hate these ads?, log in or register to hide them
Page 170 of 179 FirstFirst ... 70120160167168169170171172173 ... LastLast
Results 3,381 to 3,400 of 3571

Thread: What Video Card Should I Buy?

  1. #3381
    Movember 2011Movember 2012 Nordstern's Avatar
    Join Date
    April 10, 2011
    Posts
    9,881
    AMD: Hey, let's sell a "refined" 8GB RX580 for $279 and call it the RX590.
    Retailers: Cool, we'll carry it. By the way, we dropped the price on 8GB RX580s to $170.
    AMD: shockedpikachu.jpg


    brilliant move, AMD
    Last edited by Nordstern; December 19 2018 at 03:24:00 AM.
    "Holy shit, I ask you to stop being autistic and you debate what autistic is." - spasm
    Quote Originally Posted by Larkonis Trassler View Post
    WTF I hate white people now...

  2. #3382
    Donor erichkknaar's Avatar
    Join Date
    April 9, 2011
    Posts
    12,325
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by elmicker View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Even without NVLink there are still plenty of workloads that benefit from just chucking 8x 2080Ti's in a server. Here's a link to something similar for the 1080Ti.
    Why chuck 8 x 2080Tis in the server when you chuck 8 of these bad boys in the same server.
    Ding ding ding.

    People buying these in bulk aren't constrained by the dollar cost of the cards. $50k for a box? Rounding error.
    Of course they are. Performance per dollar is a thing you know. Sure, some people won't care and be prepared to pay the stupidly high premium to get slightly better performance, but for most people economics are a thing. If you scale out the difference in price becomes more important, not less, because you're quickly talking about millions of dollars. About the only time it makes sense is if your main constraint is rack space, but that is rarely the case. Being constrained by available power / cooling capacity is much more common, and boxes of these pull so much power that you won't be able to completely fill up a rack anyway in the vast majority of data centres.
    How about the difference in time?

  3. #3383
    Specially Pegged Donor Overspark's Avatar
    Join Date
    April 10, 2011
    Location
    NL fuck yeah
    Posts
    3,418
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by elmicker View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Even without NVLink there are still plenty of workloads that benefit from just chucking 8x 2080Ti's in a server. Here's a link to something similar for the 1080Ti.
    Why chuck 8 x 2080Tis in the server when you chuck 8 of these bad boys in the same server.
    Ding ding ding.

    People buying these in bulk aren't constrained by the dollar cost of the cards. $50k for a box? Rounding error.
    Of course they are. Performance per dollar is a thing you know. Sure, some people won't care and be prepared to pay the stupidly high premium to get slightly better performance, but for most people economics are a thing. If you scale out the difference in price becomes more important, not less, because you're quickly talking about millions of dollars. About the only time it makes sense is if your main constraint is rack space, but that is rarely the case. Being constrained by available power / cooling capacity is much more common, and boxes of these pull so much power that you won't be able to completely fill up a rack anyway in the vast majority of data centres.
    How about the difference in time?
    Usually solved by scaling out instead of scaling up (i.e. more relatively cheap servers instead of faster but much more expensive servers), but if you have a workload that can only scale up you're fucked and will have to pay the price. Also known as the CCP method of instancing an MMO.

  4. #3384
    Donor erichkknaar's Avatar
    Join Date
    April 9, 2011
    Posts
    12,325
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by elmicker View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Even without NVLink there are still plenty of workloads that benefit from just chucking 8x 2080Ti's in a server. Here's a link to something similar for the 1080Ti.
    Why chuck 8 x 2080Tis in the server when you chuck 8 of these bad boys in the same server.
    Ding ding ding.

    People buying these in bulk aren't constrained by the dollar cost of the cards. $50k for a box? Rounding error.
    Of course they are. Performance per dollar is a thing you know. Sure, some people won't care and be prepared to pay the stupidly high premium to get slightly better performance, but for most people economics are a thing. If you scale out the difference in price becomes more important, not less, because you're quickly talking about millions of dollars. About the only time it makes sense is if your main constraint is rack space, but that is rarely the case. Being constrained by available power / cooling capacity is much more common, and boxes of these pull so much power that you won't be able to completely fill up a rack anyway in the vast majority of data centres.
    How about the difference in time?
    Usually solved by scaling out instead of scaling up (i.e. more relatively cheap servers instead of faster but much more expensive servers), but if you have a workload that can only scale up you're fucked and will have to pay the price. Also known as the CCP method of instancing an MMO.
    No, but I think its fair to say most people try to achieve both, if you actually have a problem that warrants building and maintaining a physical stack for this, rather than just renting them.
    meh

