AMD GDC2016 Thread

[iiiankiii]

Although I hate multi-GPUs, I wouldn't mind grabbing 2X Polaris if:

~Fury performance @ ~100W @ $349 or lower.

That would be pretty awesome. That should be good enough for 4k 60fps gaming @ medium settings in most games. 4k 60FPS gaming while consuming under ~350 watts would be sweet. I'd buy that.

 

[boozzer]

i dont speak american , so i have no idea when the fall is. Polaris is due mid year and its box is dead bang in the middle of the year on the marketing slide.

remembering meeting back to school requires AMD to ship parts to vendors and laptop manufactures months eailer then when we see the product on the shelf.

damn, you are a nice person.

FYI, his misunderstanding is deliberate :sneaky: I just want to give you a heads up, I hate to see a nice guy wasting his time and energy.

 

[maddie]

Polaris 10 from what little info they release, should perform higher than 390X. 390X can't sustain 60 fps in Hitman at 1440p, even in DX12.

So my guess are Polaris 10 #1 $349 giving ~Fury performance, while Polaris 10 #2 $299, giving ~$390X performance.

So they would raise the bar, $299, 290/970 -> 390X performance. Not by much though.

I don't expect them to be priced lower, it would be too generous and they do need $$ badly.

Polaris 11, no way will it match 290/390 levels. ~120mm2 die is just too small for that kind of performance levels.

ps. Ofc I could be very wrong, and AMD suddenly becomes a charity and price 390X beating performance for $249 or something and 390X class performance for $199.. if they do, I would immediate get 2x Polaris 10 for a CF setup.

Edit: For rough performance estimate, you can calculate it on die size: 232mm2 * 2.4x density = equivalent of ~556mm2 GCN on 28nm. Add uarch improvements, there is potential there for ~Fiji performance. So it should land above 390X easily. For Polaris 11, 120mm2 * 2.4x density = equivalent of ~288mm2 GCN on 28nm. Well below Hawaii, which is ~438mm2. With uarch changes, it can approach Tonga, 380X performance. No more than that.

The problem I see with the above prices is that this statement by Taylor cannot be true.

So we're going to have to make it possible to run good quality VR at a much lower price. And I'm confident with Polaris we're going to have a big impact to help that.

Either we accept his statement or totally disregard it. We both agree that Polaris 11 should not be able to run with the 290/970 class card, unless AMD has miraculously transformed their shader performance.

 

[3DVagabond]

so fall is right now :sneaky:

ps. i'm in Australia it is currently autumn.

And you should see Polaris early winter. Or about the second week that your temps stay consistently below 30°C. :lol:

 

[itsmydamnation]

And you should see Polaris early winter. Or about the second week that your temps stay consistently below 30°C. :lol:

Off topic but where i live it goes from max of about 44C in summer to around -8C min in winter, if you count wind chill even colder. Would have to be up there for greatest temp differentials for an area that isn't a desert.

damn, you are a nice person.

FYI, his misunderstanding is deliberate I just want to give you a heads up, I hate to see a nice guy wasting his time and energy.

i take silence as victory

 

[Silverforce11]

The problem I see with the above prices is that this statement by Taylor cannot be true.

So we're going to have to make it possible to run good quality VR at a much lower price. And I'm confident with Polaris we're going to have a big impact to help that.

Either we accept his statement or totally disregard it. We both agree that Polaris 11 should not be able to run with the 290/970 class card, unless AMD has miraculously transformed their shader performance.

The problem is PR people will exaggerate. Raja is an engineering guy so I put more credibility to what he says.

I just don't see them massively shifting the perf/$ scene when they need to have good margins and profits.

It's more reasonable to expect $299 for 390X class performance, as that would already have shifted the perf/$ and VR entry.

Remember, Roy and such often talk about Async Compute being a huge deal, significant gains and we're talking 10-20% perf bonus only here.

 

[MrTeal]

Off topic but where i live it goes from max of about 44C in summer to around -8C min in winter, if you count wind chill even colder. Would have to be up there for greatest temp differentials for an area that isn't a desert.



i take silence as victory

That doesn't seem that extreme. Where I am the average daytime high in July is 25.3C, and the average overnight low in January is -20.7C. We're colder than anywhere in Australia, but the climate here isn't anything too unusual.

 

[3DVagabond]

That doesn't seem that extreme. Where I am the average daytime high in July is 25.3C, and the average overnight low in January is -20.7C. We're colder than anywhere in Australia, but the climate here isn't anything too unusual.

I've lived in upstate NY and in E Tx. So I've experienced similar to both of you. Personally I found the cold worse than the heat. Although living in NZ I can tell you that the southern hemisphere sun is vicious. The stories you hear about the ozone are true.

 

[maddie]

The problem is PR people will exaggerate. Raja is an engineering guy so I put more credibility to what he says.

I just don't see them massively shifting the perf/$ scene when they need to have good margins and profits.

It's more reasonable to expect $299 for 390X class performance, as that would already have shifted the perf/$ and VR entry.

Remember, Roy and such often talk about Async Compute being a huge deal, significant gains and we're talking 10-20% perf bonus only here.

I guess we will see in a few months, but I expect extremely good prices vs present stuff.

Raja in Shrout interview. 7:15 mark

You and your readers are going to be pleased and you are going to be surprised at the positioning of those cards.

I just realized that we are assuming that Samsung fabbing costs the same as TSMC. AMD might be getting a sweet deal. TSMC has a lot of customers to play against each other. Is the situation the same with Samsung/GloFlo?

 

[el etro]

I feel Polaris 10 power averaging 100W exactly while gaming. Exact same as HD7870, and notebook GPU champ.

living in NZ I can tell you that the southern hemisphere sun is vicious. The stories you hear about the ozone are true.

PS: Didn't got the NZ D:


In last winter(summer here in most USA places) last days, termomethers measuring up to 48°C of thermical sensation in Rio de Janeiro(very big city then temperature varies more). I don't even wanna know how is to live in Ecquador or even in Brazil's Northeast these days.


Geography aside, i just want to empathise how important is TDP and operating temperature question these days.

 

[3DVagabond]

I just realized that we are assuming that Samsung fabbing costs the same as TSMC. AMD might be getting a sweet deal. TSMC has a lot of customers to play against each other. Is the situation the same with Samsung/GloFlo?

Add to that AMD basically gets fabbing done for a fixed rate until they go over their wafer allocations @ GloFo. One of those, "They are already buying the wafers, might as well use them", situations. They've had to pay GloFo for not buying wafers in the past. Hopefully this will prevent that folly.

 

[Elixer]

I guess we will see in a few months, but I expect extremely good prices vs present stuff.
...
I just realized that we are assuming that Samsung fabbing costs the same as TSMC. AMD might be getting a sweet deal. TSMC has a lot of customers to play against each other. Is the situation the same with Samsung/GloFlo?

Why would AMD (or anyone else--except for Apple) be getting a sweet deal from any fab at the launch of a new node?
This is all new tech on both sides, and it is incredibly expensive, so, I don't see extremely good prices coming down the pipeline, unless they are getting fantastic yields, and the fabs aren't charging a premium for their new tech.
They need to recoup their investments.

Look at the prices for the old node, and that is with a known process, and rock bottom prices for the tech, and they are still charging (especially nvidia) a price premium for the faster cards.

 

[Silverforce11]

About HBM and GDDR5 for Polaris, it can't be. It is either/or, because the entire memory subsystem is different, the uarch would have to be designed for either ground up.

Low cost part, GDDR5 with high clocks + next-gen memory bandwidth compression should be enough for that performance profile of Polaris 10.

Just compare the 380X, 1.4ghz vram clocks, 256 bit bus. It performs quite well in modern games.

Going to 1.75ghz vram clocks, which are available for awhile now on the 770, 970 etc, even the 390 has stock 1.5ghz vram which OC quite well, allows for a higher performance profile already by default. Add better bandwidth compression, and it will easily suffice for 390X+ performance.

Why bother with expensive HBM and time consuming interposer and TSV tech for entry and low-midrange SKU that need to be price sensitive?

 

[maddie]

Why would AMD (or anyone else--except for Apple) be getting a sweet deal from any fab at the launch of a new node?
This is all new tech on both sides, and it is incredibly expensive, so, I don't see extremely good prices coming down the pipeline, unless they are getting fantastic yields, and the fabs aren't charging a premium for their new tech.
They need to recoup their investments.

Look at the prices for the old node, and that is with a known process, and rock bottom prices for the tech, and they are still charging (especially nvidia) a price premium for the faster cards.

I understand what everyone is saying, but it's RTG top staff saying we should expect to be surprised at the product placement, not me. I'm only quoting.

Everyone estimating pricing ranges are basing it one their own values and reasoning and basically ignoring what we hear coming from the producer.

Do we ignore Koduri and Taylor? They are the ones claiming a big increase in performance/$ with Polaris.

 

[3DVagabond]

I understand what everyone is saying, but it's RTG top staff saying we should expect to be surprised at the product placement, not me. I'm only quoting.

Everyone estimating pricing ranges are basing it one their own values and reasoning and basically ignoring what we hear coming from the producer.

Do we ignore Koduri and Taylor? They are the ones claiming a big increase in performance/$ with Polaris.

Good points.

 

[maddie]

Add to that AMD basically gets fabbing done for a fixed rate until they go over their wafer allocations @ GloFo. One of those, "They are already buying the wafers, might as well use them", situations. They've had to pay GloFo for not buying wafers in the past. Hopefully this will prevent that folly.

That should really save them a lot of money.

 

[Silverforce11]

I understand what everyone is saying, but it's RTG top staff saying we should expect to be surprised at the product placement, not me. I'm only quoting.

Everyone estimating pricing ranges are basing it one their own values and reasoning and basically ignoring what we hear coming from the producer.

Do we ignore Koduri and Taylor? They are the ones claiming a big increase in performance/$ with Polaris.

Sure, but understand it's still PR and should be taken with a grain of salt.

Ofc, it would be great if they really move the perf/$ forward a lot, no doubts about that. Let's hope so too, but definitely cautious optimism is best, else your expectations may be too high and it will be a disappointment. Kinda like how Fury X was.

 

[maddie]

About HBM and GDDR5 for Polaris, it can't be. It is either/or, because the entire memory subsystem is different, the uarch would have to be designed for either ground up.

Low cost part, GDDR5 with high clocks + next-gen memory bandwidth compression should be enough for that performance profile of Polaris 10.

Just compare the 380X, 1.4ghz vram clocks, 256 bit bus. It performs quite well in modern games.

Going to 1.75ghz vram clocks, which are available for awhile now on the 770, 970 etc, even the 390 has stock 1.5ghz vram which OC quite well, allows for a higher performance profile already by default. Add better bandwidth compression, and it will easily suffice for 390X+ performance.

Why bother with expensive HBM and time consuming interposer and TSV tech for entry and low-midrange SKU that need to be price sensitive?

If the cards are giving a ~2.5 performance/watt of present gen and they are using high clocked GDDR5 for memory, what power the GPU core must be using. The core GPU must have a much larger efficiency increase than 2.5 times as the memory power budget should be close to constant.

I hope we get good analysis on release.

 

[IllogicalGlory]

On their graphic, Vega has a slightly higher performance per watt compared to Polaris. Perhaps that's simply because it's HBM?

 

[Silverforce11]

On their graphic, Vega has a slightly higher performance per watt compared to Polaris. Perhaps that's simply because it's HBM?

Yes, HBM2 is even more efficient. Removing the MC saves on die space and improves the perf/w of the chip. It's the major reason Fiji can have its efficiency gains over Hawaii.

Polaris GDDR5 being 2.5x efficient is easily 2x from the node change itself, the 0.5x will come from improved GCN.

 

[AtenRa]

1. Why HBM 1 will be significantly more expensive than GDDR-5 at 1.75GHz in 2H 2016 ?

2. At low TDPs (30-50W TDP) GDDR5 1.75GHz could use more power than the GPU chip itself.

3. Same memory type for more than Fury will be better for games and VR optimization. Especially for VR since AMD are going full throttle on it.

4. HMB 1 + 14nm FF = Mobile Heaven. Smaller package, lower power, same software optimization with big Desktop GPUs. AMD could take a huge market share with this even if HBM1 alone is a little bit more expensive than GDDR-5.

5. 14nm FF + HBM 1 = heaven for Desktop OEM systems as well. HBM1 will allow for huge performance SFF OEM desktop Systems with half-size Graphics Cards.

Im not saying Polaris is HBM 1 but im not going to discard the possibility. GDDR 5 was originally debuted on the HD4870 ($300) in June 2008 and less than a year later it was used on the first 40nm HD4770 $110 GPU (April 2009).

 

[Silverforce11]

It's not just the HBM chips.

IIRC, the Interposer process adds ~$30 to the PCB cost (read this somewhere legit awhile back). TSV I do not know the prices, but that process itself was the source for Fiji's low volume as the company that did it only went into volume production a month after Fury X was released.

 

[AtenRa]

Why HBM 2 will be cheaper than HBM 1 ??? they use the same tech and the same production (TSV, Interposers etc etc).

Just because HBM 1 was expensive in early 2015 and had low volume availability (Interposer etc etc) doesnt mean the situation is the same in 2H 2016.

Anyway we really have to wait and see, for now we are only speculate.

 

[Silverforce11]

Nobody said HBM2 will be cheaper.

Raja said HBM2 is expensive, watch the video.

Robert Hallock also said awhile ago, HBM2 will be used where the market allows it due to cost issues.

Small chips that are aimed at perf/$ just can't go HBM tech at this time. AMD have been quite clear on that.

 

[AtenRa]

Raja said they will bring HBM 2 when it will make sense to be used in mainstream. That only tells us that HBM 2 as HBM 1 needs more volume in order to drop the price and initial cost investment.

HBM was made to replace GDDR-5, if they cannot bring the cost close to the same level between the two, then HBM wouldn’t be in production and used in an actual product by now.