Not me I have a 1080. If vega has the performance I would sell mine and get it. There are lot of people out there like me that are stuck with NV but have a freesync monitor just because AMD doesn't have higher performance product. I would jump ships if it has the price/performance. For me it doesn't have to beat the pants off NV, just be priced right and have similar or better performance and I will make the jump.
![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
Radeon RX Vega Discussion Thread
- Thread starter DPI
- Start date
Not me I have a 1080. If vega has the performance I would sell mine and get it. There are lot of people out there like me that are stuck with NV but have a freesync monitor just because AMD doesn't have higher performance product. I would jump ships if it has the price/performance. For me it doesn't have to beat the pants off NV, just be priced right and have similar or better performance and I will make the jump.
Nope way off:
- 36 compute units Polaris is 232mm2
- 64 compute unit Vega? You would be 1.8 X area of Polaris -> 412mm2 if that was the case
- Vega is estimated to be 530mm^2 to 560mm^2 by images taken
cybereality
[H]F Junkie
- Joined
- Mar 22, 2008
- Messages
- 8,789
I think Vega will blow it out the water. AMD is too quiet, sort of like a poker player with pocket aces not wanting to tip their hand.
Reminds me of the R300 days just before the 9700pro was launched. The only clue that it was special was that in Doom 3 it blew the 5800fx out of the water before Doom 3 was launched, Carmack was letting folks know why he was using the hidden ATi card vice Nvidia for shows. Fun time.
I always keep my hopes low with AMD graphics devision but Nvidia releasing 1080ti and then releasing new Titan Xp that is fully enabled, looks like they saw a chance to milk people as much as they can before Vega comes out. They might know more about it then we do lol. If they expect Vega to compete hard it only makes sense to release the 1080ti asap. I know it because when AMD release fury x, Nvidia waited what like a week before to announce the 980 ti?
They could have easily rolled on with 1080 at same price and Titan x but for them to release 1080ti way before vega and update the titan X to Titan Xp is quiet telling.
SighTurtle
[H]ard|Gawd
- Joined
- Jul 29, 2016
- Messages
- 1,410
Or AMD has nothing and Nvidia is just operating on their own timetable? But it is sorta strange for Nvidia to release 1080ti and drop a unexpected Titan Xp so close by.
cybereality
[H]F Junkie
- Joined
- Mar 22, 2008
- Messages
- 8,789
The Titan Xp coming so soon after 1080 Ti makes me think it's a preemptive strike. Meaning Vega could land as something competitive to 1080 Ti.
Well since we know Polaris has been revised and sits 580 and down -> which will basically make the 1060 trail and the 3gb version pointless. What sits above it? The 590 and 590 XT or 595 whatever AMD decides to call the 1070/1080 level cards. That I would expect to be Vega 11 since it is clear Vega 11 is not the replacement for Polaris. It is kinda funny in that we really do not know for sure which Vega was being demoed in the machine - The chip shown I think was told to be Vega 10.
So with Nvidia poorer performance with RyZen (then it should) especially in DX 12 I can see a perfect setting. Two RyZen machines, one with a 1080Ti or Titan Xp, DX 12 games and Vega and Nvidia Cards OC with Vega doing extremely well. Could be a perfect storm. Then add in a kick ass price and then you have mobs (internet chatter) talking about how Nvidia has ripped people off for the last year
.You can't extrapolate units like that, the front end changes of Polaris then add in DP
. Polaris doesn't have DP units. Need to account for that too. So pretty much yeah its going to end up ~475nm.No Vega from what they showed was 470-500nm. Its around Fiji size, any more there is just no space with the bracket. The bracket looked the same size as Fiji's too. So right around that size.
Not strange, the 1080ti made the old Titan P obsolete at its price. nV is only competing with themselves.
june 2nd for 980ti
and june 24th for fury x.