  5. #3385
    Specially Pegged Donor Overspark's Avatar
    Join Date
    April 10, 2011
    Location
    NL fuck yeah
    Posts
    3,418
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by elmicker View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Even without NVLink there are still plenty of workloads that benefit from just chucking 8x 2080Ti's in a server. Here's a link to something similar for the 1080Ti.
    Why chuck 8 x 2080Tis in the server when you chuck 8 of these bad boys in the same server.
    Ding ding ding.

    People buying these in bulk aren't constrained by the dollar cost of the cards. $50k for a box? Rounding error.
    Of course they are. Performance per dollar is a thing you know. Sure, some people won't care and be prepared to pay the stupidly high premium to get slightly better performance, but for most people economics are a thing. If you scale out the difference in price becomes more important, not less, because you're quickly talking about millions of dollars. About the only time it makes sense is if your main constraint is rack space, but that is rarely the case. Being constrained by available power / cooling capacity is much more common, and boxes of these pull so much power that you won't be able to completely fill up a rack anyway in the vast majority of data centres.
    How about the difference in time?
    Usually solved by scaling out instead of scaling up (i.e. more relatively cheap servers instead of faster but much more expensive servers), but if you have a workload that can only scale up you're fucked and will have to pay the price. Also known as the CCP method of instancing an MMO.
    No, but I think its fair to say most people try to achieve both, if you actually have a problem that warrants building and maintaining a physical stack for this, rather than just renting them.
    Like I've been saying from the beginning, it depends on your workload. Why would you pay for severely overpriced cards if your workload can scale out and thereby achieve a short completion time?

  6. #3386
    Donor erichkknaar's Avatar
    Join Date
    April 9, 2011
    Posts
    12,325
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by elmicker View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Even without NVLink there are still plenty of workloads that benefit from just chucking 8x 2080Ti's in a server. Here's a link to something similar for the 1080Ti.
    Why chuck 8 x 2080Tis in the server when you chuck 8 of these bad boys in the same server.
    Ding ding ding.

    People buying these in bulk aren't constrained by the dollar cost of the cards. $50k for a box? Rounding error.
    Of course they are. Performance per dollar is a thing you know. Sure, some people won't care and be prepared to pay the stupidly high premium to get slightly better performance, but for most people economics are a thing. If you scale out the difference in price becomes more important, not less, because you're quickly talking about millions of dollars. About the only time it makes sense is if your main constraint is rack space, but that is rarely the case. Being constrained by available power / cooling capacity is much more common, and boxes of these pull so much power that you won't be able to completely fill up a rack anyway in the vast majority of data centres.
    How about the difference in time?
    Usually solved by scaling out instead of scaling up (i.e. more relatively cheap servers instead of faster but much more expensive servers), but if you have a workload that can only scale up you're fucked and will have to pay the price. Also known as the CCP method of instancing an MMO.
    No, but I think its fair to say most people try to achieve both, if you actually have a problem that warrants building and maintaining a physical stack for this, rather than just renting them.
    Like I've been saying from the beginning, it depends on your workload. Why would you pay for severely overpriced cards if your workload can scale out and thereby achieve a short completion time?
    Because if you have a time constraint (or you are building a cloud for others to rent, who are probably the primary customers for this), they aren't really overpriced, and if your workload can scale out, it can scale out much further with these.
    meh

  7. #3387
    Specially Pegged Donor Overspark's Avatar
    Join Date
    April 10, 2011
    Location
    NL fuck yeah
    Posts
    3,418
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by elmicker View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Even without NVLink there are still plenty of workloads that benefit from just chucking 8x 2080Ti's in a server. Here's a link to something similar for the 1080Ti.
    Why chuck 8 x 2080Tis in the server when you chuck 8 of these bad boys in the same server.
    Ding ding ding.

