|
hobbesmaster posted:Depending on how they implement it there may be a massive performance issue. 12GB is probably perfectly fine for the 4070’s performance targets and isn’t a limit right now based on the 3080 10GB and 12GB doing okayish on the recent bad ports. The 3080 has considerably more memory bandwidth than the 4070. Josh Lyman posted:This D3 discussion is somehow giving me buyer's remorse about ordering a 4070 this morning so I did some back of the napkin calculations. I don't have a crystal ball, but my guess is that the 7800xt will be on par with the 6950xt but significantly more power efficient and likely better in ray tracing. The 7900xt is getting pushed below $800, the 7800xt will have to slot in below that, price wise. Fsr is a thing, though dlss is better at the moment; frame generation for the rdna 3 is also and eventuality, though who knows when. Raytracing will be better on an rtx, but it requires vram which is already a problem in some titles--what I am getting at is there is no slam dunk like the 3080 or 1080ti in this generation of cards.
|
# ? May 9, 2023 20:36 |
|
|
# ? May 30, 2024 14:35 |
|
Yudo posted:The 3080 has considerably more memory bandwidth than the 4070. So speaking of that here’s a pretty good video with benchmarks. https://youtu.be/1IKHrdQecfA
|
# ? May 9, 2023 20:38 |
|
We're still pretty in the dark about what to expect from the Navi 32 chips that'll be going into the 7800 XT / 7700 XT, most of the info I can find is pure speculation but the consensus seems to be that the chips will have 60 compute units on the 7800 XT and 54 on the 7700 XT (compared to 96 on the 7900 XTX and 84 on the 7900 XT) and 16gb / 12gb memory with 256-bit / 192-bit buses and 64mb / 48mb infinity caches, respectively It's gonna be hard to speculate on their value until we know some prices and have seen some reviews
|
# ? May 9, 2023 20:40 |
|
DoombatINC posted:We're still pretty in the dark about what to expect from the Navi 32 chips that'll be going into the 7800 XT / 7700 XT, most of the info I can find is pure speculation but the consensus seems to be that the chips will have 60 compute units on the 7800 XT and 54 on the 7700 XT (compared to 96 on the 7900 XTX and 84 on the 7900 XT) and 16gb / 12gb memory with 256-bit / 192-bit buses and 64mb / 48mb infinity caches, respectively
|
# ? May 9, 2023 21:13 |
|
Josh Lyman posted:I was under the impression the 7800XT would be Navi 31 like with the 6000 series? Nope, at least according to the internet rumor mill - Navi 32 for the 7800 and 7700, Navi 33 (which is gonna be a monolithic) for the 7600 and below
|
# ? May 9, 2023 21:16 |
|
Yudo posted:The 3080 has considerably more memory bandwidth than the 4070. 1080 Ti has considerably more memory bandwidth than a 2070S. 5700XT has more bandwidth than a 6700XT. There is more to a design than memory bandwidth, or die size, or any of the metrics that enthusiasts are fussing about the number going down on. "Effective memory bandwidth" is kind of the number that matters, but it also depends on cache hitrates and occupancy (because unused L1/L2 is also used as another cache level too). It's complicated. Like yes, a design based on exploiting TSMC’s sense SRAM/cache looks very different than a fat wide Samsung die on a 10nm-family smartphone node with half the logic density and 1/4 the SRAM density. That is not surprising at all. RDNA3 lets AMD physically fan out quite a bit more than Ada in general. And that’s both more Infinity Links on die (because they’re small) and also more space to route the signals to the PHY (because the package is big). Nvidia is definitely constrained by the G6/G6X modules only coming in 16Gbit density at most, they have to use clamshell to hit 16gb+ on their smaller packages, and they rely on the tsmc cache density to deliver more bandwidth and make up for less actual PHYs. But it’s also physically smaller and more efficient. Every approach has its pluses and minuses. Nvidia really needs something like GDDR6W, where you can stack multiple memory dies into a single package for higher density. Clamshell is not ideal and will drive up cost/drive down partner margins, but it’s also the only thing nvidia can do with a package with literally 4 gddr channels attached. Paul MaudDib fucked around with this message at 21:50 on May 9, 2023 |
# ? May 9, 2023 21:28 |
|
Yudo posted:The 3080 has considerably more memory bandwidth than the 4070. But a miss is a miss and regardless of how fast the copy is once it begins that’s still a delay. If a game expects everything to be in VRAM as it is on a console then you may have a lot of those misses - all that bandwidth doesn’t help if you need a context switch or whatever.
|
# ? May 9, 2023 21:43 |
|
Tangent but regarding fanout, MLID had an interesting bit recently from a vendor who does design work with ODMs and they said multiple ODMs were looking at the idea of ripping out the dGPUs from some of their ultrabook models and replacing them with more battery, because they can get a 1.4x increase in capacity from eliminating a 4060 and a 1.9x increase from eliminating a 4070. And if you double the battery life of their current ultrabooks, suddenly you are in competition with a MBP for battery life. And then he had another bit pointing out the constraints around the PHYs and I think those are correlated. The way I read it is, NVIDIA is doing the math on how tiny they can make the packages because the tinier the package the less ODMs will gain from ripping them out and replacing them with battery. So you have this focus on small packages with high cache density and extremely high clocks/transfer rates/encoding (GDDR6X with PAM4) and high (lossless) delta compression ratios to make up for narrow memory buses and few memory chips. Like it's all very optimized to produce the smallest package for ODMs for laptops. But you get a 4060 that has one tiny GPU die and four IC packages, and with 16gbit/2GB packages that leaves you doing clamshell to try and hit enthusiast-level capacity on desktop. Long term GDDR6W or 24gbit packages are almost a requirement for NVIDIA, I think. And really they need to look at doing L3 stacking probably too. Put a cache die under/above the GPU die, you can still do that even monolithic and it'll help stretch the memory further. But yeah it's either in cache or not, and if not you incur going to memory. If it's not in memory, you have to go out to the SSD for it with RDMA, or to the host CPU memory, and PCIe is slow. How much do you get across a PCIe 4.0x8 bus in 16ms frametime? Not all that much actually! Fury X found this out the hard way, same for Vega with HBCC. Swapping isn't a substitute for VRAM capacity, unless PCIe gets a lot faster. Also, I am pretty sure transfers are the unit of importance in PCIe. If you have a slow SSD or a slow motherboard, it will eat up more bus utilization than a PCIe 4 SSD/GPU/etc. And the 4060/4060 Ti (and likely AMD's) are also x8. We'll see how much that becomes a bottleneck when it's not just transferring drawcalls but also swapping tons of textures across the same link. I would not buy x8 without a PCIe 4.0 setup imo, mobo+CPU+SSD. Paul MaudDib fucked around with this message at 22:01 on May 9, 2023 |
# ? May 9, 2023 21:58 |
|
Arivia posted:it's kinda crazy that GPU ram needs to equal an entire console's RAM to be competitive all of a sudden This is not all that sudden, is it? I feel like we went through a similar thing with 4GB GPUs when games made to take full advantage of the PS4 and Xbox One started getting ported to the PC, and Nvidia had to do a major course correction with Pascal. If you have less VRAM than what the consoles have available (which is around 12GB for the current gen), then you're not guaranteed to always achieve full console parity with every game. Especially when developers have been as careless and lazy with their PC ports as they have been recently.
|
# ? May 9, 2023 22:13 |
|
Dr. Video Games 0031 posted:This is not all that sudden, is it? I feel like we went through a similar thing with 4GB GPUs when games made to take full advantage of the PS4 and Xbox One started getting ported to the PC, and Nvidia had to do a major course correction with Pascal. If you have less VRAM than what the consoles have available (which is around 12GB for the current gen), then you're not guaranteed to always achieve full console parity with every game. Especially when developers have been as careless and lazy with their PC ports as they have been recently. How long has the ps5 been a thing...since 2019? That is long enough for a game to be developed entirely targeting a console with 16gb of vram (12gb in practice as you note).
