Rx 470 TDP and Performance Revealed

[DocSwag]

2 hours ago, WereCat said:

Not all the memory chips can handle 8Gbps speed or even 7Gbps so the 4GB version of 480 can be just 8GB version with those 4 other chips not meeting the specifications disabled.

 

Also for 480 it is not required to have memory speed of 8Gbps .. that is just for reference 8GB cards, the 4GB reference version is clocked on 7Gbps only. Cards from AIB partners can range anywhere between 7-8Gbps.

Last year the 980 was a 4gb card with a 256 bit bus and it had 7gbps memory, so the 480 can too. 

 

That just doesn't make sense... It only raises costs for amd for the 4gb card and doesn't offer any advantages. 

[mariushm]

The reviewer cards were 8 GB models with memory running at 8gbps made by Samsung. 

 

The reviewers received a bios that made the card use only half of each chips and reduce the frequency to 7 gbps in order to imitate how a rx 480 4GB card would perform in real world.

 

True RX 480 4GB cards will use 4 GB of memory and manufacturers have only the restriction to use at least 7gbps memory, so they can use memories from ANY memory manufacturer as long as they're capable of 7gbps.

 

The 8gbps memories used on RX480 8GB are made by Samsung and as far as I know they're the only ones capable of those speeds, so nVidia uses the same chips on 1070.

 

RX470 probably has the same conditions imposed by AMD, so chances are small you'll see cards with those Samsung made 8gbps chips but you could see high performance 7gbps chips overclockable to 7.5 gbps or something like that.

 

The RX470 4GB is supposed to be at around $160 msrp , I doubt there's going to be a RX 470 8 GB model because it gets very close to RX480 4 GB and may eat up from rx 480's sales.

[Doobeedoo]

It's a ok card and won't be priced much. It will have buyers that can't reach for higher priced one.

[ClobberXD]

"Will it have the same PCIe power delivery violation issue? Will it taste like Polyjuice potion? We will have to wait and watch... This is me, signing off from linustechtips.com"

[Derangel]

22 hours ago, Energycore said:

If RX 480 = GTX 970

 

then RX 470 = 88% GTX 970 = just over R9 380X.

 

Probably can pass for VR if overclocked

Pretty sure performance doesn't scale at a 1:1 ratio like that. I could be wrong, but we'll see.

[DocSwag]

1 hour ago, Derangel said:

Pretty sure performance doesn't scale at a 1:1 ratio like that. I could be wrong, but we'll see.

Yeah, it's not exactly that but it's relatively close.

[Energycore]

19 hours ago, Derangel said:

Pretty sure performance doesn't scale at a 1:1 ratio like that. I could be wrong, but we'll see.

It's closer when it's the same architecture (this is even the same chip). It's exactly 88% the size, probably with one SP block out of nine disabled in each shader engine. If it's got the same amount of ROPs and TMUs and all the other crap then it should be pretty close in performance to the 480.

 

Probably not equivalent numbers on all those though.

[DocSwag]

1 hour ago, Energycore said:

It's closer when it's the same architecture (this is even the same chip). It's exactly 88% the size, probably with one SP block out of nine disabled in each shader engine. If it's got the same amount of ROPs and TMUs and all the other crap then it should be pretty close in performance to the 480.

 

Probably not equivalent numbers on all those though.

Thing is, I think if may be closer to the 480 than 88 or even 90% (if the clocks are the same).... I feel that it is often case that the 480 is being rop bound. In pure shader power the 480 has more power than the 390 yet it loses to it most of the time, which is probably cuz it has half the ROPs. That's probably also why the 480 beats the 390 in a few games... Those games are probably more shader bound. So perhaps the 470 will be closer to the 480 than just 90%...

 

Thats just my theory; take it for what you want!

[Energycore]

13 minutes ago, DocSwag said:

Thing is, I think if may be closer to the 480 than 88 or even 90% (if the clocks are the same).... I feel that it is often case that the 480 is being rop bound. In pure shader power the 480 has more power than the 390 yet it loses to it most of the time, which is probably cuz it has half the ROPs. That's probably also why the 480 beats the 390 in a few games... Those games are probably more shader bound. So perhaps the 470 will be closer to the 480 than just 90%...

 

Thats just my theory; take it for what you want!

I'm pretty sure AMD fucked up when they decided to add only 32 ROPs to Polaris 10. It's marketed as a VR premium card ffs

 

Like, that mistake is much bigger than allowing a couple games to make the card pull 80W off a PCI-E slot.

 

Now I'm not sure the 32CU number is confirmed. But if it is, we might be looking at a $150 entry point for VR, not $200.

[JamGorby]

7 minutes ago, Energycore said:

I'm pretty sure AMD fucked up when they decided to add only 32 ROPs to Polaris 10. It's marketed as a VR premium card ffs

No, it's not. It's marketed as affordable VR. AMD never said anything about 'premium'. 

