Support NeoGAF

[laxu]

Hey, I really enjoyed DOOM 2016. Thing is, I already played through it more than a year ago now on my old 980 Ti which was actually able to sustain a solid 50-60 fps most of the time in 4K. So I'm not exactly excited about how it performs now, more than a year after I played it.

Exactly. It's one of those games that runs pretty well on any half-decent GPU. Plus as a game it does not have huge open world levels or complex physics so it's no wonder it runs well. Still a good game though.

I suppose it's relevant for Vulkan tests because few games use it at the moment but AMD really just loves to bring it up because it happens to run well on their hardware. Which really isn't that great if Nvidia's cards run all games well instead of just ones specifically optimized for their cards or ones that happen to fit with how the GPU architecture handles things.

 

[spyshagg]

Toms didn't find it "terrible" by any metric, one just needs to remember that feature that is reducing fps inevitably... reduces fps .

http://www.tomshardware.com/reviews/amd-radeon-chill-ocat-relive,4846-2.html

Neither did techpowerup:


https://www.techpowerup.com/reviews/AMD/Radeon_Crimson_ReLive_Drivers/5.html


And it seems toms didn't realize that one could set target fps per game (nor did techpowerup). I'm not sure if it simply wasn't there by the time they were reviewing it, or whether they didn't bother.

Its a great piece of tech. Me myself being a natural born camper in BF4/BF1 with little screen action, chill would save me a couple ºC in my room temperature after a long session.

 

[DemonCleaner]

some interesting points he's making:

https://www.youtube.com/watch?v=ejVNhOvwtuI

 

[iswasdoes]

I'd like to than this thread for pushing me over the edge to get a 1080Ti

 

[michaelius]

Exactly. It's one of those games that runs pretty well on any half-decent GPU. Plus as a game it does not have huge open world levels or complex physics so it's no wonder it runs well. Still a good game though.

I suppose it's relevant for Vulkan tests because few games use it at the moment but AMD really just loves to bring it up because it happens to run well on their hardware. Which really isn't that great if Nvidia's cards run all games well instead of just ones specifically optimized for their cards or ones that happen to fit with how the GPU architecture handles things.

Well let's look at positive side - at least they moved from using Ashes to Doom

 

[thelastword]

Toms didn't find it "terrible" by any metric, one just needs to remember that feature that is reducing fps inevitably... reduces fps .

http://www.tomshardware.com/reviews/amd-radeon-chill-ocat-relive,4846-2.html

Neither did techpowerup:


https://www.techpowerup.com/reviews/AMD/Radeon_Crimson_ReLive_Drivers/5.html


And it seems toms didn't realize that one could set target fps per game (nor did techpowerup). I'm not sure if it simply wasn't there by the time they were reviewing it, or whether they didn't bother.

Adored TV explored Radeon Chill in his review of the RX 580. He was impressed...

https://www.youtube.com/watch?v=MQ9ro5pwfXY

 

[dgrdsv]

Well let's look at positive side - at least they moved from using Ashes to Doom

Likely has something to do with the fact that Vega 10 has less effective bandwidth for compute than Fiji and thus Ashes will probably perform somewhat unspectacular on Vega.

 

[Delusibeta]

Overclockers UK update: they're going to sell Vega cards for twice RRP, and then hide a 50% off Vega cards voucher somewhere on their forums. Of course, the odds of said voucher staying hidden is very slim, but the effort is appreciated.

 

[Coulomb_Barrier]

Overclockers UK update: they're going to sell Vega cards for twice RRP, and then hide a 50% off Vega cards voucher somewhere on their forums. Of course, the odds of said voucher staying hidden is very slim, but the effort is appreciated.

Joke?

Just saw Gibbo the staff member has been perma-banned. Yikes. Obviously shouldn't have made those comments about mining performance.

 

[Delusibeta]

Joke?

Yeah, sorry for not including source: it's from this PCGamesN article.

Just saw Gibbo the staff member has been perma-banned. Yikes. Obviously shouldn't have made those comments about mining performance.

You cannot merely ban the admins...

 

[rtcn63]

RX Vega |OT| Mine the Gap

EDIT: Was this an OT title for something already I'm not sure

 

[RootCause]

RX Vega |OT| Mine the Gap

EDIT: Was this an OT title for something already I'm not sure

That's a great tittle. 👍

 

[PFD]

Please don't give it a title based on a rumor

 

[rtcn63]

Please don't give it a title based on a rumor

If the rumor does turn out to be true, there won't be a need for a Vega thread since not a single one will be bought for gaming. ;_;

 

[Unknown Soldier]

AMD aren't dummies, they gave Vega such good mining ability on purpose so it would be guaranteed to sell out regardless of how gaming performs.

 

[Sir Abacus]

AMD aren't dummies, they gave Vega such good mining ability on purpose so it would be guaranteed to sell out regardless of how gaming performs.

No. AMD have always been really good to people on using compute on gaming cards. Nvidia on the other hand have tried to stop their gaming cards from cannibalizing their own Quadro cards. Witness the Titan Xp drivers released a few days ago that triple the performance when Vega was going to eat their lunch on compute.

 

[Unknown Soldier]

No. AMD have always been really good to people on using compute on gaming cards. Nvidia on the other hand have tried to stop their gaming cards from cannibalizing their own Quadro cards. Witness the Titan Xp drivers released a few days ago that triple the performance when Vega was going to eat their lunch on compute.

Remember everybody:

Nvidia performance uplift = OMG INTENTIONAL GIMPING PURE EEEEEVIL DEATH TO EVIL NVIDIA
AMD performance uplift = FineWine(tm), WE LOVE AMD, THEY ARE THE BEST

 

[Rushersauce]

Remember everybody:

Nvidia performance uplift = OMG INTENTIONAL GIMPING PURE EEEEEVIL DEATH TO EVIL NVIDIA
AMD performance uplift = FineWine(tm), WE LOVE AMD, THEY ARE THE BEST

Cut the shit. This was definitely scummy by Nvdia.

 

[Sir Abacus]

Remember everybody:

Nvidia performance uplift = OMG INTENTIONAL GIMPING PURE EEEEEVIL DEATH TO EVIL NVIDIA
AMD performance uplift = FineWine(tm), WE LOVE AMD, THEY ARE THE BEST

No. This was gimped. There is no fine wine argument on the other side. AMD cards have always been strong on compute, no fine wine drivers or said driver arguments needed. Nvidia has always gimped their consumer gear for compute to make sure their gaming cards weren't cheap workstation gear and this was a big one.

On the plus side, I'm a shill for AMD now. Where do I hand in my Gsync monitor and 1080 Ti?

 

[snap]

Remember everybody:

Nvidia performance uplift = OMG INTENTIONAL GIMPING PURE EEEEEVIL DEATH TO EVIL NVIDIA
AMD performance uplift = FineWine(tm), WE LOVE AMD, THEY ARE THE BEST

cuz when it happens to nvidia, it's cuz they held back the performance on purpose and knew what they were doing.

when it happens to amd, it's cuz amd is inept and can barely keep it together. so over time they fixed their broken work.

 

[Unknown Soldier]

No. This was gimped. There is no fine wine argument on the other side. AMD cards have always been strong on compute, no fine wine drivers or said driver arguments needed. Nvidia has always gimped their consumer gear for compute to make sure their gaming cards weren't cheap workstation gear and this was a big one.

On the plus side, I'm a shill for AMD now. Where do I hand in my Gsync monitor and 1080 Ti?

Okay, let's actually stop and think for 2 seconds instead of instantly going to our respective Green and Red fortresses and flinging mortars at each other over the walls.

The argument, such as it is, is that Nvidia intentionally gimped performance of Titan X Pascal and Titan Xp to sell more Quadro. This is why that argument is a fallacy.

(1) For the purposes of professional applications, the big draw of Quadro is driver certification. Nvidia actually works with vendors of important software packages like Maya, 3D Studio MAX, AutoCAD, etc. to certify their drivers work correctly and contain no performance killing or crashing bugs. This not cheap. Most of the Quadro tax is paid for certified drivers.

https://www.reddit.com/r/nvidia/comments/6qrrde/nvidia_improves_titan_xp_performance_for_creative/

(2) The consumer Geforce drivers are not certified and may contain notable bugs which kill performance in untested professional applications. This is more likely than intentional gimping, especially since the same Geforce drivers are used for Titans which have the full GP102 core and the 1080 and 1070 which have a GP104 core which in hardware doesn't support full FP16 compute performance. Nvidia probably never bothered to test compute performance against drivers on Pascal-gen Titans simply because other Pascals literally don't have the same hardware for compute.

http://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-1080-and-1070-founders-edition-review/5

(3) Nvidia isn't suddenly afraid of Vega FE or even consumer Vega for professional applications, because Vega uses Radeon drivers and not certified FirePro drivers. No actual professionals will use Vega FE or RX Vega for professional applications anymore than they use Titan Xp for them. See (1). Also, Titan Xp already outperforms Vega FE in the applications that got a boost from this driver. They didn't need to do anything to already have their $1,200 Titan Xp outperform the $1,500 Vega FE in the majority of tested applications.

http://www.gamersnexus.net/hwreviews/2973-amd-vega-frontier-edition-reviewed-too-soon-to-call/page-3

(4) It's not actually 3x performance, like everything is suddenly 3 times faster. I mean Nvidia is known for exaggerating performance but this is hilarious. If they had just said it's actually like 50-100% performance boost only in certain applications and benchmarks which use that applications maybe it wouldn't look like intentional gimping. But Nvidia has never cared about this because why would they?

https://www.reddit.com/r/nvidia/comments/6qrrde/nvidia_improves_titan_xp_performance_for_creative/

(5) The Quadro P6000 has 24 GB of VRAM. The Titan Xp has 12 GB of VRAM. You might not think this is a big deal but professionals do. Quadro also support 10-bit color while consumer Pascal is limited to 8-bit color. Again: you do not care. Professionals do. It turns out that you also need FirePro for 10-bit color on the AMD side, you are also limited to 8-bit color on the Radeon cards.

(6) Big secret: If you aren't an actual professional and don't need 24 GB of VRAM, the actual real-world performance of a lot of these pro tools isn't really much different between consumer Pascal and Quadros. This was true even before this driver was released.

https://www.pugetsystems.com/labs/a...CC-2017-NVIDIA-Quadro-Pascal-Performance-938/

Conclusion:

Nvidia is not likely to be intentionally gimping Titan in order to sell more Quadro, because if they really wanted to make sure no one could use Titan for compute they would have just sold a GPU core which literally didn't have the hardware for it like 1080 and 1070. It's more likely that the consumer Geforce drivers simply didn't have functional code for some compute hardware that was never even in normal consumer Pascal but now it's been fixed or even added from the Quadro branch. Furthermore as I have mentioned, many of the benefits of Quadro have always been for professional applications and performance has never been that dissimilar when a non-pro is using a consumer card to run some of these pro software programs.

You're welcome to question the timing of this sudden miraculous performance boost, but Nvidia seems pretty determined to do things their way and it's not like AMD is in a competitive position to stop them.

What's interesting is the original Titan X Pascal benefits from this too:
https://www.reddit.com/r/nvidia/comments/6qtf9b/titan_x_pascal_benches_with_the_new_38512_driver/

Which implies the 1080 Ti might also benefit, although no one seems to have benchmarked it yet.

 

[icecold1983]

Okay, let's actually stop and think for 2 seconds instead of instantly going to our respective Green and Red fortresses and flinging mortars at each other over the walls.

The argument, such as it is, is that Nvidia intentionally gimped performance of Titan X Pascal and Titan Xp to sell more Quadro. This is why that argument is a fallacy.

(1) For the purposes of professional applications, the big draw of Quadro is driver certification. Nvidia actually works with vendors of important software packages like Maya, 3D Studio MAX, AutoCAD, etc. to certify their drivers work correctly and contain no performance killing or crashing bugs. This not cheap. Most of the Quadro tax is paid for certified drivers.

https://www.reddit.com/r/nvidia/comments/6qrrde/nvidia_improves_titan_xp_performance_for_creative/

(2) The consumer Geforce drivers are not certified and may contain notable bugs which kill performance in untested professional applications. This is more likely than intentional gimping, especially since the same Geforce drivers are used for Titans which have the full GP102 core and the 1080 and 1070 which have a GP104 core which in hardware doesn't support full FP16 compute performance. Nvidia probably never bothered to test compute performance against drivers on Pascal-gen Titans simply because other Pascals literally don't have the same hardware for compute.

http://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-1080-and-1070-founders-edition-review/5

(3) Nvidia isn't suddenly afraid of Vega FE or even consumer Vega for professional applications, because Vega uses Radeon drivers and not certified FirePro drivers. No actual professionals will use Vega FE or RX Vega for professional applications anymore than they use Titan Xp for them. See (1). Also, Titan Xp already outperforms Vega FE in the applications that got a boost from this driver. They didn't need to do anything to already have their $1,200 Titan Xp outperform the $1,500 Vega FE in the majority of tested applications.