|
# ? May 9, 2023 22:42 |
|
Yudo posted:How long has the ps5 been a thing...since 2019? That is long enough for a game to be developed entirely targeting a console with 16gb of vram (12gb in practice as you note). Yet games that actually took advantage of the new console generation didn't really start coming out in earnest until the past 6 months or so.
|
# ? May 9, 2023 22:44 |
|
Yudo posted:How long has the ps5 been a thing...since 2019? That is long enough for a game to be developed entirely targeting a console with 16gb of vram (12gb in practice as you note). Holy poo poo has it really been over 3 years already
|
# ? May 9, 2023 23:14 |
|
Yudo posted:How long has the ps5 been a thing...since 2019? That is long enough for a game to be developed entirely targeting a console with 16gb of vram (12gb in practice as you note). So the PS5 has 16GB of RAM....total. Not just VRAM. It's shared (plus an extra 512MB for system stuff, so the real playground IS probably closer to 16 and not 12, but I would imagine you'd never budget for more than around 12GB for graphics at any given time). So importantly its unlikely that games actually need, from a technical perspective, that much VRAM. However, it seems like the approach that is being used when porting is not being very intelligent about what is being stored in VRAM and when. Importantly, one thing that is killing me, is how these ports in particular DON'T look all that much better despite having higher requirements. I think if better engineering approaches were taken we could see that bar drop dramatically, but I don't know if its the game developers or Sony's tools or what.
|
# ? May 9, 2023 23:14 |
|
Lockback posted:So the PS5 has 16GB of RAM....total. Not just VRAM. It's shared (plus an extra 512MB for system stuff, so the real playground IS probably closer to 16 and not 12, but I would imagine you'd never budget for more than around 12GB for graphics at any given time). So importantly its unlikely that games actually need, from a technical perspective, that much VRAM. However, it seems like the approach that is being used when porting is not being very intelligent about what is being stored in VRAM and when. That doesn't mean everything has to be a TLOU-style bin fire but it's probably safe to assume no PC port will ever be as well optimised as a console game, so that will always balloon the requirements a bit Zephro fucked around with this message at 23:26 on May 9, 2023 |
# ? May 9, 2023 23:23 |
|
ijyt posted:Holy poo poo has it really been over 3 years already Preorders started in September 2019, IIRC. Deliveries began sometime in 2020. Don't remember specifically when, only that for me it was shortly after lockdown began (so April or May?)
|
# ? May 9, 2023 23:32 |
|
It was November 2020. Both PS5 and XSX only hit the shelves in November 2020, and there was no pre-order period. This generation has also had an unusually long "cross-gen" period because a lot of games that were supposed to come out earlier got delayed into 2022/2023. Aside from first-party stuff, we're only now seeing the first wave of games designed with the new consoles in mind, I think. If we look at the recent games that have shown issues with 8GB of VRAM, it's mostly stuff that primarily targeted the new consoles, like Forspoken, HogLeg, TLOU Part 1, RE4R (at least, when ray tracing is enabled), and the Spider-Man remasters (at high resolutions with RT enabled). Some of these are scalable to be more reasonable at lower resolutions/without RT, and others are just awful no matter what because they're poo poo ports.
Dr. Video Games 0031 fucked around with this message at 23:44 on May 9, 2023 |
# ? May 9, 2023 23:37 |
|
Dr. Video Games 0031 posted:It was November 2020. Both PS5 and XSX only hit the shelves in November 2020. There was no pre-order period. Holy crap you're right. I was so sure about that memory, but I looked up my order history and I placed my order on September 17 2020 and it was delivered November 14. I blame COVID time dilation.
|
# ? May 9, 2023 23:46 |
|
mdxi posted:Holy crap you're right. I was so sure about that memory, but I looked up my order history and I placed my order on September 17 2020 and it was delivered November 14. Off-topic, but where did you place your order? I'm certain that there was no pre-order period in most parts of the world because it was such a mad scramble to get one when it finally released. Were there actually retailers out there who had the common sense to use a preorder/queue system? edit: Or I'm guessing that shops like Gamestop must have had them live for a minute before they sold out and you got lucky?
|
# ? May 9, 2023 23:48 |
|
Dr. Video Games 0031 posted:Off-topic, but where did you place your order? I'm certain that there was no pre-order period in most parts of the world because it was such a mad scramble to get one when it finally released. Were there actually retailers out there who had the common sense to use a preorder/queue system? I pre ordered mine from Amazon when they had them up for like an hour before they sold out. Ordered it on sept 16th Did the same with the Xbox from Costco.
|
# ? May 10, 2023 02:02 |
|
mdxi posted:Holy crap you're right. I was so sure about that memory, but I looked up my order history and I placed my order on September 17 2020 and it was delivered November 14. My sense of time is completely shot. I thought it hit the shelves late 2019.
|
# ? May 10, 2023 04:10 |
|
Zephro posted:Well, it doesn't until it does. If you start to bump up against the limits of the VRAM then it affects performance pretty dramatically. Forecasting the future is a mug's game but given that most new AAA games will be targeting the Series X and the PS5 and dropping support for older consoles, I'd be a little uncomfortable with 12GB if I was planning to hold onto the card for more than a couple of years. It might be fine, but even if it is it'll be a bit of tight squeeze. Yeah I mostly play VR poo poo nowadays and things like VRChat especially are VRAM hogs, because nobody believes in optimized texture work. I have a 3080 12GB and I can watch my frame rate literally slash in half as soon as I fill my VRAM and textures wind up cached in system RAM. Even worse is if you somehow wind up overflowing your VRAM significantly, which seems to pretty much be impossible outside of poo poo like VRChat, you can eat up all your system RAM with textures that didn't fit in VRAM, and you can crash your PC.
|
# ? May 10, 2023 04:26 |
|
orange juche posted:Yeah I mostly play VR poo poo nowadays and things like VRChat especially are VRAM hogs, because nobody believes in optimized texture work. I have a 3080 12GB and I can watch my frame rate literally slash in half as soon as I fill my VRAM and textures wind up cached in system RAM. Even worse is if you somehow wind up overflowing your VRAM significantly, which seems to pretty much be impossible outside of poo poo like VRChat, you can eat up all your system RAM with textures that didn't fit in VRAM, and you can crash your PC. one time i heard someone bitching about 16gb not being enough because of modded skyrim at some point that's on you buddy, you're shoving bloated texture packs into age old games and pretending that means "fidelity"
|
# ? May 10, 2023 05:02 |
|
Shipon posted:one time i heard someone bitching about 16gb not being enough because of modded skyrim I've run across poo poo where people are using 8k textures for skinning models, individual 8k textures for each material instead of realizing that 4k is quite probably fine, and they could also atlas the textures to save space in vram. Some people build avatars which consume an entire GB of VRAM just for the textures on one model. It's what happens when you give people who know gently caress all about best practices for 3d modelling the ability to import whatever the gently caress they want into your game.
|
# ? May 10, 2023 05:19 |
|
Dr. Video Games 0031 posted:Off-topic, but where did you place your order? I'm certain that there was no pre-order period in most parts of the world because it was such a mad scramble to get one when it finally released. Were there actually retailers out there who had the common sense to use a preorder/queue system? Best Buy. They and, like, Walmart were doing the functional equivalent of pre-sales on what I suppose were their initial allocations from Sony. I happened to be up late one night, saw someone post about a new batch showing up and thought "can't hurt to try", and actually landed one. Didn't say anything to anyone about it for months because I felt so guilty as the magnitude of the shortage became apparent.
|
# ? May 10, 2023 05:33 |
|
The 6900xt is $580 on Amazon. I think that qualifies as a firesale. I have wonder how much rdna 2 inventory is left.
|
# ? May 10, 2023 06:35 |
|
https://videocardz.com/newz/nvidia-geforce-rtx-4060-ti-8gb-to-launch-on-may-24th-a-day-before-radeon-rx-7600 4060 Ti 8GB release: May 24th 7600 8GB release: May 25th 4060 8GB release: 1H July 4060 Ti 16GB release: 2H July My condolences to reviewers who have to review two GPUs at the same time.
|
# ? May 10, 2023 12:42 |
|
Dr. Video Games 0031 posted:https://videocardz.com/newz/nvidia-geforce-rtx-4060-ti-8gb-to-launch-on-may-24th-a-day-before-radeon-rx-7600 8gb of vram, gross
|
# ? May 10, 2023 14:14 |
|
ah, there's my route in 2025. a nice 5060 Ti 14GB for $650
|
# ? May 10, 2023 14:37 |
|
The "4060" is really going to be an old-school x50 product, the poo poo you buy for video decoding/multiple outputs/etc or maybe for a kid to play a few 5 year old games. In that context 8gb of vram is perfectly fine, the problem is it's going to sell for $400+ instead of the relatively trivial cost those products used to have where sometimes you could justify it for a specific use case. The 4060 Ti we'll see but that's probably going to be a product that never should have existed and is a bad buy to sell the 4070, so I'm not sure the VRAM will even matter but it depends on price. AMD's positioning is actually super weird and seems like it will make no sense, with the 7600 probably being a GPU that really should have more VRAM. It's an odd decision because AMD is still on G6 and it's not that expensive, must be driven by a desire to avoid higher memory on a lower end GPU at a moment when GPU VRAM appears relatively more important. This generation continues to really, really suck outside of the 4090. None of these cards are going to be fun to own 3 years from now, and they all cost way too much to buy for such a short term.
|
# ? May 10, 2023 14:41 |
|
Cavauro posted:ah, there's my route in 2025. a nice 5060 Ti 14GB for $650 Bold of you to assume it’ll be that affordable.
|
# ? May 10, 2023 14:42 |
|
Cavauro posted:ah, there's my route in 2025. a nice 5060 Ti 14GB for $950
|
# ? May 10, 2023 14:44 |
|
I'm thinking of selling my 6800XT and getting a 4070. I'd like to be able to do raytracing (even if it needs upscaling/frame gen), and mostly I'd like to be able to mess around with AI stuff, which is just a nightmare with AMD cards. In either case the lower VRAM kind of sucks but drat the 4080 is expensive. I'm also thinking I'm being an idiot because the 6800XT is fine, RT is a gimmick at anything but absolute high end, and I should save up for a 5090 or whatever and rent GPUs online to do ML... I really wish NVidia had been a little more competitive this round so that I'd have a better excuse to dump my AMD card, instead of being tempted to do a ~$100 "side-grade" to play with some tech.
|
# ? May 10, 2023 15:57 |
|
I have heard from several people itt that it saves both time and money to rent GPU time for ML, unless you are doing it literally all the time, and then it is worth the multiple thousands to get one of the Quadro cards.
|
# ? May 10, 2023 16:19 |
|
wargames posted:8gb of vram, gross This really needs to stop, ffs
|
# ? May 10, 2023 16:24 |
|
a 5060 will be 12GB with a 96-bit memory bus or 12GB with a 64b bus using the 3GB GDDR7 chips
|
# ? May 10, 2023 16:26 |
|
we love the state of AAA gaming don't we folks https://twitter.com/opinali/status/1656318296407392257?t
|
# ? May 10, 2023 16:53 |
|
steam shows two file sizes for patches—the download size and the size of the files the patch modifies. he's only showing the latter
|
# ? May 10, 2023 17:01 |
|
Yeah the patch is only 4.3GB apparently.
|
# ? May 10, 2023 17:09 |
|
oh, that's no fun to make fun of in real news, asus is launching a new 4090 variant with the smaller cooler from their 3090ti which might be good for space constrained cases https://twitter.com/VideoCardz/status/1656304449856458753
|
# ? May 10, 2023 17:14 |
|
|
# ? May 30, 2024 14:35 |
|
MagusDraco posted:Yeah the patch is only 4.3GB apparently. I only downloaded 1.75GB on the EA app lol
|
# ? May 10, 2023 17:28 |