|
kefkafloyd posted:Congrats. Cambridge MC is my local store, I live about twenty minutes away. I wish they would do some of this in evening hours for us working stiffs. Unless I take a week off from work there's no chance I'd be able to spend the time to be able to do it. theyve done saturday drops at other ones...
|
#
?
Jun 28, 2021 22:18
|
|
|
|
| # ? Apr 28, 2024 16:20 |
|
|
it'll be interesting to see how it goes because the distributors (and in some cases the brands themselves) have marked everything way up and retailers are going to be loathe to take a big loss. it is definitely true that (especially towards the end) a lot of retailers were taking "dumb" prices just to get cards. Basically pre-scalped, pretty much the same prices you would have paid on ebay, and they had to try and sell them, or just not take the cards. to some extent the loss is unavoidable though because prices are what they are, if retailer A gets new inventory at $1700 then retailer B can't just hold onto their inventory at $2000 forever hoping for a sucker because they just won't sell any ever. people don't really think about it that much in tech because prices normally don't move up and down all that much, but the reality is that the market clearance price is what it is, if you paid $1500 for a GPU and tomorrow prices increase and you can sell it for $2000, then the inventory you hold has appreciated by $500. And vice versa, if prices go down then you lost money on that inventory. To some extent the money is already gone, although in some cases you can find a sucker who will buy at least at a slightly inflated price. it's much more common in stuff like gas stations - you're holding $30k of gasoline just to do your daily business. If prices rise? You made a bunch of profit on the gas sitting in your tank. If prices fall? Lol you just lost a bunch of money. if things get too wacky maybe you'll see NVIDIA step in and rebate the retailers for some of the value they've lost in their inventory, although there's also the possibility they're not too receptive to that after all the scalping shenanigans (10 months of 10x the usual margin and you're crying about being stuck with 50 cards you'll have to take a $200 loss on? lol get hosed)
|
|
#
?
Jun 28, 2021 22:27
|
|
|
Anyone know how using a laptop with a discrete GPU, with an eGPU, would work as far as drivers and the like? RTX3050ti as the built in, RTX3070 as the eGPU, if it matters.
|
|
#
?
Jun 28, 2021 22:33
|
|
|
shrike82 posted:All video cards are deep learning accelerators - tensor cores specifically does optimised lower precision math. Nvidia still doesn’t support bfloat16 on Ampere so it’s not like they’re leading in lower precision inference support either.
|
|
#
?
Jun 28, 2021 23:08
|
|
|
ilkhan posted:Anyone know how using a laptop with a discrete GPU, with an eGPU, would work as far as drivers and the like? RTX3050ti as the built in, RTX3070 as the eGPU, if it matters. It should just work. That is if all the YouTube videos I’ve seen are to be believed.
|
|
#
?
Jun 28, 2021 23:11
|
|
|
Harik posted:wait are "tensor cores" really nothing but nVidia lifting the 1/64th rate limitation on fp16 for a handful of their SMs because lol if so. The Tensor/RT cores are sprinkled in each SM, not lumped together in specific SMs as i understand it. But in general, my extreme lay person understanding is that CUDA/Tensor/RT are all similar on the silicon side, they are all just ALUs. But the Tensor units are a fixed-function ALU cluster with shared controlling logic (making them more efficient for the die area), vs individually addressable and mixed-precision ALUs in the case of CUDA cores. The magic from what I can tell is how it all works at the subcore and scheduler level which is way out of my depth. AT put it like: quote:And if this sounds familiar to a normal GPU ALU pipeline, then it should. Tensor cores, while being brand-new to the GPU space, are not all that far removed from standard ALU pipelines. The density has changed – they're now operating on sizable matricies instead of SIMD-packed scalar values – but the math has not. At the end of the day there's a relatively straightforward tradeoff here between flexibility (tensor cores would be terrible at scalar operations) and throughput, as tensor cores can pack many more operations into the same die area since they are so rigid and require a fraction of the controlling logic when that cost is divided up per ALU. https://www.anandtech.com/show/12673/titan-v-deep-learning-deep-dive/3
|
|
#
?
Jun 29, 2021 00:18
|
|
|
Also DG2 is apparently sampling. We might see products sooner rather than later, but I almost guarantee they are laptops to start. https://twitter.com/VideoCardz/status/1409551050995085312
|
|
#
?
Jun 29, 2021 00:25
|
|
|
The 1:64 cap is applied to FP64 ops, Nvidia caps FP16 performance in a different way by handicapping FP32 accumulation throughput effectively lowering Geforce FP16 performance by 10-20% versus their enterprise cards. The tensor cores are distinct from SMs and are heavily optimized for the lower precision math that's used in deep learning. With Ampere for example, they've tweaked the cores further to do sparse math efficiently. Deep Learning is a young research field where the AI architectures tend to follow from available hardware i.e., researcher design their models around whatever gear they have. I guess my point here is that people shouldn't assume that tensor cores (and its optimizations) are the only way to run stuff like DLSS efficiently.
|
|
#
?