http://www.gamersnexus.net/hwreviews/2973-amd-vega-frontier-edition-reviewed-too-soon-to-call/page-3

(4) It's not actually 3x performance, like everything is suddenly 3 times faster. I mean Nvidia is known for exaggerating performance but this is hilarious. If they had just said it's actually like 50-100% performance boost only in certain applications and benchmarks which use that applications maybe it wouldn't look like intentional gimping. But Nvidia has never cared about this because why would they?

https://www.reddit.com/r/nvidia/comments/6qrrde/nvidia_improves_titan_xp_performance_for_creative/

(5) The Quadro P6000 has 24 GB of VRAM. The Titan Xp has 12 GB of VRAM. You might not think this is a big deal but professionals do. Quadro also support 10-bit color while consumer Pascal is limited to 8-bit color. Again: you do not care. Professionals do. It turns out that you also need FirePro for 10-bit color on the AMD side, you are also limited to 8-bit color on the Radeon cards.

(6) Big secret: If you aren't an actual professional and don't need 24 GB of VRAM, the actual real-world performance of a lot of these pro tools isn't really much different between consumer Pascal and Quadros. This was true even before this driver was released.

https://www.pugetsystems.com/labs/a...CC-2017-NVIDIA-Quadro-Pascal-Performance-938/

Conclusion:

Nvidia is not likely to be intentionally gimping Titan in order to sell more Quadro, because if they really wanted to make sure no one could use Titan for compute they would have just sold a GPU core which literally didn't have the hardware for it like 1080 and 1070. It's more likely that the consumer Geforce drivers simply didn't have functional code for some compute hardware that was never even in normal consumer Pascal but now it's been fixed or even added from the Quadro branch. Furthermore as I have mentioned, many of the benefits of Quadro have always been for professional applications and performance has never been that dissimilar when a non-pro is using a consumer card to run some of these pro software programs.

You're welcome to question the timing of this sudden miraculous performance boost, but Nvidia seems pretty determined to do things their way and it's not like AMD is in a competitive position to stop them.

What's interesting is the original Titan X Pascal benefits from this too:
https://www.reddit.com/r/nvidia/comments/6qtf9b/titan_x_pascal_benches_with_the_new_38512_driver/

Which implies the 1080 Ti might also benefit, although no one seems to have benchmarked it yet.

Do you ever have even the slightest idea what you are takking about? What hardware does a titanxp have that a 1080 does not?

 

[00ich]

Do you ever have even the slightest idea what you are takking about? What hardware does a titanxp have that a 1080 does not?

absolutely, Quadro != Tesla

 

[icecold1983]

absolutely, Quadro != Tesla

Umm what??

 

[dgrdsv]

So far Vega 10's mining capacity seems to be pretty weak, as in about three times lower than that of 1070 per watt for ETH. There's a huge chance that Polaris (470 4GB specifically) will remain to be the most efficient mining GPU on AMD's side even after Vega's launch.

 

[DieH@rd]

Professional mining software developer here.
Ethash (the proof-of-work used in Ethereum) is memory bandwidth bound.
1 Ethash requires 8KB of memory bandwidth, 1 MH requires ~7.8 GB of memory bandwidth (1000000*8KB/1048576 ~= 7.8GB).
Vega56 has ~409.6 GB/s; 409.6/7.8 ~= 52.5 MH/s
Vega64 has ~483.8 GB/s; 483.8/7.8 ~= 63.4 MH/s
Both versions use SKHynix's HBM2 memory, which is a 1.6Gbps part @ 1.2volts (SKHynix's Q2'17 Databook). This is the memory speed in Vega56. The Vega64 runs the same part @ 1.89Gbps meaning it is factory overclocked and unlikely to go faster.
Tldr: Peak Ethash performance unlikely to exceed 65 MH/s on Vega64 and 55 MH/s on stock Vega56.

https://www.reddit.com/r/Amd/comments/6rm0cs/vega64_mining_is_unlikely_to_exceed_65_mhs/

 

[dgrdsv]

https://www.reddit.com/r/Amd/comments/6rm0cs/vega64_mining_is_unlikely_to_exceed_65_mhs/

Vega 10 uses Samsung's HBM2 and not SK Hynix.

It's also somewhat doubtful that ETH miners are solely memory b/w bound as Fiji isn't a very good mining card and it has more bandwidth than even Vega 10.

 

[Xyphie]

I don't get where this "Vega will be amazing at mining" meme making the tech news rounds is coming from. Is there a rational reason to think mining performance is going to be higher than the ~30-35 MH/s FE owners are reporting?

 

[Sir Abacus]

Wow. I vote that all Nvidia vs AMD threads be conducted purely by thelastword on the AMD side and icecold1983 on the Nvidia side.

The mental gymnastics we would see on display would be olympic calibre.

 

[Delusibeta]

I don't get where this "Vega will be amazing at mining" meme making the tech news rounds is coming from. Is there a rational reason to think mining performance is going to be higher than the ~30-35 MH/s FE owners are reporting?

The primary source seems to be Overclockers UK, and they are absolutely in the position to know this sort of thing, being one of the larger PC component webstores in the UK.

 

[snap]

I don't get where this "Vega will be amazing at mining" meme making the tech news rounds is coming from. Is there a rational reason to think mining performance is going to be higher than the ~30-35 MH/s FE owners are reporting?

There was a rumor that a driver in beta increased that to 70-100 MH/s

 

[ZOONAMI]

For the power consumption though you're better off going with 2 1070s to get 60mh/s

Or hell 4 1060s would get you closed to 100 mh/s and would pull close to or less power and are readily available

 

[tuxfool]

For the power consumption though you're better off going with 2 1070s to get 60mh/s

Or hell 4 1060s would get you closed to 100 mh/s and would pull close to or less power and are readily available

In most cases power consumption isn't the most significant contributor to cost, unless you live somewhere with high energy prices. In places with higher than average energy prices, those will also have inflated hardware prices.

 

[ZOONAMI]

In most cases power consumption isn't the most significant contributor to cost, unless you live somewhere with high energy prices. In places with higher than average energy prices, those will also have inflated hardware prices.

Well yeah but if it's less power consumption and the prices of the cards aren't inflated like amds cards are, it's cheaper all around to just go with Nvidia for mining right now

 

[Unknown Soldier]

The primary source seems to be Overclockers UK, and they are absolutely in the position to know this sort of thing, being one of the larger PC component webstores in the UK.

The weird thing is that the guy who posted it, a well-known OCUK employee, was subsequently perma-banned from there shortly afterwards. Something is indeed rotten in the state of Denmark.

 

[icecold1983]

Wow. I vote that all Nvidia vs AMD threads be conducted purely by thelastword on the AMD side and icecold1983 on the Nvidia side.

The mental gymnastics we would see on display would be olympic calibre.

are you maybe mixing me up with someone else?

 

[Sir Abacus]

My bad. Unknown Soldier. Me fail English.

 

[riflen]

The weird thing is that the guy who posted it, a well-known OCUK employee, was subsequently perma-banned from there shortly afterwards. Something is indeed rotten in the state of Denmark.

I'm pretty sure that Gibbo is the purchasing manager at OCUK and that the 'ban' was a joke.

 

[RootCause]

So what will be the expected performance for next years budget cards? And when is the approximate release date?

 

[Sir Abacus]

So what will be the expected performance for next years budget cards? And when is the approximate release date?

What's the definition of "budget"?

 

[Delusibeta]

The weird thing is that the guy who posted it, a well-known OCUK employee, was subsequently perma-banned from there shortly afterwards. Something is indeed rotten in the state of Denmark.

That did happen, for about an hour. He subsequently unbanned himself.

 

[RootCause]

What's the definition of "budget"?

My bad. $200-$300 range.

 

[Unknown Soldier]

That did happen, for about an hour. He subsequently unbanned himself.

Oh, how amusing. He's definitely messing with people, but if he's right then Vega sales will be amazing. Just not to gamers. AMD doesn't really care one way or the other though, they need to drive revenue and profits and if miners are the ones buying the cards then so be it.

 

[Sir Abacus]

My bad. $200-$300 range.