    People buying these in bulk aren't constrained by the dollar cost of the cards. $50k for a box? Rounding error.
    Of course they are. Performance per dollar is a thing you know. Sure, some people won't care and be prepared to pay the stupidly high premium to get slightly better performance, but for most people economics are a thing. If you scale out the difference in price becomes more important, not less, because you're quickly talking about millions of dollars. About the only time it makes sense is if your main constraint is rack space, but that is rarely the case. Being constrained by available power / cooling capacity is much more common, and boxes of these pull so much power that you won't be able to completely fill up a rack anyway in the vast majority of data centres.
    How about the difference in time?
    Usually solved by scaling out instead of scaling up (i.e. more relatively cheap servers instead of faster but much more expensive servers), but if you have a workload that can only scale up you're fucked and will have to pay the price. Also known as the CCP method of instancing an MMO.
    No, but I think its fair to say most people try to achieve both, if you actually have a problem that warrants building and maintaining a physical stack for this, rather than just renting them.
    Like I've been saying from the beginning, it depends on your workload. Why would you pay for severely overpriced cards if your workload can scale out and thereby achieve a short completion time?
    Because if you have a time constraint (or you are building a cloud for others to rent, who are probably the primary customers for this), they aren't really overpriced, and if your workload can scale out, it can scale out much further with these.
    No you can't, that's not scaling out. The Titan RTX cards are more than twice as expensive for about 5-20% more processing power than the 2080Ti's. If you buy 2080Ti's you can scale out almost twice the amount you can with Titan RTX cards. The only benefits the Titan RTX cards have is 1) far more memory 2) NVLink 3) a tiny amount of scaling up. If your workloads depends on point 1 or 2 then by all means go ahead, but scaling up is stupid at these prices and scaling out heavily favours the 2080Ti's unless you have very unusual requirements (like rack space instead of power, which sort of prevents scaling out in the first place).

  8. #3388
    Donor erichkknaar's Avatar
    Join Date
    April 9, 2011
    Posts
    12,325
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Quote Originally Posted by elmicker View Post
    Quote Originally Posted by erichkknaar View Post
    Quote Originally Posted by Overspark View Post
    Even without NVLink there are still plenty of workloads that benefit from just chucking 8x 2080Ti's in a server. Here's a link to something similar for the 1080Ti.
    Why chuck 8 x 2080Tis in the server when you chuck 8 of these bad boys in the same server.
    Ding ding ding.

    People buying these in bulk aren't constrained by the dollar cost of the cards. $50k for a box? Rounding error.
    Of course they are. Performance per dollar is a thing you know. Sure, some people won't care and be prepared to pay the stupidly high premium to get slightly better performance, but for most people economics are a thing. If you scale out the difference in price becomes more important, not less, because you're quickly talking about millions of dollars. About the only time it makes sense is if your main constraint is rack space, but that is rarely the case. Being constrained by available power / cooling capacity is much more common, and boxes of these pull so much power that you won't be able to completely fill up a rack anyway in the vast majority of data centres.
    How about the difference in time?
    Usually solved by scaling out instead of scaling up (i.e. more relatively cheap servers instead of faster but much more expensive servers), but if you have a workload that can only scale up you're fucked and will have to pay the price. Also known as the CCP method of instancing an MMO.
    No, but I think its fair to say most people try to achieve both, if you actually have a problem that warrants building and maintaining a physical stack for this, rather than just renting them.
    Like I've been saying from the beginning, it depends on your workload. Why would you pay for severely overpriced cards if your workload can scale out and thereby achieve a short completion time?
    Because if you have a time constraint (or you are building a cloud for others to rent, who are probably the primary customers for this), they aren't really overpriced, and if your workload can scale out, it can scale out much further with these.
    No you can't, that's not scaling out. The Titan RTX cards are more than twice as expensive for about 5-20% more processing power than the 2080Ti's. If you buy 2080Ti's you can scale out almost twice the amount you can with Titan RTX cards. The only benefits the Titan RTX cards have is 1) far more memory 2) NVLink 3) a tiny amount of scaling up. If your workloads depends on point 1 or 2 then by all means go ahead, but scaling up is stupid at these prices and scaling out heavily favours the 2080Ti's unless you have very unusual requirements (like rack space instead of power, which sort of prevents scaling out in the first place).
    At what point do you saturate your interconnect and have to start spending real money?