[Energycore]

5 minutes ago, JamGorby said:

No, it's not. It's marketed as affordable VR. AMD never said anything about 'premium'. 

If you wanna argue whether or not AMD actually "meant" VR Premium when they "said" VR Premium, feel free to do so without including me in that conversation.

[Delicieuxz]

Maybe the "premium" VR AMD refers to is in low graphical-intensity games, like DOTA2 and LOL:

 

 

That image suggests the RX 470 and 480 GPUs are 1440p-gaming GPUs... with the qualification that they get 60 FPS at 1440p in games like DOTA2, and LOL, which, in my opinion, makes the 470 and 480 1080p cards, which are capable of 1440p in select games.

[Energycore]

2 minutes ago, Delicieuxz said:

Maybe the "premium" VR AMD refers to is in low graphical-intensity games, like DOTA2 and LOL:

 

I think I remember Raja talking about their wording and making a distinction between easy to set up VR (stuff with your phone, like Google Cardboard) and expensive, buy-a-headsset VR, which they called Premium VR.

 

Still, they fucked up bad by having 32 ROPs on a card that's supposed to drive games at 2160x1200

[DocSwag]

15 hours ago, Energycore said:

I'm pretty sure AMD fucked up when they decided to add only 32 ROPs to Polaris 10. It's marketed as a VR premium card ffs

 

Like, that mistake is much bigger than allowing a couple games to make the card pull 80W off a PCI-E slot.

 

Now I'm not sure the 32CU number is confirmed. But if it is, we might be looking at a $150 entry point for VR, not $200.

I know right... I don't know what in the world they were thinking when they put 32 ROPs on there instead of 64! If they put 64 ROPs I could easily see stock 480 trading blows with the stock 390x.... It would probably be beating 390 and 970 in nearly every game.

 

If you watch AdoredTV, in this video (I think it was this one but I could be wrong):

He said that he expected the 480 to have 64 ROPs. That sure didn't happen, and it sure was stupid. Would it really have taken up that much die size to double the ROPs? I'd imagine at most it would make Polaris 10 around 260 mm^2, though I could be wrong.

[Humbug]

I expect 380x performance at GTX 950 price point. Pretty good.

 

But things will really get interesting when nvidia releases their products as well because then their is some price competition. For now there is no price pressure on AMD.

[Trixanity]

Yeah, it's quite clear in some benchmarks against the 390(X) that Polaris is starved. We're talking less shaders, less TMUs and half the ROPs.

The shaders are improved in Polaris but the amount is also lower, so we get 10% less shaders but 15% increased shader efficiency (vs the 390, 18% less shaders vs the 390X). The ROPs should be pretty much equal but halved in number.

 

So that would indicate it simply lacks the resources to clearly beat the older card. Obviously, the improvements makes it better in other aspects, particularly in GameWorks titles where AMD has done a pretty damn good job of combating Nvidia's shenanigans.

 

AMD needs a 3000+ shader/64+ ROP/160+ TMU/8 GB G5X card on GCN4 and have it priced at like $300-350. That would be a really good card for those who want more performance but still a good price.

[captain cactus]

24 minutes ago, Trixanity said:

Yeah, it's quite clear in some benchmarks against the 390(X) that Polaris is starved. We're talking less shaders, less TMUs and half the ROPs.

The shaders are improved in Polaris but the amount is also lower, so we get 10% less shaders but 15% increased shader efficiency (vs the 390, 18% less shaders vs the 390X). The ROPs should be pretty much equal but halved in number.

 

So that would indicate it simply lacks the resources to clearly beat the older card. Obviously, the improvements makes it better in other aspects, particularly in GameWorks titles where AMD has done a pretty damn good job of combating Nvidia's shenanigans.

 

AMD needs a 3000+ shader/64+ ROP/160+ TMU/8 GB G5X card on GCN4 and have it priced at like $300-350. That would be a really good card for those who want more performance but still a good price.

Soooo... Vega 10? But instead of G5X it's HBM2.

[Trixanity]

29 minutes ago, lots of unexplainable lag said:

Soooo... Vega 10? But instead of G5X it's HBM2.

I haven't seen any confirmation if there really is a Vega 10 AND a Vega 11. I mean Vega 10 is confirmed as per this tweet from Raja. Vega 11 though? Haven't seen it except for rumors which means inaccurate and made up as fuck.

 

But if there is only one, it'll be a huge chip with HBM2. 4096 shaders at least. Which means my envisioned $300-350 card doesn't exist.

 

If there are two chips, we should have a smaller one and I don't think AMD will go for HBM2 on two chips due to pricing both in regards to BOM cost AND the MSRP. AMD will probably embrace G5X at one point. It's still a bit expensive and low availability but eventually the performance and power improvements (along with increased availability causing drops in price) will warrant a shift, particularly on the mid-range and above.

And yes, I know the slide for Vega says HBM2 as a feature but it doesn't necessarily mean for all chips in the family although it could be possible but I just don't see it happening on a card at that price. I think even a $400 card would be unlikely to see HBM2. I think $500 is the minimum price for an HBM2 card. But that's pure speculation.

[DocSwag]

3 minutes ago, Trixanity said:

I haven't seen any confirmation if there really is a Vega 10 AND a Vega 11. I mean Vega 10 is confirmed as per this tweet from Raja. Vega 11 though? Haven't seen it except for rumors which means inaccurate and made up as fuck.

 

But if there is only one, it'll be a huge chip with HBM2. 4096 shaders at least. Which means my envisioned $300-350 card doesn't exist.

 

If there are two chips, we should have a smaller one and I don't think AMD will go for HBM2 on two chips due to pricing both in regards to BOM cost AND the MSRP. AMD will probably embrace G5X at one point. It's still a bit expensive and low availability but eventually the performance and power improvements (along with increased availability causing drops in price) will warrant a shift, particularly on the mid-range and above.

And yes, I know the slide for Vega says HBM2 as a feature but it doesn't necessarily mean for all chips in the family although it could be possible but I just don't see it happening on a card at that price. I think even a $400 card would be unlikely to see HBM2. I think $500 is the minimum price for an HBM2 card. But that's pure speculation.

Yup, that's basically the same thing as my guess! I'm guessing that next year vega 10 will be the new fury and fury X at 500<, vega 11 will be the 390/X replacement at 300-450 dollars, and p10 and p11 will stay where they are. Vega 10 would have hbm 2 and vega 11 g5x. Of course vega 11 isn't confirmed but it just doesn't make sense for AMD to leave a whole section of the market uncontested.

[manikyath]

i just wanna add that it caught my eye they use the term "future proof"

 

which to me is a sad, sad statement about how the company tries to advertise.

unless if the card magically gets faster every year until infinity, it is NOT future proof.

 

fun to see this as their very first disclaimer at the bottom:

Quote

[1] Statement of “future-proof” refers to support of current and upcoming technology standards including 14nm FinFET process technology, DirectX®12 and Vulkan™ API support, new display technology, and experiences such as VR. “Future-proof” statement is not meant to serve as a warranty or indicate that users will never have to upgrade their graphics technology again. Support of current and upcoming technology standards described above has the potential to reduce frequency of graphics upgrades for some users.

in other words, "future proof" is now the new marketing term after "eco" and "modern"

[captain cactus]

3 minutes ago, DocSwag said:

Yup, that's basically the same thing as my guess! I'm guessing that next year vega 10 will be the new fury and fury X at 500<, vega 11 will be the 390/X replacement at 300-450 dollars, and p10 and p11 will stay where they are. Vega 10 would have hbm 2 and vega 11 g5x. Of course vega 11 isn't confirmed but it just doesn't make sense for AMD to leave a whole section of the market uncontested.

Except Vega 10 is the small GPU (390X-Fury non-X replacement) and Vega 11 is the big GPU (Fury X replacement).

[DocSwag]

1 minute ago, lots of unexplainable lag said:

Except Vega 10 is the small GPU (390X-Fury non-X replacement) and Vega 11 is the big GPU (Fury X replacement).

Really? God AMD.... Why do you have to switch when you go from Polaris to vega?!?

[ace_cheaply]

11 minutes ago, DocSwag said:

Really? God AMD.... Why do you have to switch when you go from Polaris to vega?!?

Because Polaris 10 came first, and then they decided to make the smaller, slower Polaris 11.   They started on Vega 10 first, and realized they wanted an uber gpu, and so they have Vega 11.  It's the sequence of development

[DocSwag]

3 minutes ago, ace_cheaply said:

Because Polaris 10 came first, and then they decided to make the smaller, slower Polaris 11.   They started on Vega 10 first, and realized they wanted an uber gpu, and so they have Vega 11.  It's the sequence of development

Oh god.... I swear that's gonna confuse the crap out of a lot of people! Why couldn't they just make the big GPU the 10 and the smaller one the 11 universal or vice versa instead of confusing everybody?!? *cries*

[Trixanity]

24 minutes ago, manikyath said:

i just wanna add that it caught my eye they use the term "future proof"

 

which to me is a sad, sad statement about how the company tries to advertise.

unless if the card magically gets faster every year until infinity, it is NOT future proof.

 

fun to see this as their very first disclaimer at the bottom:

in other words, "future proof" is now the new marketing term after "eco" and "modern"

Given AMD's track record, I think it'll get at least 10% faster in the course of its life span. AMD's GPUs have a tendency to start off a bit less than expected but improving over time - a slow start if you well. Nvidia has a tendency to start with a bang and impressing everyone all around but very little improvement during its life span and eventually performing less than expected when driver support trickles to a halt.

 

I think that's a fair presentation of the two. One could argue that AMD should be better prepared when launching their cards and I'll agree on that. The inconsistent launch performance hurts their sales.