The RX580 was just refreshed a few months ago (April) but we could be seeing a GV106(?) as early as the beginning of next year. Any theoretical GV106, educated guess based on the precedence of xx6 chips being 1/3 of a fully enabled version we're looking at 28 SMs for a total of 1792 CUDA cores. Depending on how hard it clocks it's probably 1070 level of performance with lower power draw. The 1070 overclocks ridiculously well so any theoretical 1160 would have to clock into the stratosphere (like 2.1-2.2GHz) to be outright better than an 1070 but they'll probably be neck and neck all the way.

Now is actually a pretty good time to buy a $200-$300 card if you can get something like an 8GB RX580 for RRP ($229). Special order it from a local PC place or Best Buy, it'll take a week, and you won't get screwed on price.

 

[icecold1983]

The RX580 was just refreshed a few months ago (April) but we could be seeing a GV106(?) as early as the beginning of next year. Any theoretical GV106, educated guess based on the precedence of xx6 chips being 1/3 of a fully enabled version we're looking at 28 SMs for a total of 1792 CUDA cores. Depending on how hard it clocks it's probably 1070 level of performance with lower power draw. The 1070 overclocks ridiculously well so any theoretical 1160 would have to clock into the stratosphere (like 2.1-2.2GHz) to be outright better than an 1070 but they'll probably be neck and neck all the way.

Now is actually a pretty good time to buy a $200-$300 card if you can get something like an 8GB RX580 for RRP ($229). Special order it from a local PC place or Best Buy, it'll take a week, and you won't get screwed on price.

Its unclear how much of the volta architecture of gv100 will carry over to consumers. Its not impossible that nvidia just releases a pascal chip with more SMs since unlike gcn we havent yet hit a point where nvidia gpus dont scale well with more units

 

[Atomicfireman]

Oh, how amusing. He's definitely messing with people, but if he's right then Vega sales will be amazing. Just not to gamers. AMD doesn't really care one way or the other though, they need to drive revenue and profits and if miners are the ones buying the cards then so be it.

AMD would much rather sell to gamers. If the cryptocurrency market falls below profitibilty, miners will sell their cards used by the thousands, destroying the new gpu market. In addition, low market share in the gaming space will discourage developers from optimising for their hardware, further reducing performance on AMD's cards.

AMD will look for a way to get cards in gamer's hands before miners.

 

[Sir Abacus]

Its unclear how much of the volta architecture of gv100 will carry over to consumers. Its not impossible that nvidia just releases a pascal chip with more SMs since unlike gcn we havent yet hit a point where nvidia gpus dont scale well with more units

Highly unlikely. They're doing Volta on a new process (12nm TSMC) so it would be crazy to have to go with bigger, more expensive dies for less expensive products. Especially since a) if you're going to be dealing with yield issues on new products with bigger dies why wouldn't you just do it on the new process (i.e. GV102)? and b) the consumer stack from top to bottom is the same basic chip just with parts lopped off or disabled to keep yields high so why run a different die layout and process from GV102? Especially since this method has been tried and tested going on multiple generations now.

We're probably going to lose stuff like the tensor units and 1/2 FP64 going from GV100 to GV102 but I expect everything in the consumer stack to be some layout of the same rough Volta GPC design.

 

[icecold1983]

Highly unlikely. They're doing Volta on a new process (12nm TSMC) so it would be crazy to have to go with bigger, more expensive dies for less expensive products. Especially since a) if you're going to be dealing with yield issues on new products with bigger dies why wouldn't you just do it on the new process (i.e. GV102)? and b) the consumer stack from top to bottom is the same basic chip just with parts lopped off or disabled to keep yields high so why run a different die layout and process from GV102? Especially since this method has been tried and tested going on multiple generations now.

We're probably going to lose stuff like the tensor units and 1/2 FP64 going from GV100 to GV102 but I expect everything in the consumer stack to be some layout of the same rough Volta GPC design.

I think the new L0 cache is the most likely to carry over. I dont expect the smaller number of SPs per SM to carry over along with the things you mentioned. Fp16x2 is probably somewhat likely

 

[Drakhoran]

If we can believe Wccftech

Spoiler

(and we probably can't)

, there could be one Vega card capable of beating the 1080Ti.

 

[Sir Abacus]

I think the new L0 cache is the most likely to carry over. I dont expect the smaller number of SPs per SM to carry over along with the things you mentioned. Fp16x2 is probably somewhat likely

It changed? I thought it was still 64 per SM like Pascal. The layout just changed from 2x32 FP32 core mini modules to 4x16 FP32 core mini modules inside the SM. Oh and doubling the LD/ST units in each SM.