    Also, you keep talking about constrained power in racks, and my contention there is, why do you actually need physical servers then. If you don’t have a big enough operation to place it in a non power constrained location, why wouldn’t you just rent the G/TPUs for the time you need it?
    Last edited by erichkknaar; December 23 2018 at 03:35:12 PM.
    meh

  9. #3389
    Specially Pegged Donor Overspark's Avatar
    Join Date
    April 10, 2011
    Location
    NL fuck yeah
    Posts
    3,418
    Saturating your interconnect depends on your workload again. Lots of workloads are parallel enough not to run into that issue.

    This whole discussion is irrelevant if you're renting processing time, someone else will have made the decision for you.

    If you're not renting processing time and building the servers yourself you're likely to be constrained by something. In my experience that usually is power, since you're either renting a number of racks with a fixed amount of power per rack, or you own/rent the entire building and are actually constrained by how much power can be delivered to the location. And if it's not power it's usually money, in which case the earlier arguments apply.

    I was watching a Titan RTX review from Gamers Nexus yesterday, according to them the gaming performance is 3-7% faster than a 2080Ti. Their take was that it was only worth it to buy the RTX if you have a workload that doesn't fit into 11 GB memory. Apparently the Titan RTX even has the same PCB as the 2080Ti, just a slightly different GPU and more memory. And the cooling is a massive issue on the Titan RTX, as it tries to draw air from the side, where another card will be if you're filling your box with them. You'd need a 3rd party cooler with a different design for that, but since there aren't going to be 3rd party Titan RTX cards (and there obviously already are 3rd party 2080Ti's) that means it's almost impossible to fill a box with them. The only solution would be to dismantle the cooler and replace it with something else yourself, which you're probably not going to do on a $2500 card.

    So that leaves you with either going back to 3rd party 2080Ti cards or shelling out for their actual datacenter products, such as the DGX-2, which comes with 16x interconnected Tesla V100 32GB cards (and costs $250k IIRC). Or building something yourself with Tesla T4 cards, which are newer than the V100's.

    The Titan RTX seems relegated to workstation loads that need more than 11 GB of memory. That seems to be a very thin niche, but I'm sure NVIDIA know what they're doing in that regard. What it isn't this time round is a better card for gamers with too much money, as it just doesn't make any sense in that scenario. SLI overclocked 3rd party cooled 2080Ti's would do way better and cost about the same. You can SLI two Titan RTXes too of course, but I doubt you'll get much more performance out of them unless you go custom coolers.

  10. #3390
    Donor Pattern's Avatar
    Join Date
    April 9, 2011
    Posts
    6,823
    If it is of any interest, most gpu render farms I've come across use/used gtx1080tis in their nodes.

  11. #3391

    Join Date
    April 13, 2011
    Posts
    6,899
    The workloads driving the price of these cards don't [easily] scale out beyond a single machine.

    Quote Originally Posted by erichkknaar View Post
    why wouldn’t you just rent the G/TPUs for the time you need it?
    Staggeringly expensive.

  12. #3392
    Donor erichkknaar's Avatar
    Join Date
    April 9, 2011
    Posts
    12,325
    Quote Originally Posted by Overspark View Post
    What it isn't this time round is a better card for gamers with too much money, as it just doesn't make any sense in that scenario.
    On this we 100% agree.
    meh

  13. #3393
    Movember 2011Movember 2012 Nordstern's Avatar
    Join Date
    April 10, 2011
    Posts
    9,881
    At CES 2019, Nvidia revealed the newest card in its series of GeForce RTX GPUs, the RTX 2060. Rumors had pointed toward the possibility of the release of a GTX 1160, which might forego the advanced features present in RTX. But instead, Nvidia has developed and announced a $349 version of its RTX graphics.

    Nvidia CEO Jensen Huang spent the first hour of the presentation recapping on the past year of RTX technology, highlighting games like Battlefield V and Anthem. These are two very high-profile games that feature RTX features. Battlefield V supports both ray tracing and DLSS, while Anthem supports only the latter.

    The card looks nearly identical to its older brothers, the RTX 2070, 2080, and 2080 Ti, with a dual-fan setup and clean design. Nvidia’s RTX 2060 features 240 tensor cores with 52 teraflops of deep learning power, 5 giga-rays of real-time ray tracing, and 6GB of GDDR6 RAM.

    ...

