agree there. But stuff like being unstable under heavy compiling loads is gonna have to be ironed out before the production order is placed. Oh you sweet summer child
|
# ¿ Jun 4, 2017 21:16 |
|
Rastor posted:Without violating NDAs, I know what I'm talking about : so do I It is as if different providers and customers have different requirements even inside a cloud....
|
# ¿ Jun 5, 2017 03:08 |
|
SwissArmyDruid posted:If there is a criticism that I have continually levelled at AMD, it is that they appear to be hell-bent on leaving as much money on the table as possible. AMD needs market share just as much as revenue.
|
# ¿ Jun 7, 2017 00:42 |
|
Scarecow posted:Yeah they really need to build market and mind share at this stage given their current HEDT market % is what.....0? also it could very well be true that making the ripper chips is really cheap and they are getting really high yields so even at half the price they are still turning a nice profit while also making people stop and go "gently caress 16 cores for that much I cant say no" Totally, 32 threads for $850 is enough for me to turn my 5820k into the worlds most overpowered NAS
|
# ¿ Jun 7, 2017 00:51 |
|
Watermelon Daiquiri posted:Gah good loving lord that 'quantum well' gate growing a layer of si in the middle of all that? granted, you really only need one mask for all of that metal, but still. Yeah a combo of iii-v and gaa seems a bit too much at the same time I suspect Intel is going to do one or the other and then the next step will be to.do both However Samsung TSMC and the relentless pace of mobile have basically brought a second and third fab competitor to the game so who knows
|
# ¿ Jun 10, 2017 15:17 |
|
Actually yeah that qw fet is a iii-v finfet
|
# ¿ Jun 10, 2017 15:19 |
|
Much like most tech wait for the 1.1 revision Can't wait for the threadripper v2 how many cores can we got on a dinner plate sized mcm
|
# ¿ Jun 15, 2017 12:41 |
|
Combat Pretzel posted:A while ago I've seen a video of some Linux dude testing it and forcing errors by loving with RAM voltages. It actually reported single bit errors under Linux. And caused a kernel panic on worse errors. I figure it'll do the same in Windows, too. r-rip my threads AMDaddy
|
# ¿ Jun 29, 2017 21:42 |
|
Generic Monk posted:considering the RAM is integrated into the package the latency is probably pretty low This The capacitance and inductance from the multiple DIMMs and long traces limits max frequency That said package integrated dram is best and maybe Intel can get off its rear end or and can make a MCM with 16 gb of hbm Malcolm XML fucked around with this message at 20:02 on Jul 4, 2017 |
# ¿ Jul 4, 2017 19:52 |
|
Making nand is printing money at the moment since the demand is insane and supply is constrained to 3-4 makers Toshiba's entire net worth is their nand division
|
# ¿ Jul 15, 2017 20:31 |
|
Subjunctive posted:From mutual friends, I've heard that Tesla is paying him much, much more than AMD or Apple did. I strongly suspect Jim Keller has a large long position in AMD That said musk and Tesla are in a position to pay him a lot more
|
# ¿ Jul 25, 2017 02:48 |
|
Hmm Vega looks ok and my new monitor has free sync by accident. Might shift my 980ti over to act as a cuda slave on a Linux box for NN training. AMD breaking 15 on afterhours
|
# ¿ Jul 26, 2017 10:55 |
|
Obv gonna wait for the reviews and ideally the fire sale after it tanks
|
# ¿ Jul 26, 2017 11:39 |
|
The layout people for asrock deserve a goddamn medal
|
# ¿ Jul 28, 2017 15:59 |
|
welp vega is apparently a mining beast so good for my amd stock if they can plow that revenue into a decent Navi then I will be happy
|
# ¿ Aug 3, 2017 22:52 |
|
Twerk from Home posted:If you've got an hour to spare, the current state of the art is work-stealing algorithms and doing everything possible to minimize synchronization. Game developers discover ancient multicore secrets
|
# ¿ Aug 9, 2017 19:46 |
|
ShinAli posted:Not to mention we only have what we're given to work with, meaning either exposed APIs are too low level to immediately plug and play vs. a too high level API that doesn't serve the needs that we're looking for. i always thought a good dataflow architecture would make a lot of sense but i wouldn't know much tbh
|
# ¿ Aug 10, 2017 00:20 |
|
Mofabio posted:Do you folks think it's fair to say AMD is the primary innovator in chip design, and Intel's the innovator in fab process design? Looking at 21st century CPU improvements, it seems like (correct me if I'm wrong) AMD was first to market for most of the major innovations: x86-64, multi-core, on-die GPU (Intel HD graphics came with Westmere, but AMD bought ATI in '06 and talked at the time of doing integrated graphics), integrating the northbridge, and now MCM to squeeze more performance while chips butt up against Moore's Law. AMD is forced to go the cheapass route since they have no money which occasionally makes them invest in new tech to save bux And then they piss it away on canards like HBM
|
# ¿ Aug 24, 2017 00:33 |
|
EdEddnEddy posted:So I am still catching up on the thread (page 38 still) but wanted to check. Anyone see that Tesla appears to be working with AMD to make an AI chip for their cars? hell yeah bois up uP UP!!!! FaustianQ posted:I wonder if they could get something like 50% of the benefits of moving to "10nm" and this is why Nvidia and AMD basically laughed off 10nm for their products. "Hmm, yes redesign everything, new masks included, for a nearly dead node or, or, now bear with me, we just mature the current node further and spend a gigantic assload less for making it halfway." nodes are marketing names at this point this is very likely "10nm" FEOL with "14nm" BEOL or somesuch like tsmc 16nm used older BEOL tech Heck intel is basically hosed with 10nm, 14+++ is gonna be better than 10 for everything but extreme density/power needs (apparently SemiAccurate has the scoop on whats wrong but i am not spending 1k on that)
|
# ¿ Sep 20, 2017 23:27 |
|
Watermelon Daiquiri posted:Well, yeah, both intel-10 and gf-7 are banking on EUV working out, which so far it isnt doing so hot p sure intel 10 is still multi patterning only
|
# ¿ Sep 21, 2017 20:19 |
|
also we're getting into nodes that need multi-patterned EUV which is gonna be hella fun
|
# ¿ Sep 21, 2017 20:20 |
|
AMD needs volume volume volume to get market share so them aggressively binning is no surprise to get the cheaper stuff out there They really need that loving Raven ridge though to get the real volume And they need a modular Navi core too to erase the stink of Vega
|
# ¿ Oct 5, 2017 19:00 |
|
Glofo 7 has been rumored to be quite good so I'm not surprised that amd would want to jump on it
|
# ¿ Oct 26, 2017 11:02 |
|
hmm perfect for my bunker oil burning generator powered wanking station
|
# ¿ Oct 28, 2017 18:34 |
|
crazypenguin posted:GPUs can support the newer cable standards without needing any internal IO upgrade (this is why the ports are directly on the card: so all that data doesn't have to go over any internal bus.) pcie4 is totally pointless for gaming, since gaming is not pcie bandwidth bound. maybe there's some latency improvement but it's not the orders of magnitude that'd be necessary and pcie5 is like 2 years out after that so i suspect amd/nvidia will just wait for that for mass market stuff and for GPGPU/deep learning where it IS pcie bandwidth bound jump on pcie4 and then pcie5
|
# ¿ Oct 30, 2017 01:07 |
|
A loving hdd in 2017 way to shoot yourself in the brain and
|
# ¿ Nov 5, 2017 03:03 |
|
ohgodwhat posted:Oh cool based on your recollection we can all start violating our non competes, no matter the jurisdiction or structure of the agreement. In California they are not.
|
# ¿ Nov 10, 2017 22:20 |
|
Palladium posted:Best outcome would be an AMD-NV merger, selling RTG assets to Intel with interim licensing of said IP until NV-based APUs come to play. Would likely not pass antitrust for a number of reasons. GPU consolidation (Intel has 0% of the dGPU market right now) and the x86 license falls apart if AMD is acquired iirc.
|
# ¿ Nov 12, 2017 15:31 |
|
Watermelon Daiquiri posted:To be fair they explicitly said that intel would get dgpu ip doesn't really matter even if you're using the standard DoJ "do consumers pay more" criteria -- nvidia is twisting the screws and as a shareholder, i am glad of it
|
# ¿ Nov 12, 2017 16:05 |
|
Kazinsal posted:Phoronix is the perfect benchmarking/etc site for your stereotypical Linux Uber Alles type PC master race nerd. The benchmarks mean poo poo when comparing to any other platform for myriad reasons so naturally they look really good on Linux systems. Dude knows exactly what he's doing and somehow manages to sustain the site and I guess himself off it. phoronix owns if u want someone else to read the mailing lists for you and paste them into posts with ads on them
|
# ¿ Nov 25, 2017 00:32 |
|
Yeah instead of using ioctls like windows they mount efivars as rw without access control so any dumbass script can brick the system? Lmao Doesn't ryzen put the EFI / bios on the chip itself like a SoC? Bricking a CPU would be ace
|
# ¿ Nov 25, 2017 10:35 |
|
|
# ¿ May 8, 2024 00:49 |
|
So it turns out that intel 10nm is turbofucked, meaning that AMD + glofo 7nm could actually be ahead/competitive with intel on a process level. AMD still needs a good GPU but getting that sweet sweet datacenter cpu cash would go a long way to funding RTG well....
|
# ¿ Dec 14, 2017 14:32 |