|
gently caress Moore's law, great integrated graphics for less than $90 is where it's loving at Gordon Moore can suck it, we're a society of web browsing weebs now, give us our integrated auxiliary acceleration resources I want to watch scott manley send green men to space in 4K 60FPS and I want to do it on a machine that costs less than an Ikea mattress Potato Salad fucked around with this message at 19:09 on May 5, 2018 |
#
?
May 5, 2018 19:06
|
|
|
|
| # ? Apr 26, 2024 21:09 |
|
|
Moore just cares about transistor count. We know there are more factors in modern performance that just that count. It’s really not a big deal.
|
|
#
?
May 5, 2018 20:57
|
|
|
Just give me the hardware to run VRchat and I'm happy
|
|
#
?
May 5, 2018 21:46
|
|
|
SourKraut posted:I'm sure the answer is "We don't know", but is there any sense of how much of an improvement Ryzen 2 might be over Ryzen+? The R7 2700X is enticing... Likely a significant improvement TBH, Pinnacle Ridge is merely what AMD wanted Summit Ridge to be, Matisse looks to be a major overhaul.
|
|
#
?
May 5, 2018 22:54
|
|
|
SourKraut posted:I'm sure the answer is "We don't know", but is there any sense of how much of an improvement Ryzen 2 might be over Ryzen+? The R7 2700X is enticing... We don’t know.
|
|
#
?
May 6, 2018 00:03
|
|
|
Subjunctive posted:Moore just cares about transistor count. We know there are more factors in modern performance that just that count. It’s really not a big deal. That's the best kind of correct, but the practical effect of doubling transistor count was doubling performance. The world of processors now lines up with the world of batteries: single digit percent gains instead of 200% gains
|
|
#
?
May 6, 2018 00:06
|
|
|
Rastor posted:That's the best kind of correct, but the practical effect of doubling transistor count was doubling performance. In Moore’s paper, he talks about doubling every year in the context of chip complexity as it relates to yield and ultimately cost. The double-perf-every-18-months comes from a guy at Intel who was talking about the combination of transistor count and performance per transistor. (Moore’s paper is fun to read, though shallow. Recommend!) We see double-digit-or-more gains in performance between generations in GPUs (500-1100% on DL tasks between Pascal and Volta) and mobile SoCs (25-50% for A7-A8). A large portion of the world’s processor applications are still seeing excellent gains.
|
|
#
?
May 6, 2018 00:35
|
|
|
Subjunctive posted:500-1100% on DL tasks between Pascal and Volta) LMAO, I'm curious what "DL tasks" these are. Most benchmarks show a 10-30% difference for training/inferencing between the Titan V and the 1080 Ti. shrike82 fucked around with this message at 01:17 on May 6, 2018 |
|
#
?
May 6, 2018 01:13
|
|
|
shrike82 posted:LMAO, I'm curious what "DL tasks" these are. Looks like ResNet-50, though I can’t find the same numbers. Still up to 3x, though. https://www.nvidia.com/en-us/data-center/tensorcore/
|
|
#
?
May 6, 2018 01:23
|
|
|
yeah that's bullshit marketing which isn't borne out in actual benchmarks. We've actually looked at switching from 1080 Tis to the Titan Vs for our Dev ML servers, the latter aren't much faster. A big part is the fact that Nvidia crippled FP16 throughput for consumer cards - but that's not used in actual training/inferencing.
|
|
#
?
May 6, 2018 01:28
|
|
|
Do you mean that the ResNet results aren’t going to be reproducible, or that they aren’t representative (possibly because most workloads don’t code to the tensor cores)?
|
|
#
?
May 6, 2018 01:29
|
|
|
Well I'm guessing you're not a technical guy - 1. Don't take marketing "benchmarks" at face value especially when the X-axis is "time to solution" (presumably convergence) and that can be manipulated 2. They compare the V100 to the P100 when the 1080 Ti is faster than the P100 and cheaper from a TCO standpoint if you're doing dev ML work 3. This is an example of a real-world benchmark  4. ResNet-50 is a pretty small (i.e., toy) image architecture by modern standards
|
|
#
?
May 6, 2018 01:39
|
|
|
Thanks. I hadn’t thought about manipulation of convergence. Should that affect the inference results? I only know it in terms of establishing weights, but my ML experience is years old and poorly remembered. The point of comparing P100 to V100 was to isolate the inter-generation gap, but I suppose the numbering doesn’t really ensure that. Maxwell-vs-Pascal might have been better because I could have compared the resulting top end of both lines.
|
|
#
?
May 6, 2018 01:48
|
|
|
Rastor posted:That's the best kind of correct, but the practical effect of doubling transistor count was doubling performance. The reason CPU performance has stalled isn't because transistor counts are stalled, as Subjunctive pointed out other areas like GPUs are still seeing big gains every generation. If you could just keep adding more cores like you can with a GPU we'd still be seeing big gains, but that approach doesn't work for CPUs because you need to balance single and multi-threaded performance. CPU performance gains have stalled because clock speed improvements have stalled and IPC gains are not even close to being linear with increased transistor counts of the individual cores. Basically it comes down to the fact that further increases in single-threaded IPC at this point are really difficult. MaxxBot fucked around with this message at 07:43 on May 6, 2018 |
|
#
?
May 6, 2018 07:37
|
|
|
MaxxBot posted:The reason CPU performance has stalled isn't because transistor counts are stalled, as Subjunctive pointed out other areas like GPUs are still seeing big gains every generation. If you could just keep adding more cores like you can with a GPU we'd still be seeing big gains, but that approach doesn't work for CPUs because you need to balance single and multi-threaded performance. CPU performance gains have stalled because clock speed improvements have stalled and IPC gains are not even close to being linear with increased transistor counts of the individual cores. Basically it comes down to the fact that further increases in single-threaded IPC at this point are really difficult. In other words: Moore's Law is dead.
|
|
#
?
May 6, 2018 11:23
|
|
|
I wonder what the Intel engineers that designed the 8086 back in 1976 would say if they could see how far past the point of diminishing returns we've carried their baby.
|
|
#
?
May 7, 2018 02:22
|
|
|
Probably quite confused about meltdown.
|
|
#
?
May 7, 2018 02:29
|
|
|
This is probably a dumb question but anyone have any idea why fully modular motherboard designs like this are not becoming more of a thing yet? Or is it just due to the motherboard and cpu cartel and/or engineering difficulty? Seems like a great way to make everything backwards/forward compatible, significantly reduce costs, and reduce waste as you could just keep adding on stuff as tech progresses. But, I'm guessing getting an actual frankenstein computer to work would be a true nightmare in terms of ram frequency differences, somehow getting to completely different cpu sockets to work together, and everything else.
|
|
#
?
May 7, 2018 04:37
|
|
|
I would think that a fully modular motherboard like that would be both an engineering and software nightmare due to the mishmash of connection types, PCB traces, and hardware variables.
|
|
#
?
May 7, 2018 04:46
|
|
|
Avalanche posted:This is probably a dumb question but anyone have any idea why fully modular motherboard designs like this are not becoming more of a thing yet? Or is it just due to the motherboard and cpu cartel and/or engineering difficulty? It can only last as many CPU upgrade cycles as a normal motherboard can, since you'd need to upgrade most of the parts to not be wasting a CPU upgrade anyway. Plus you'd have to pay more for all the parts from the start, broken out as they are. Really, things like that are designed to support goals like "fit into odd shaped case for a particular task" or "provide easy component replacement for ease of repair".
|
|
#
?
May 7, 2018 04:48
|
|
|
Bloody Antlers posted:I wonder what the Intel engineers that designed the 8086 back in 1976 would say if they could see how far past the point of diminishing returns we've carried their baby. It was only 40 years ago. Stephen Morse, Bill Pohlman, and Bruce Ravenel are still alive. Morse did an interview with PC World for the 30th anniversary. He doesn't talk much about modern CPUs but did mention that if they had known that x86 was going to stick around rather than just be a stopgap until the 432 could be released, they'd have done some things differently.
|
|
#
?
May 7, 2018 05:23
|
|
|
Theris posted:It was only 40 years ago. Stephen Morse, Bill Pohlman, and Bruce Ravenel are still alive. Morse did an interview with PC World for the 30th anniversary. He doesn't talk much about modern CPUs but did mention that if they had known that x86 was going to stick around rather than just be a stopgap until the 432 could be released, they'd have done some things differently. I worked with Steve Morse post-Intel (mid/late 90s). He’s much better at hardware than software!
|
|
#
?
May 7, 2018 09:38
|
|
|
Avalanche posted:This is probably a dumb question but anyone have any idea why fully modular motherboard designs like this are not becoming more of a thing yet? Or is it just due to the motherboard and cpu cartel and/or engineering difficulty? Most mobo manufactures have done this for years. Nothing new bro. Only difference is the manufacturer does it, not you.
|
|
#
?
May 8, 2018 00:21
|
|
|
Rastor posted:It's just now moving into higher power situations (Examples: Qualcomm Centriq, Cavium ThunderX2), which may be part of why Intel feels threatened enough to bring in Jim Keller. This was a few pages back, but uhhh Qualcomm Plans Exit From Server Chips
|
|
#
?
May 9, 2018 11:13
|
|
|
repiv posted:This was a few pages back, but uhhh Interesting, maybe Carlyle Group would be interested in adding it to their Ampere Computing startup. Yesterday Qualcomm announced a new smartwatch SoC, so I'm guessing they see the low-power market as their strength and decided to focus on it.
|
|
#
?
May 9, 2018 11:40
|
|
|
Relevant to the ARM server discussion, STH just released their testing of a ThunderX2 system and it's actually.... really good. Beats both Epyc and Scalable in perf/$ and perf/watt in a lot of tests. Kinda didn't expect that, but thats my own x86 bias.quote:The Cavium ThunderX2 that we see today has its origins in the Broadcom Project Vulcan and so many of the features we saw in Cavium ThunderX, such as 40GbE ports, are not present. Instead, we have an Arm chip that can go toe-to-toe with Intel and AMD and come out ahead in some cases. Best of all, the list price of the 32 core top-bin CN9980 part is $1795 about half of the competitive Intel and AMD chips. https://www.servethehome.com/cavium-thunderx2-review-benchmarks-real-arm-server-option/ But since I'm just a nerd and not in this for a living, baby, i wanna see them glamor shots. Huge rear end socket oh baby yeah:  
|
|
#
?
May 9, 2018 22:50
|
|
|
So how are we supposed to benefit from the 128 pcie lanes on Epyc? What sort of performance could be squeezed out of the hardware?
|
|
#
?
May 10, 2018 01:47
|
|
|
A gargantuan number of PCIe SSDs. 
|
|
#
?
May 10, 2018 03:14
|
|
|
Split out the pcie lanes and mine with 100+ cards at once, this is a fantastic idea.
|
|
#
?
May 10, 2018 03:17
|
|
|
GRINDCORE MEGGIDO posted:Split out the pcie lanes and mine with 100+ cards at once, this is a fantastic idea. I could see 128 threads in Epyc being used to produce bulk dried strawberries.
|
|
#
?
May 10, 2018 05:45
|
|
|
Cygni posted:Relevant to the ARM server discussion, STH just released their testing of a ThunderX2 system and it's actually.... really good. Beats both Epyc and Scalable in perf/$ and perf/watt in a lot of tests. Kinda didn't expect that, but thats my own x86 bias. Kind buried an important part here. quote:We were told that our power consumption results may be 10-15% higher than shipping parts. quote:Our Gigabyte/ Cavium ThunderX2 Saber development platform hit a peak of 823W at 100% load. That's dual socket, what's the known peak draw for Intel and EPYC in dual socket? I feel like it's about 500W, but don't know. So while cheap and effective, it seems insanely hungry. Also, those single thread results are pretty good considering it's clock deficiency. https://www.servethehome.com/amd-epyc-7601-dual-socket-early-power-consumption-observations/ - 483W for EPYC dual socket 7601. I'm assuming Intel is going to be comparable or better with an 8160 or 8180 Platinum. EmpyreanFlux fucked around with this message at 07:09 on May 10, 2018 |
|
#
?
May 10, 2018 07:00
|
|
|
Devian666 posted:I could see 128 threads in Epyc being used to produce bulk dried strawberries. You joke, but. https://www.youtube.com/watch?v=1WLIm4XLPAE
|
|
#
?
May 10, 2018 14:47
|
|
|
Possible 2950X in the wild http://ranker.sisoftware.net/show_run.php?q=c2ffc9ef8eefd2ebdfecdaeec8ba87b791f491ac9cbac9f4cc&l=en 3.89Ghz, 16C/32T, 117W 
|
|
#
?
May 10, 2018 15:26
|
|
|
FaustianQ posted:Possible 2950X in the wild http://ranker.sisoftware.net/show_run.php?q=c2ffc9ef8eefd2ebdfecdaeec8ba87b791f491ac9cbac9f4cc&l=en Inter-Core Latency 179.4ns That seems high even for amd with its inter ccx hops.
|
|
#
?
May 10, 2018 15:59
|
|
|
Pretty normal values for TR looking at their 1950X benchmarks, it kind of is a rebadged server product after all. The target audience won't mind the increased latency and there's the option of disabling half the cores to keep latency sensitive applications on one die.
|
|
#
?
May 10, 2018 16:24
|
|
|
Kazinsal posted:A gargantuan number of PCIe SSDs. That's legit one of the best uses of the thing, a ridiculous cache box that sits right at the provider edge, with a 1TB of ram, 32 Nvme drives, and enough processing power to flood the twin 40GbE ports. 400k+ IOPS, 57+ GB/sec of aggregate throughput at 128KB, 7 million IOPS and 53 GB/sec at 4k random.
|
|
#
?
May 10, 2018 17:21
|
|
|
FaustianQ posted:Possible 2950X in the wild http://ranker.sisoftware.net/show_run.php?q=c2ffc9ef8eefd2ebdfecdaeec8ba87b791f491ac9cbac9f4cc&l=en So what freq would that be at the full beans (180W)?
|
|
#
?
May 10, 2018 18:47
|
|
|
Seamonster posted:So what freq would that be at the full beans (180W)? Probably 4.4Ghz on like 8 cores.
|
|
#
?
May 10, 2018 18:54
|
|
|
I don't think there's an easy formula for something like that, but if you compare the 1950X to 1800X and then look at how much improvement there was from 1800X to 2700X I wouldn't be surprised to see a 2950X that can turbo to 4.5.
|
|
#
?
May 10, 2018 18:54
|
|
|
|
| # ? Apr 26, 2024 21:09 |
|
|
I watched that last night. From the building work that I do those old radiators are mostly rust with a little bit of metal. I wasn't surprised to see the state of the water after heating that storage room.
|
|
#
?
May 10, 2018 23:07
|
|