|
Paul MaudDib posted:Does Xeon-W really have half a ring? I thought it (same silicon as SKL-X) both had mesh. What's the implication of that in terms of AVX clockdown and/or whatever test they were doing? I just meant that Xeon-W clocks down much less under load, so since the second test was done on Xeon-W silicon it would show less impact. Malcolm XML posted:Reminder most software engineers have no idea that cpus are vastly more complicated than what they are taught. Most think that the in order scalar cpus of the 80s are basically it. One of the most famous answers on stack overflow explains saturation branch prediction. In 2008. In 2018 there's a tiny perceptron on amd cpus doing that. One guy is a CS professor!! Malcolm XML posted:Now incorporate multiple caching hierarchies, NUMA on chip with the ring bus/mesh, NUMA off chip, cache coherence protocols, out of order execution with functional units, DVFS and other modern cpu features and Joe Javascript's head explodes. Malcolm XML posted:the principal of least surprise says that AVX-512 shouldn't cause massive frequency changes since it's billed as just another ISA extension though
|
#
?
Apr 25, 2018 02:51
|
|
|
|
| # ? Apr 28, 2024 04:40 |
|
|
Dumb question: what is it about avx instructions that causes so much additional heat?
|
|
#
?
Apr 25, 2018 05:37
|
|
|
GRINDCORE MEGGIDO posted:Dumb question: what is it about avx instructions that causes so much additional heat? Really dense matrix calculations done on lots of data per clock tick, more or less. There are AVX specific registers, and you can do a ton of floating point vector math in not a whole lot of time. Same way a welder has a duty cycle, the processor has more burst capacity than steady state capacity, so really computationally dense instructions running for long periods will thermal throttle it.
|
|
#
?
Apr 25, 2018 05:48
|
|
|
GRINDCORE MEGGIDO posted:Dumb question: what is it about avx instructions that causes so much additional heat? They do a lot of work. An avx 512 operation could be multiplying 16 pairs of floats in parallel, thats going to light up a lot of silicon
|
|
#
?
Apr 25, 2018 05:49
|
|
|
Thank you, folks.
|
|
#
?
Apr 25, 2018 06:12
|
|
|
For consumer workloads I thought it was always assumed AVX throttled but the performance gains made it not matter. Like you might lose 10% clock speeds but it is 75% faster so who cares.
|
|
#
?
Apr 25, 2018 12:06
|
|
|
craig588 posted:For consumer workloads I thought it was always assumed AVX throttled but the performance gains made it not matter. Like you might lose 10% clock speeds but it is 75% faster so who cares. This is configurable in the bios and usually called "AVX offset". It is very important to understand when overclocking because you might have a stable chip at 5ghz or whatever but then you start running AVX instructions and you start thermal throttling. When I was overclocking my 8700k, setting the avx offset to 3 would drop the cpu 300mhz when performing avx instructions.
|
|
#
?
Apr 25, 2018 14:06
|
|
|
The newest CPU I have is a Haswell so not for me! That was introduced with Broadwell and got more granular with Skylake. I just run my Haswell at 4.3GHz since it keeps it under 70C with AVX/FMA instructions, I don't need the performance of 4.7GHz, not for the extra noise and heat it gives off.
|
|
#
?
Apr 25, 2018 14:13
|
|
|
JawnV6 posted:I dunno, I'm going to get myself sued here, but one of the client/server differences back in my day was a fully bidirectional ring vs. one-way. I don't know what "mesh" means. I thought you were pointing out the difference in the testing setup was a client CPU (Skylake-SP) vs. server CPU (Xeon-W). But I don't deal with those names any more and I'd sorta like to maintain that streak... Yeah I'm not gonna defend these guys but finding software engineers who understand the hardware is really really difficult. Being a cs professor is probably a knock since cs has little to do with actual physical computers. It would be nice if the blog wasnt acting like it actually knew anything
|
|
#
?
Apr 25, 2018 14:13
|
|
|
Methylethylaldehyde posted:Really dense matrix calculations done on lots of data per clock tick, more or less. There are AVX specific registers, and you can do a ton of floating point vector math in not a whole lot of time. These morons decided to run avx-512 (throughput Optimized) for latency sensitive code
|
|
#
?
Apr 25, 2018 14:32
|
|
|
TSMC 7nm is in revenue production Sic transit gloria INTC
|
|
#
?
Apr 25, 2018 18:39
|
|
|
Malcolm XML posted:Sic transit gloria INTC F
|
|
#
?
Apr 25, 2018 18:54
|
|
|
Intel confirms X399 and Z390 chipsets: Also, they're calling AMD's bluff on chipset naming. "Help, my Core i9 won't work on my X399 board!"
|
|
#
?
Apr 25, 2018 20:12
|
|
|
Interesting that the table seems to imply Cannon Lake CPUs might actually come out for those platforms... which were of course originally designed for Cannon Lake before The Great 10nm gently caress Up.
|
|
#
?
Apr 25, 2018 20:52
|
|
|
Is it actually literal Coffee/Cannon Lake, or did they rename the HEDT variant of Cascade Lake because marketing?
|
|
#
?
Apr 25, 2018 22:36
|
|
|
Malcolm XML posted:Yeah wider = slower but it can have throughput. AVX is magical because wider doesn't necessarily mean slower, the latest implementations have 32 registers, and if you have code with a lot of vector math, especially math that can be chained, you can have the entire AVX silicon stack chewing on stuff 24/7.
|
|
#
?
Apr 25, 2018 23:49
|
|
|
Methylethylaldehyde posted:AVX is magical because wider doesn't necessarily mean slower, the latest implementations have 32 registers, and if you have code with a lot of vector math, especially math that can be chained, you can have the entire AVX silicon stack chewing on stuff 24/7. except the point of the whole article was that avx-512 is so intensive it causes core throttling. so even though each instruction processes 16-32x you process instructions slower...
|
|
#
?
Apr 26, 2018 00:28
|
|
|
Ahahahaquote:“Today is Jim Keller’s last day at Tesla, where he has overseen low-voltage hardware, Autopilot software and infotainment. Prior to joining Tesla, Jim’s core passion was microprocessor engineering and he’s now joining a company where he’ll be able to once again focus on this exclusively. We appreciate his contributions to Tesla and wish him the best.” Ahahaha
|
|
#
?
Apr 26, 2018 06:03
|
|
|
Jim Keller: the wandering optimization samurai without a home Do people think it's CPU side or GPU side? I could see either being valid with a man of his talents.
|
|
#
?
Apr 26, 2018 06:07
|
|
|
Place bets on when Keller winds up at Via
|
|
#
?
Apr 26, 2018 06:10
|
|
|
WhyteRyce posted:Ahahaha Holy poo poo, that's hilarious if he is joining Intel. How do I buy stock in Jim Keller? Krailor fucked around with this message at 15:31 on Apr 26, 2018 |
|
#
?
Apr 26, 2018 06:18
|
|
|
Krailor posted:How do I buy stock in Jm Keller? You have entered a satellite office of a silicon giant. The receptionist sees your road-weary party and sighs at you. [work] [party members] [enter bar] [use item] [leave] Paul MaudDib fucked around with this message at 07:39 on Apr 26, 2018 |
|
#
?
Apr 26, 2018 07:31
|
|
|
Just further proof that Intel’s been out of gas since Sandy Bridge, just like AMD with excavator. Time to call up Jim.
|
|
#
?
Apr 26, 2018 07:37
|
|
|
Not the first time Tesla's made a celebrity hire and then had them leave shortly thereafter when they realized Elon Musk is a loving idiot
|
|
#
?
Apr 26, 2018 09:08
|
|
|
Yeah, that Elon guy can't make money at all, sheesh!
|
|
#
?
Apr 26, 2018 09:09
|
|
|
I will be very surprised if this somehow doesn't end in Krzanich's ejection from corporate because holy poo poo things have got to be really dire for him to finally consider signing on.
|
|
#
?
Apr 26, 2018 09:27
|
|
|
What's the backstory to Jim Keller?
|
|
#
?
Apr 26, 2018 09:28
|
|
|
Zedsdeadbaby posted:What's the backstory to Jim Keller?
|
|
#
?
Apr 26, 2018 09:31
|
|
|
LRADIKAL posted:Yeah, that Elon guy can't make money at all, sheesh! lol if you think "has made lots of money" correlates well to "is actually smart, or ethical"
|
|
#
?
Apr 26, 2018 09:33
|
|
|
BobHoward posted:lol if you think "has made lots of money" correlates well to "is actually smart, or ethical" Who said anything about ethical? Keep telling yourself that you're the smart one and Elon is the stupid one, that'll get you far.
|
|
#
?
Apr 26, 2018 09:36
|
|
|
Zedsdeadbaby posted:What's the backstory to Jim Keller? Guy designed DEC Alpha, then reassembled his team and designed AMD Athlon, then coauthored the x86-64 spec, then was a VP at a company. Apple acquired him, and he designed a couple ARM processors for them. Then he came back to AMD and designed Zen, before leaving to Tesla to design [presumably] a bunch of deep learning poo poo for them. And now he's been acquired by Intel. The guy has personally designed about 5 of the 10 last groundbreaking compute architectures over the last 20 years. Paul MaudDib fucked around with this message at 09:43 on Apr 26, 2018 |
|
#
?
Apr 26, 2018 09:37
|
|
|
LRADIKAL posted:Who said anything about ethical? Keep telling yourself that you're the smart one and Elon is the stupid one, that'll get you far. Musk has plenty of dumb ideas in the transportation sector, at least.
|
|
#
?
Apr 26, 2018 09:39
|
|
|
Everyone has dumb ideas! All the time! What makes this guy successful though? Some kind of intelligence, right? It's not ALL luck!
|
|
#
?
Apr 26, 2018 10:02
|
|
|
Oh boy, now we're going to have to do the "great man" theory vs the "hundreds of engineers under him" theory.
|
|
#
?
Apr 26, 2018 10:05
|
|
|
Laslow posted:Just further proof that Intel’s been out of gas since Sandy Bridge, just like AMD with excavator. Time to call up Jim. To be fair, both AMD and Intel are out of ideas - their only play at the moment is shoehorning as many cores as they can onto substrate. I mean, are we going to see LGA16528 someday?
|
|
#
?
Apr 26, 2018 10:15
|
|
|
Paul MaudDib posted:Guy designed DEC Alpha, then reassembled his team and designed AMD Athlon, then coauthored the x86-64 spec, then was a VP at a company. Apple acquired him, and he designed a couple ARM processors for them. Then he came back to AMD and designed Zen, before leaving to Tesla to design [presumably] a bunch of deep learning poo poo for them. And now he's been acquired by Intel. So he really is a lone tech samurai wandering the lands
|
|
#
?
Apr 26, 2018 10:17
|
|
|
LRADIKAL posted:Who said anything about ethical? Keep telling yourself that you're the smart one and Elon is the stupid one, that'll get you far. Musk thinks that the biggest near term risk to humanity is the possibility that someone will invent a self-improving AI that will make itself into a malevolent godlike being which will then Skynet us. No, seriously, he actually thinks that, and promotes the idea with gusto. He's a dipshit. But hey, keep clinging to that just-world fallacy.
|
|
#
?
Apr 26, 2018 12:05
|
|
|
LRADIKAL posted:Who said anything about ethical? Keep telling yourself that you're the smart one and Elon is the stupid one, that'll get you far. Tesla is in deep poo poo if it can't get its manufacturing on track before 2020. Autopilot is a failure--it's less developed than off the shelf safety tech. The model 3 is a beta car that's insanely hard to produce and has poor fit and finish despite having great bones Elon's dumbass strategic leadership delayed their electric suv by years in market where sedans are dying. Need I go on? Musk has had some successes but he isn't infallible and I'm not surprised Keller bailed out.
|
|
#
?
Apr 26, 2018 12:23
|
|
|
It's like a handful of software engineers in the 90s got absurdly lucky with regards to borth time and place that they made such an absurd amount of money that they're insolated from any future failures and as such have become convinced they're hyperintelligent Supermen, masters of all, instead of the one thing they know anything about (software engineering), and seeing as it's been 20 years, probably don't know much about at this point.
|
|
#
?
Apr 26, 2018 13:01
|
|
|
|
| # ? Apr 28, 2024 04:40 |
|
|
BIG HEADLINE posted:To be fair, both AMD and Intel are out of ideas - their only play at the moment is shoehorning as many cores as they can onto substrate. I mean, are we going to see LGA16528 someday?
|
|
#
?
Apr 26, 2018 13:40
|
|