Nvidia announced it just before fury x was announced. I think specs had already leaked for fury x. So yea few weeks before. But It was pretty close to fury x release.
that is a stupid way to do business. No company just sits there and competes with themselves. They had absolutely no incentive to drop price on gtx 1080 and release 1080ti at $699 and killing titan x and releasing titan Xp. When there is absolutely no competition. All these moves are preemptive.
"1080 is selling like crazy" lets 1080ti and drop the price by a 100 on 1080 lose some profits and make titan x obsolete.
1080ti and Titan XP are so far before Vega's launch, there doesn't seem anything coicidental about it, the suprising thing for me is the price drop of the 1080 and lower price of the 1080ti, I didn't expect nV to do that, I really thought they would milk the 1080ti at 800 bucks or 850.
Yes they do, when there are no other competitors, a company has to release products to keep on selling, look at Intel, they didn't do it only because they are making money with OEM's to businesses. nV can't do that. If nV tried to stagnate the market like Intel with mediocre performance increases, they wouldn't sell any products for a generation or two. Intel has a revolving door that nV doesn't have.
The only preemptive thing they did was to screw with AMD's price on Vega, nV knows because of HBM2, AMD can't price it low, so by putting their 3rd best card at 500 bucks, is going to hurt the hell out of AMD. At 700 AMD should be making 45% margins at least, that isn't nV's plan. The 980ti came out and crushed Fury X's hopes on pricing it above 650. And yet they stll priced it at 650 even though it wasn't wroth it at that price.
ofcourse but for them to drop 1080 by a hundred and price 1080ti $699 was largely surprising. They must be expecting some competition.
When was the last time you cared about base clocks on rx series? So we are disregarding these as boost clocks? My rx 470 in second rig does 1325 24/7, does not move down even for a single second under load. Not one! This is not refernece design we are talking about. Are you really going to argue this now. Whey didn't gigabye give us 1439 boost clock before if it was just marketing and it was a lie? lol
Sometimes as much as you know, when you ignore something you don't say much.
No expect that its very close to the max, why do you think the gaming boost clocks are so close to the OC boost clocks? Seems odd doesn't it?
Aurus brand don't know if you have ever tried them, are very close to the max overclocks. They also do this with nV cards too. The only one I know that has head room to do more 100 mhz is the 1080ti..... All others are pretty much maxed out maybe you get 25mhz more if you manually over clock them.
Yes you can:You can't extrapolate units like that, the front end changes of Polaris then add in DP
No Vega from what they showed was 470-500nm. Its around Fiji size, any more there is just no space with the bracket. The bracket looked the same size as Fiji's too. So right around that size.
Hawaii 28nm process 438mm^ 44CU's
Fiji 28nm process 64CU's, 595mm^2
64cu/44cu X 438mm^ = 637mm^2 Fiji if all was relative. Except for 28nm a 637mm^2 chip is really not doable hence it was cut down with almost the same front end and back end getting below 600mm^2 or 595mm^2.
If Vega is similar to Polaris on the same process and it has 4096 Streaming units it should be 437mm^2 not 530mm^2 ()
- Can mean a number of things as mentioned before: Big design change, made for speed, more CU's, different configuration of each CU - who knows besides AMD?
Fiji didn't have dp units either man.
DP units take up a ton of space, look at GP 102 to GP 100, and then you have to factor in the bus size which is a 15% difference between the different memories. We are looking at around 20% increased space just for those DP units.
DP units take up a ton of space, look at GP 102 to GP 100, and then you have to factor in the bus size which is a 15% difference between the different memories. We are looking at around 20% increased space just for those DP units.
All cards with different modes are like that as well.
clocks for msi gtx 1080 gaming. 12mhz higher OC. 13Mhz higher in gaming mode, Stock 1607. When it could be much higher. These OC limits are nothing but whatever they can keep up with default power limits I believe out of the box. But nvidia cards hardly stay at stock they usually stay around boost clocks as well.