    As mentioned, only some games support RTX features. Luckily, you’ll get one of those games with the card. Nvidia is offering a bundle with either Anthem or Battlefield V (your choice) bundled with the card. That offer will also be available for the RTX 2070. RTX 2080 and 2080 Ti buyers get an even better bundle, as they receive both games instead of having to choose.
    AMD's CES keynote is Wednesday, 9AM.
    Last edited by Nordstern; January 7 2019 at 03:48:54 PM.
    "Holy shit, I ask you to stop being autistic and you debate what autistic is." - spasm
    Quote Originally Posted by Larkonis Trassler View Post
    WTF I hate white people now...

  14. #3394
    Specially Pegged Donor Overspark's Avatar
    Join Date
    April 10, 2011
    Location
    NL fuck yeah
    Posts
    3,418
    The much bigger news is that Nvidia is starting to support Freesync, although they're careful to call it VESA adaptive sync. Only 12/400 tested monitors automatically enabled for now, but you can enable it yourself in the driver options.

  15. #3395
    Mortvvs's Avatar
    Join Date
    April 10, 2011
    Posts
    1,286
    Quote Originally Posted by Overspark View Post
    The much bigger news is that Nvidia is starting to support Freesync, although they're careful to call it VESA adaptive sync. Only 12/400 tested monitors automatically enabled for now, but you can enable it yourself in the driver options.
    So glad I didn't upgrade my screens yet, almost went with the Asus PG279Q. Looking at the Acer XV272UP now, supposedly it's using a newer panel than all the other 144hz WQHD IPS screens.

  16. #3396
    Straight Hustlin's Avatar
    Join Date
    April 14, 2011
    Posts
    10,485
    Quote Originally Posted by Mortvvs View Post
    Quote Originally Posted by Overspark View Post
    The much bigger news is that Nvidia is starting to support Freesync, although they're careful to call it VESA adaptive sync. Only 12/400 tested monitors automatically enabled for now, but you can enable it yourself in the driver options.
    So glad I didn't upgrade my screens yet, almost went with the Asus PG279Q. Looking at the Acer XV272UP now, supposedly it's using a newer panel than all the other 144hz WQHD IPS screens.
    This is dope, fingers crossed it works on my Samsung

  17. #3397
    Straight Hustlin's Avatar
    Join Date
    April 14, 2011
    Posts
    10,485
    Quote Originally Posted by Mortvvs View Post
    Quote Originally Posted by Overspark View Post
    The much bigger news is that Nvidia is starting to support Freesync, although they're careful to call it VESA adaptive sync. Only 12/400 tested monitors automatically enabled for now, but you can enable it yourself in the driver options.
    So glad I didn't upgrade my screens yet, almost went with the Asus PG279Q. Looking at the Acer XV272UP now, supposedly it's using a newer panel than all the other 144hz WQHD IPS screens.
    This is dope, fingers crossed it works on my Samsung

  18. #3398
    Lana Torrin's Avatar
    Join Date
    April 13, 2011
    Location
    Bonding around
    Posts
    19,554
    2060 looks like its a 1070 successor FINALLY. Really interested to see if the 1160 actually turns up and if so how much cheaper it is.
    Quote Originally Posted by lubica
    And her name was Limul Azgoden, a lowly peasant girl.

  19. #3399

    Join Date
    November 5, 2011
    Posts
    12,064
    Quote Originally Posted by Lana Torrin View Post
    2060 looks like its a 1070 successor FINALLY. Really interested to see if the 1160 actually turns up and if so how much cheaper it is.
    Kinda hate that they're cheaping out and sticking with 6gb vram tho. It should be able to do max settings at some decent resolutions in modern games np so you'll be hitting up right close to that amount already. :/

  20. #3400

    Join Date
    April 18, 2011
    Posts
    2,935
    Quote Originally Posted by Lana Torrin View Post
    2060 looks like its a 1070 successor FINALLY. Really interested to see if the 1160 actually turns up and if so how much cheaper it is.
    Seems ok allthough a 15-20% performance increase over the 1070 at the same price point is a bit meh. Think of it like this, it's essentially the performance jump from the RX 480 to the RX590 and no one was impressed with that. And prospective buyers better know what version they're buying if they release the GDDR5 versions as well.

    DLSS seems like a killer feature though. Could be really good one day.

Bookmarks

Bookmarks

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •