AMD Polaris Thread: Radeon RX 480, RX 470 & RX 460 launching June 29th

[Yakk]

The importance of gf beeing able to use the samsung process can not be understated.

From all we have seen so far, I agree it cannot be understated. The deal with Samsung looks like it was a HUGE win for Global Foundries.

 

[Despoiler]

Well if gibbo is to be trusted and they have 500 cards for day one. Half the amount of 1080 for all days. Its the engineering miracle of this decade imo. Gf turning out finfet in spades. A company that was 3 years behind tsmc.
I dont really care if perf or eff is plus minus 10%. If this is a sign what is in front of us it will give the biggest boost to compettition and price perf for consumers in a decade. The importance of gf beeing able to use the samsung process can not be understated.

It helps that the Polaris chips are smaller. More good chips per wafer even if the yields are still ramping. I think the smaller chip first is the correct approach. Especially given a new node. It is great if GF finally managed to painlessly deliver.

 

[AstuteCobra]

The October rumour comes from two forum posts from 3DCenter. AMD itself says Vega is 2017.

The bottom right also has "AMD roadmaps are subject to change without notice or obligations to notify of changes."

 

[Headfoot]

Just a consideration,

There are only two things unchanged from Hawaii to Polaris as seen in the slide bellow and those are the Rasterizers and ROPs. So with the same amount of Rasterizers (4x) and ROPs (64x) count on both Hawaii (R9 390X) and Polaris RX 480, the performance in games/synthetics that are Raster/ROP limited will be the same.

I believe this is what we see in 3D Mark FireStrike.

I'd assume the rasterizers and ROPS are changed from Hawaii.... only because they were changed slightly in Tonga/Fiji and GCN4 is likely based off of those and not based off of Hawaii

 

[digitaldurandal]

It helps that the Polaris chips are smaller. More good chips per wafer even if the yields are still ramping. I think the smaller chip first is the correct approach. Especially given a new node. It is great if GF finally managed to painlessly deliver.

It also helps that Samsung had already developed the process.

 

[AtenRa]

I'd assume the rasterizers and ROPS are changed from Hawaii.... only because they were changed slightly in Tonga/Fiji and GCN4 is likely based off of those and not based off of Hawaii

I dont believe Rasters and ROPs changed from Hawaii to Tonga/Fiji, only the amount has changed between those chips.

 

[el etro]

The "Perf/mm2 game" has no meaning. It is not relevant to anything. Perf/Watt is something that affects customers. Perf/mm2 has zero impact on anything.

It has its meaning if you're trying to extract every single penny of margin from the sold product. It don't look that matters much, but you understand the meaning once you analyse the Fury X(that uses HBM/Interposer(that even adds Interposing yield rate) and a AIO cooler) margins versus GTX 980Ti(6GB 384-bit bus, needs less power delivery circuitry due to smaller power requirements) margins. Not only Nvidia sells more 980Tis, but also the greens are cheaper to manufacture, giving better margins. Once the profits reach enough levels to amortise the architecture higher cost, all new card sold goes only for the margins. That's why even with sucessful products(Hawaii is awesome!) AMD graphics division operates at loss, while Nvidia consumer graphics division operates at profit.

 

[sirmo]

It has its meaning if you're trying to extract every single penny of margin from the sold product. It don't look that matters much, but you understand the meaning once you analyse the Fury X(that uses HBM/Interposer(that even adds Interposing yield rate) and a AIO cooler) margins versus GTX 980Ti(6GB 384-bit bus, needs less power delivery circuitry due to smaller power requirements) margins. Not only Nvidia sells more 980Tis, but also the greens are cheaper to manufacture, giving better margins. Once the profits reach enough levels to amortise the architecture higher cost, all new card sold goes only for the margins. That's why even with sucessful products(Hawaii is awesome!) AMD graphics division operates at loss, while Nvidia consumer graphics division operates at profit.

Not buying any of it.. completely different fabs. GlobalFundries owned by same people who own 10% of AMD.. also AMD is forecasted to report earnings in the next two quarters.. so what value do you bring to this thread?

 

[Thala]

If you mean perf/watt, no, I don't think they are limiting Radeon RX 480 performance to improve perf/watt on their fastest desktop part. I think that's the best they got out of their mainstream chip, and they priced it accordingly.

It is irrelevant what you think. Trading off perf/area for perf/w is a very typical architectural decision.
Regarding price, it is beating anything left and right in perf/price so it is not priced "accordingly" but gives exceptional value, in particular compared to the NVidia offerings.

 

[Sweepr]

It is irrelevant what you think.

''Did it occur to you that AMD may have sacrificed perf/mm2 to improve perf/watt?''

It was a response to this question. Try reading carefully next time.

Regarding price, it is beating anything left and right in perf/price so it is not priced "accordingly" but gives exceptional value, in particular compared to the NVidia offerings.

Sure, just like Geforce GTX 1070 beats Fury X left and right in perf/price. See what I did here? Easy to come up with claims like this when there's no direct competitor on the market yet.

 

[Face2Face]

This is what I'm seeing on the USA distribution end. Prices seem a bit off, and no stock has been uploaded as of yet.

Looks like MSRP for the XFX model is $249. Could It be the model with the back plate?

 

[trane]

It has its meaning if you're trying to extract every single penny of margin from the sold product. It don't look that matters much, but you understand the meaning once you analyse the Fury X(that uses HBM/Interposer(that even adds Interposing yield rate) and a AIO cooler) margins versus GTX 980Ti(6GB 384-bit bus, needs less power delivery circuitry due to smaller power requirements) margins. Not only Nvidia sells more 980Tis, but also the greens are cheaper to manufacture, giving better margins. Once the profits reach enough levels to amortise the architecture higher cost, all new card sold goes only for the margins. That's why even with sucessful products(Hawaii is awesome!) AMD graphics division operates at loss, while Nvidia consumer graphics division operates at profit.

AMD had a perf/mm^2 lead from 2007 to 2014, sometimes very significant. (It's true that GK104 had greater perf/mm^2 than Tahiti, because GK104 lacked compute features; but elsewhere AMD still won.) It is only with Maxwell that Nvidia retook the lead they lost 7 years prior. IIRC, at their peak, RV770 was nearly 2x perf/mm^2 of GT200. It's hard to imagine today, but Nvidia actually sold a ~576mm^2 GPU for $200 in 2008. Even at launch day GTX 260-216 was priced $270. Competing against a ~250mm^2 4870.

Yet, Nvidia remained much more profitable throughout.

 

[el etro]

Not buying any of it.. completely different fabs. GlobalFundries owned by same people who own 10% of AMD.. also AMD is forecasted to report earnings in the next two quarters.. so what value do you bring to this thread?

I am a Pro-AMD poster, i have no intention of to bash AMD here. My avatar pick didn't tell it?

 

[el etro]

AMD had a perf/mm^2 lead from 2007 to 2014, sometimes very significant. (It's true that GK104 had greater perf/mm^2 than Tahiti, because GK104 lacked compute features; but elsewhere AMD still won.) It is only with Maxwell that Nvidia retook the lead they lost 7 years prior. IIRC, at their peak, RV770 was nearly 2x perf/mm^2 of GT200. It's hard to imagine today, but Nvidia actually sold a ~576mm^2 GPU for $200 in 2008. Even at launch day GTX 260-216 was priced $270. Competing against a ~250mm^2 4870.

Yet, Nvidia remained much more profitable throughout.

Both ones was operating at a loss at that time. After GF104 Nvidia became profitable, and with the overpriced Kepler cards Nvidia profits skyrocketed. Don't know why AMD did not operate at profit these times, maybe the GPUs are covering the losses on the CPU business.

 

[96Firebird]

Looks like MSRP for the XFX model is $249. Could It be the model with the back plate?

I believe that backplate is part of the reference cooler. Both use the same coolers, XFX just has a higher MSRP because they think they can sell for $10 more.

Whoops, nevermind, looks like the backplate is missing from the reference model.

 

[William Gaatjes]

It helps that the Polaris chips are smaller. More good chips per wafer even if the yields are still ramping. I think the smaller chip first is the correct approach. Especially given a new node. It is great if GF finally managed to painlessly deliver.

Maybe the more knowledgeable can verify this to be true, because it comes from the dark corners of my memory :
ATI (before AMD) always did this in the past. On a new process, come with the small chips first, then optimize the process and go for the big chips. I once had a ATI RADEON 9600 XT. I believe i read in the reviews back then they started at that time with the RV360 first on the new process, and then the bigger chips later. A very good approach.


EDIT:
(Note that Nvidia did the same starting with GM107)

 

[crisium]

My concern with performance per mm2 (ppm) is not necessarily on this product, but with what implications it could fortell about the high end battles since ~600mm2 appears to be the practical limit (610mm2 GP100, 601mm2 GM200, 596mm2 Fiji).

ATi/AMD had a huge advantage in ppm from 2008-2011, but (out of what I guess was a reaction to 2900XT) deliberately allowed Nvidia to make far, far larger dies and often surrendered the performance crown to them. I don’t see Nvidia doing this.

So here’s why it’s a concern if AMD did not improve here as much as Nvidia. GP102 is rumoured to be 50% larger than GP104 in everything – same ratio as GM200-204. If the die size ratio stays similar, that means ~474mm2 for the next tier as a 3840 Cuda core card. This should work out well for the +30% a year trend if it hits in 2017.

Now, if big Vega cannot come close to matching GP102 ppm and it’s well into the 500s or even approaching 600mm2, then that’s as big as they can get until real architectural improvements in Navi or whatever the future arch holds.

Nvidia still has room. Notice the 610mm2 GP100? Re-engineer that to full single precision CUDA cores, and we could be talking 5000 Cuda cores or so.

Now as captain Negative Nancy from before, I have calmed down a bit and am open to the thought that Polaris’ size may not necessarily dicate the size of the two Vega cards. But P10 is 53% the size of Hawaii and contains 82% of the SP’s. Small Vega is rumoured to contain 100% of the SP’s as Fiji, so we’re looking at likely more than 53% of the size (53% is already GP104 size).

Meaning, the GP104 competitor, small Vega, will be larger. And thus you can get into the “what if” game I outlined above where Nvidia will always have the performance crown as of now unless they allow AMD to make a much bigger chip (or GF 14nmFF just easily allows well over 600mm2 and ppm is not a burden).

 

[xthetenth]

SP's aren't the only part of the chip though. At the very least you should attempt to norm your variables by constructing a hypothetical GP 106 using the scaling from GM 204 to GM 206 so you can compare apples to apples when comparing P10 to Hawaii by comparing that GP 106 to GM 204 and get a proper scaling measured for both chips.

 

[William Gaatjes]

You know what would be funny ?
If AMD optimized GCN4 in the same way as Nvidia with their gpu architecture did when transitioning from Kepler to Maxwell. Maxwell is a beast in graphics but was not good at fp64 computations for cuda.
Would be fun if AMD has done the same for the polaris chips, going back to the graphics route. Would be fun if asynch compute would still be fast but less than the previous generations of gcn when only doing compute stuff. So Miners maybe have less benefits from the new cards.

Disclaimer : This is a guess.

 

[Lonyo]

It has its meaning if you're trying to extract every single penny of margin from the sold product. It don't look that matters much, but you understand the meaning once you analyse the Fury X(that uses HBM/Interposer(that even adds Interposing yield rate) and a AIO cooler) margins versus GTX 980Ti(6GB 384-bit bus, needs less power delivery circuitry due to smaller power requirements) margins. Not only Nvidia sells more 980Tis, but also the greens are cheaper to manufacture, giving better margins. Once the profits reach enough levels to amortise the architecture higher cost, all new card sold goes only for the margins. That's why even with sucessful products(Hawaii is awesome!) AMD graphics division operates at loss, while Nvidia consumer graphics division operates at profit.

It has other meaning too. If you are launching a mid-range chip and are worse in terms of performance/mm^2, if limits your absolute performance with a higher end GPU, since there's a maximum reasonable size a GPU can be, and assuming both NV and AMD can max out at <600mm^2, AMD won't match NV's performance with a large GPU.

We've had 550mm^2 GPUs before. GP104 is 314mm^2. If NV released something with 50% more die size, AMD would be approaching the limits of a reasonable die to match performance.

 

[crisium]

SP's aren't the only part of the chip though. At the very least you should attempt to norm your variables by constructing a hypothetical GP 106 using the scaling from GM 204 to GM 206 so you can compare apples to apples when comparing P10 to Hawaii by comparing that GP 106 to GM 204 and get a proper scaling measured for both chips.

Yeah, we have too much incomplete info I'm so I'm just trying to gauge it off last gen. P10 is 256-bit, rumoured 64 ROPs, and Hawaii is 512 bit 64 ROPs. I'm not sure how this compares to a comparison between Fiji HBM and Vega HBM2, and the 4096SP Vega is just based off of a LinkedIn account, so who the hell really knows what it'll be.

And all of my stuff is speculation from rumours and last gen stuff, so who knows. Please no one take it as even a near certainty.

A GP106 based on the same design as GM206 (half of the x04) would be around 179mm2, btw, simply using the same ratios which may not be accurate when you get that small anyway (probably need more room when you get very small, so likely bigger). But who knows if they will follow the same path. 192-bit and even 256-bit rumours suggest otherwise.

 

[krumme]

It has other meaning too. If you are launching a mid-range chip and are worse in terms of performance/mm^2, if limits your absolute performance with a higher end GPU, since there's a maximum reasonable size a GPU can be, and assuming both NV and AMD can max out at <600mm^2, AMD won't match NV's performance with a large GPU.

We've had 550mm^2 GPUs before. GP104 is 314mm^2. If NV released something with 50% more die size, AMD would be approaching the limits of a reasonable die to match performance.

All on dx11 games or dx11 games with some dx12 patched on. I think the change will come because of new api using the compute part far more. Put those ace to good use.
Look at history. The gcn advantage will only continue. Pascal is a bandaid solution.

 

[Pheran]

My personal concern is whether or not the RX 480 will be decent at 1440p, as that will be the determining factor as to whether I need a 480 or a 1070. If it really performs like a 390x, then I'm afraid I'll probably need to get a 1070, due to charts like this (stolen from the 1070 thread, thanks Sweepr).

 

[3DVagabond]

Would probably be 250-260mm² in TSCM's process, that's already ~80% of GP104's die size. If it only matches Radeon R9 390X, it's nowhere near 80% GP104's performance.

Who cares about perf/mm². Especially across two different processes?

 

[Head1985]

My personal concern is whether or not the RX 480 will be decent at 1440p, as that will be the determining factor as to whether I need a 480 or a 1070. If it really performs like a 390x, then I'm afraid I'll probably need to get a 1070, due to charts like this (stolen from the 1070 thread, thanks Sweepr).

https://www.youtube.com/watch?v=lIzE5g2WlBo