- 8GB 256-Bit GDDR5X
- Core Clock 1632 MHz (OC Mode)
- 1620 MHz (Gaming Mode)
- 1607 MHz (Silent Mode)
- Boost Clock 1771 MHz (OC Mode)
- 1759 MHz (Gaming Mode)
- 1733 MHz (Silent Mode)
- 1 x DL-DVI-D 1 x HDMI 2.0 3 x DisplayPort 1.4
- 2560 CUDA Cores
- PCI Express 3.0 x16
I'm talking about Aorus, and how they max out clocks in their OC mode. As I stated, all their cards are close to max clocks (based on what power connectors they use). They are ment to go up against EVA's top end ICX ftw2 or classified cards.
Yea that is a 173mhz clock increase vs stock rx 480. its a 13% clock increase out of the box. If it can max out around 1500 thats 18% bump from rx 480.
We will find out soon enough I guess.
Exactly to reduce the size less than 600mm^2. If you took Hawaii up to 4096 shaders with 2:1 ratio for DP it would be 637mm^2
Vega is too big to compare to Polaris regardless if you add in DP. DP is not going to add 100mm^2
, I think you are gasping at straws man .In short the Rx 570 will be as fast if not faster than the 1060
. One problem with Pascal is their boost clock or clock they run at is not to far away from their max clock. There is not a whole bunch of mhz to push from them. Anything above 2100mhz is more luck than not.Exactly to reduce the size less than 600mm^2. If you took Hawaii up with 2:1 ratio for DP it would be 637mm^2
Vega is too big to compare to Polaris regardless if you add in DP. DP is not going to add 100mm^2
It added 125mm2 to GP 102. to GP100
So if you have 412mm2 (by your calculations from Polaris 10 to Vega based on CU count ) + 125mm2 that is damn close to 530nm2 lol. But the die probably isn't that big, because when you have a die shot like the one we saw its very hard to see the silicon edge vs the packaging cause of the lighting.
And this is considering the ram bus differences.
Any case should stated this earlier we shouldn't even consider there is a Vega bigger than 4096 ALU's anyways at this point, and we know what the one guy stated on Vega's ALU counts too, the guy that worked on both Vega and Polaris architectures. We know the TFLOPS of the instinct cards, the only way its more than that is if its got less then 1500 mhz boost clocks.
Last edited:
True, only thing I am wondering now is AMD did say Vega NCU are designed for higher clocks, and we have instinct card rated at 12.5 TFLOPS translating to roughly 1500mhz. I am wondering if in top end vega card we are looking at 1500mhz plus on the core.
Not True:
For one Vega from AMD own slide has 1/16 DFFP so no extra precious DP units. If DP units really took up a lot of space Fiji would be much smaller than 596mm^2 with both frontend and backend not significantly upgraded and less DP units. With your logic Fiji should have been less than 500mm^2
#2 - DP units add a significant amount of heat and is not needed for a gaming card, thus get rid of them or don't use them you can rev up your GPU like what Nvidia did
#3 - The package size to support a 4096 HBM2 memory configuration limits how small you can go to hook stuff up. You're not going to have a 4096 bus and a 200mm^2 gpu at this time. GP 100 size most likely deals with supporting HBM2 with a 4096 bit bus. GP 102 only had a 384 bit DDR5(x) bus to hook up to requiring less pins so to speak and could be smaller.
#4 - At all of AMD presentations DP was left out because Vega is going to hit Nvidia in the deep learning space - fastest growing segment (start throwing 4 Vega 10's in a cube with multiple cubes Nvidia has nothing to answer with).
The die size is bigger then it should be by a significant amount. It is now how well AMD can execute and if Vega is accurate in it's name. As far as we know it could lay toast to Nvidia Pascal line or be an utter dude.