Jun 29, 2021 00:31
|
|
|
You're almost certainly correct. Even if the specific implementation of DLSS sees huge performance benefits from tensor cores, it's unlikely that an implementation optimized for general purpose GPU compute would be that much slower or worse. The question that's more interesting to me is the power efficiency one - how energy efficient is DLSS? And how will Nvidia (and AMD) allocate silicon on future generations of GPUs?
|
|
#
?
Jun 29, 2021 00:35
|
|
|
I wonder if Facebook are still working on that DLSS-alike they published last year It was way too slow to really be usable back then, but they've had a full year and infinity money to keep hammering on it since then if they're so inclined
|
|
#
?
Jun 29, 2021 00:45
|
|
|
Paul MaudDib posted:it's much more common in stuff like gas stations - you're holding $30k of gasoline just to do your daily business. If prices rise? You made a bunch of profit on the gas sitting in your tank. If prices fall? Lol you just lost a bunch of money. The same will happen with GPUs: as wholesale prices top increasing and start to go down a bit, retailers may initially drop to match because nobody wants to be the only one selling at the old peak prices, but after the first correction or two everybody is going to take their time letting retail prices fall to match the new normal.
|
|
#
?
Jun 29, 2021 04:11
|
|
|
shrike82 posted:The 1:64 cap is applied to FP64 ops, Nvidia caps FP16 performance in a different way by handicapping FP32 accumulation throughput effectively lowering Geforce FP16 performance by 10-20% versus their enterprise cards. there's also been a cap on fp16 performance, it was 1:128 of fp32 on pascal for some ops, 1:64 on others. They may have relaxed that on the CUDA cores in newer architectures though. Notably the fp16 performance on a 1080 was so dogshit it was only usable to make sure your code had no bugs before running it on "real" compute hardware.
|
|
#
?
Jun 29, 2021 05:46
|
|
|
Harik posted:there's also been a cap on fp16 performance, it was 1:128 of fp32 on pascal for some ops, 1:64 on others. They may have relaxed that on the CUDA cores in newer architectures though. Notably the fp16 performance on a 1080 was so dogshit it was only usable to make sure your code had no bugs before running it on "real" compute hardware. do games just not use these datatypes at all or something?
|
|
#
?
Jun 29, 2021 05:51
|
|
|
Harik posted:there's also been a cap on fp16 performance, it was 1:128 of fp32 on pascal for some ops, 1:64 on others. They may have relaxed that on the CUDA cores in newer architectures though. Notably the fp16 performance on a 1080 was so dogshit it was only usable to make sure your code had no bugs before running it on "real" compute hardware. yes specifically FP32 accumulation like i mentioned which is a step in most mixed precision training for AI models. the numbers sound bad but in real world use-cases results in a 10-20% hit in throughput
|
|
#
?
Jun 29, 2021 05:57
|
|
|
Shipon posted:do games just not use these datatypes at all or something? They would if perf wasn't dogshit  There's quite a few cases where you don't really need the full FP32 precision and FP16 would in theory be a perf boost, in practice because of the way FP16 is handicapped on desktop, it's only really ever relevant on mobile.
|
|
#
?
Jun 29, 2021 09:04
|
|
|
repiv posted:I wonder if Facebook are still working on that DLSS-alike they published last year Anything Zuckerberg can & should be dismissed out of hand
|
|
#
?
Jun 29, 2021 09:50
|
|
|
Zedsdeadbaby posted:Anything Zuckerberg can & should be dismissed out of hand This is a silly take in a world where Facebook owns
|
|
#
?
Jun 29, 2021 10:58
|
|
|
nah, oculus should also be dismissed, don't buy into the propaganda of those 2 insane vr thread posters lol
|
|
#
?
Jun 29, 2021 11:49
|
|
|
Zedsdeadbaby posted:Anything Zuckerberg can & should be dismissed out of hand On second thoughts they probably aren't pursuing that anyway after their hard pivot to standalone/mobile VR, they'd not only have to replicate DLSS but also make it an order of magnitude or two faster for it to make any sense on the Quest
|
|
#
?
Jun 29, 2021 12:16
|
|
|
They have the resources & funding to pull it off, I just don't trust them as a company at all. It's hard enough looking after my privacy/data while using microsoft/google software, I don't need to add facebook to the mix
|
|
#
?
Jun 29, 2021 12:47
|
|
|
blunt posted:This is a silly take in a world where Facebook owns Is that what it's actually called? That's very obviously a villainous company name. More than the regular capitalism kind.
|
|
#
?
Jun 29, 2021 13:34
|
|
|
Falcorum posted:They would if perf wasn't dogshit And the entire reason this type of math is performance-capped is to push purchases of their ripoff Quadros and such right? No technical reason for it otherwise?
|
|
#
?
Jun 29, 2021 16:34
|
|
|
Partially yes, FP64 is likely overkill for games, and FP16 (AMD called it rapid packed math for vega) is basically not used because it's not fast. It also would require extensive engine changes as you can't just swap the two datatypes. And yes, it's product segmentation as to why the standard video cards don't have it.
|
|
#
?
Jun 29, 2021 16:58
|
|
|
NewFatMike posted:Is that what it's actually called? That's very obviously a villainous company name. More than the regular capitalism kind. Zuck's "philanthropic" arm is called The Zuckerberg Chan Initiative and that's definitely the name of an organization that won't get you nerve-stapled and sent to the offworld colonies.
|
|
#
?
Jun 29, 2021 16:59
|
|
|
Good news, Newegg has graphics cards for sale again! https://www.newegg.com/asrock-radeo...&quicklink=true Also it is kind of annoying that the exact model 3060 ti I signed up to be notified about from EVGA is in the shuffle today. I don't even need it anymore but could have given it to a friend or something.
|
|
#
?
Jun 29, 2021 18:05
|
|
|
Sweet only $2300! That's only like a 130% markup right?
|
|
#
?
Jun 29, 2021 18:43
|
|
|
Anyone want to purchase a 3090 ftw3 hydro copper? I just got a notify.
|
|
#
?
Jun 29, 2021 19:13
|
|
|
Dibs depending on price? What do you have in mind
|
|
#
?
Jun 29, 2021 20:13
|
|
|
Am I confused about DLSS or just being misled? I've tried it in a couple games and it only seems to make things worse. I thought it had been worded somewhere that it would help offset the tax of raytracing. I usually smash up against the 144fps limit @ 2560 in Doom Eternal, just downloaded the new patch and raytracing seems to cap me at 60, DLSS brings that down to 23-30. This is on a 2080.
|
|
#
?
Jun 29, 2021 21:32
|
|
|
I was just coming here to post about that, the Doom Eternal update has a performance cliff on 8GB cards They can just about handle it on Ultra Nightmare normally but adding raytracing on top pushes it over the edge and kills performance https://www.nvidia.com/en-us/geforce/news/doom-eternal-ray-tracing-nvidia-dlss-upgrade-available-now/ quote:When applying graphics presets like “Ultra Nightmare,” enabling ray tracing for GPUs with 8 GB or 6 GB frame buffers requires enabling DLSS and adjusting the texture pool size to fit within allocated memory. Try backing off "texture pool size" and "shadow quality" to just Ultra pack it in 3070ailures repiv fucked around with this message at 22:14 on Jun 29, 2021 |
|
#
?
Jun 29, 2021 21:40
|
|
|
DE raytracing is a mixed bag, on the one hand it doesn't improve the visuals that much (it feels like the artists went out of their way to avoid super shiny surfaces despite the sci-fi setting, so the default cheap raster reflections wouldn't stick out) but on the other hand I'm getting >100fps at 1440p on a 2070 Super so why not enable it
|
|
#
?
Jun 29, 2021 22:36
|
|
|
Doom Eternal really is a polished, optimized game. God drat.
|
|
|
#
?
Jun 29, 2021 22:39
|
|
|
I love this dumb little AI toy
|
|
#
?
Jun 29, 2021 22:39
|
|
|
Cream-of-Plenty posted:Doom Eternal really is a polished, optimized game. God drat. I should note that 100fps number is with DLSS Quality enabled (so 960p internal) The DLSS implementation seems solid from the bit I've played though, looks nice
|
|
#
?
Jun 29, 2021 22:43
|
|
|
repiv posted:Try backing off "texture pool size" and "shadow quality" to just Ultra Ok that worked
|
|
#
?
Jun 29, 2021 22:49
|
|
|
https://videocardz.com/newz/battlefield-2042-will-support-nvidia-dlss-and-reflex-technologies DLSS and Nvidia Reflex confirmed but absolute crickets when it comes to RTX  quote:
|
|
#
?
Jun 29, 2021 23:07
|
|
|
If you mean raytracing, that will be a novelty side feature for a good long while. It has to be implemented on the engine level and whole workflows will have to be adjusted around it.
|
|
#
?
Jun 29, 2021 23:09
|
|
|
Raytracing isn't really an "nvidia feature" anymore so it's possible they just omitted it there They had it in the last Battlefield so it would be weird for it to get dropped
|
|
#
?
Jun 29, 2021 23:12
|
|
|
Ah, that's a good point, too. I haven't played a Battlefield in ages.
|
|
#
?
Jun 29, 2021 23:12
|
|
|
|
| # ? Apr 28, 2024 16:20 |
|
|
I have a feeling they don't want to be pigeon-holed into including ray-tracing for the sake of it, it's horribly expensive and they're trying to get as much performance as they can for those intensive 128 player matches. I'm sure it'll have some form of it perhaps, but I would not be surprised if they focused on performance and left ray-tracing out.
|
|
#
?
Jun 29, 2021 23:17
|
|