Last edited:
Some notes:
Vega geometry rate is up 2.6x over Fiji. Fiji 4 geometry pipeline was designed for 4 polygons per clock - Vega is 11 polygons per clock. That is a huge increase. If the clock speed is indeed 1526mhz or around for the rated TFlops - geometry restrictions will be a thing of the past. AMD did work the front end and it looks like a lot.
The backend or pixel rendering end is now connected to the L2 cache vice the memory controller, this should really speed up deferred rendering and free up the memory accesses for other operations.
CU is still 64 shaders each. AMD indicated it is made for speed - we will see.
Vega geometry rate is up 2.6x over Fiji. Fiji 4 geometry pipeline was designed for 4 polygons per clock - Vega is 11 polygons per clock. That is a huge increase. If the clock speed is indeed 1526mhz or around for the rated TFlops - geometry restrictions will be a thing of the past. AMD did work the front end and it looks like a lot.
The backend or pixel rendering end is now connected to the L2 cache vice the memory controller, this should really speed up deferred rendering and free up the memory accesses for other operations.
CU is still 64 shaders each. AMD indicated it is made for speed - we will see.
I think the main reason for this is because they probably don't have a competitive part. AMD have been behind nVidia for quite a while and if Vega was genuinely faster than anything nVidia has and was right around the corner, AMD would be leaking like a sieve. They are staying tight lipped because a) Vega is the same or slightly worse than Pascal and b) If they leak that the card is worlds apart they risk alienating their die hard fans when the part lands and is not as great as they say.
Lessons learned I'd say because the scenario above has played out before (Fury). That part was delayed so long that I finally jumped ship after a decade on AMD and when the card finally launched I was glad I gave up waiting.
Maybe, maybe not - all conjecture. I say it is most likely the opposite, at least I hope so
.Need to consider that with Pascal they had to do this to make the Titan a true full GPU102 release, with Maxwell the Titan was released at launch as the full uncut GPU.
So with the launch of the 1080ti (without excessive cuts and they did not change core) to justify the Titan Pascal they really did need to make it the full core, which they should had done straight away but not really possible because they launched it well in advance and early into the new 14nm node.
That is the real story, how early into the 14nm cycle they launched the Pascal Titan.
Cheers
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,678
Its not a rumour. The chip have been shown. The reason is most likely packed math and other compute features. Vega 20 will sport FP64 for example and GMI links. And there goes the entire "7nm" shrink with the same 4096SP.
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,678
Fiji removed FP64 capability for example. So in that regard Fiji is much less efficient than Hawaii per transistor.
Vega 10, adds packed path and some frontend changes and you get ~530mm2 on 14nm. Vega 20 adds FP64 and GMI links and 2 extra HBM2 stacks and you are again at ~500mm2 at "7nm".
As long as AMD cant do anything but rehash the same GCN since 2012 this is what you get. GCN5.
1) Yes it should be less than 500 mm2, I was saying that wasn't lol. Fiji might have limited DP units. thought it was cut and it was using using its SP units to do DP work, thats where the 1/16 came in.
2) What are you talking about, HBM bus sizes are tiny man a 4096 bus for HBM is not the same thing as for GDDR. Not even close, HBM buses are like 15% of the size or less, and that is if we compare a 256 bit bus to 2048 bit HBM bus.
3) look at 2
4)
You were saying correct, it has 1/16 DP rate. But this tells us everything. 64 CU's, no more than 4096 units man. 12 TF ~1500 mhz with 4096 ALU's.
Right around 225 watts of power. So what ever they are using their space up for, well, something is eating up a butt load of transistors or maybe they transistor density is looser? That's probably how they got the clocks up more.
And for your speculation, they can't, just won't happen, too much info we know about their Instinct cards man, we know the max of Vega, in Tflops. No one is guessing on Vega, outside of a few things. You are thinking AMD has some magic sauce, they don't, not going to happen. AMD doesn't show or talk about the worst they can do, they show they best, cause that would be what anyone would do in their position.
Last edited: