AMD: Speculation, Rumors, and Discussion (Archive)

[Jawed]

I think it's worth working backwards from GTX980 performance with ~224GB/s. It's really not very much bandwidth. A single stack of HBM would do it.

But a simple GDDR5 setup would also do that easily.

 

[iMacmatician]

Hum.. a 384bit bus would mean:

1 - That AMD wouldn't launch a single new solution with HBM throughout the whole 2016 (Hynix and/or Samsung probably wouldn't like that)
2 - At least 12 GDDR5 chips (AFAIK each GDDR5 chip can only connect through a 32bit bus max)
3 - A large PCB to house 12 VRAM chips
4 - A lot of power consumption coming from GDDR5, specially compared to HBM

It doesn't really fit AMD's statements about all Polaris solutions being very power-efficient.
Unless there's a severe shortage of HBM production or some other terrible event, I don't think AMD will go back to using GDDR5 in configurations that would need to go over 256bit wide. Or probably even less than that.
Polaris 10 had to be very cheap so it's probably coming with 128bit GDDR5.

The only way I could see GDDR5 being used in Polaris 11 is if it was the new GDDR5X memory at 10Gbps in a 256bit configuration. That would somewhat undermine points 2, 3 and 4, but point 1 still stands.
Samsung and Hynix didn't start volume production of HBM2 without having customers already signed on to purchase the chips. And I doubt Fiji alone (even a refresh using HBM2) would be enough to justify the production.

A more probable solution IMO is Polaris 11 / 480X using two stacks of HBM2 with a clock that is lower than the maximum rated 2Gbps for increased yields.

I'm not convinced that a Polaris 11 with GDDR5 would have significantly worse power efficiency than Polaris 10 (assuming they are on the same process) since Polaris 11 with 3072 SPs and a 384-bit bus has about the same ratio as Polaris 10 with 1024 SPs and a 128-bit bus, depending on precise clocks.

That's a good point about the HBM2 volume production. While Big Pascal with HBM2 may come out in 2016, the recent rumors and potential signs of a Pascal delay don't give me confidence.

And yeah, 2 stacks at around 1.5-1.6 Gbps would give a good amount of bandwidth and would be in line with Fiji (I actually considered putting 1.5 Gbps in my above post but settled on 2.0 Gbps). If AMD goes this route I wonder if we'll also see lower-end or laptop variants, where 256 GB/s is plenty, with 1 stack at 2.0 Gbps for lower cost and (maybe) smaller interposer size.

 

[Frenetic Pony]

I'd remind everyone the rumored specs are just that, and hardly confirmation. Also, while not shown, engineering samples for the large Polaris gpu have apparently already been delivered, even before the two smaller ones were. Since anandtech reported Amd is using both Samsung and Tsmc, it's not too hard to assume the big one will be on Tsmcs bigger but more reliable process, meaning it may well come out this year as well, though perhaps missing the "summer" target of the other two.

Finally, if we take a guess at the frequencies of finfet as say, 50% higher than 850mhz and assume linear scaling for performance (replace gddr5 with 5x for ram?) We see Polaris 10 just topping an overclocked 380x for performance. How this goes depends on cost versus Nvidia. If we assume Nvidias mid level 4 teraflop gpu (970 perf) goes for say, $250, then Nvidia could drop a 7/8ths binned gpu for $199 that would slightly outperform a full Polaris 10. But if a full p10 costs just $179 or $169 it might still look very attractive.

 

[Deleted member 13524]

Also, while not shown, engineering samples for the large Polaris gpu have apparently already been delivered, even before the two smaller ones were.

There are only 2 Polaris GPUs coming in 2016. It could be that Polaris 10 and 11 are different graphics cards using the same chip, with another GPU coming later on, but there are no more than 2 GPUs.

 

[Esrever]

There are only 2 Polaris GPUs coming in 2016. It could be that Polaris 10 and 11 are different graphics cards using the same chip, with another GPU coming later on, but there are no more than 2 GPUs.

Do you mean there is only 2 different cards? Logic would dictate at least 4 cards with 2 GPUs.

 

[gamervivek]

People with a 7970M have successfully flashed its bios to a 8970M and to a M290.

It's the same chip.

That's interesting, how successful was it? And if it's the same chip as GCN1 pitcairn, why does it show up alongside GCN2 products in AMD's OpenCL2.0 conformance list?

Tonga is very different from Tahiti. It may get the same performance as a higher-clocked Tahiti, but it does so with 66% of its memory bandwidth and lower overall power consumption.
There's a new UVD/VCE, color compression, FP16 instructions, TrueAudio, FreeSync, better VSR support and up to 4x better geometry performance on tesselation at >=32x.

The performance is similar was my point. The rest of the schmaltz isn't concerned with GCN itself except for fp16. The lower power consumption stuff you can also find with the slide I linked before. The M370X which shows up as venus xt, branded as a cape verde rebrand, shows similar performance improvement in some games like Tonga does over Tahiti.

http://forums.anandtech.com/showpost.php?p=37461727&postcount=133

Die area is different, transistor count is different... It's not the same chip.

Of course it isn't. But can you judge the same by looking at the chips?

 

[silent_guy]

Does anybody see any upside in telling the world your plans for the next of the year?

I can understand to a certain extent why they want to excite their base by talking about some relatively inconsequential upcoming display techniques, especially if they work on existing products.

But what does AMD gain by claiming that they'll only have 2 pieces of silicon? Or by disclosing how insanely great and power efficient their next low-end GPU will be, not tomorrow but something like 8 months from now.

They're giving away important competitive information and osborning themselves at the same time.

And finally: why only 2 pieces of silicon? You'd think they'd be up to something after releasing a grand total of 1 new silicon last year. But no.

AMD is full of surprises.

 

[Putas]

Does anybody see any upside in telling the world your plans for the next of the year?

With their market share heading towards "irrelevant company with no future" level, they might want to hype as many people as possible to wait with purchase until new products are ready.

 

[hoom]

How awesome would it be if the 480X was a 130W card with a performance between the 390X and the Fury, with 8GB HBM2, the form factor of the Nano and a very quiet cooler?
With HDMI 2.0 and full HEVC decoding/encoding, it would be the perfect card for many in the HTPC crowd.

Screw HTPC I'd be up for that on my main PC.

Also: time for a new sig

 

[Entropy]

Screw HTPC I'd be up for that on my main PC.

Yup. Quiet, performant, reasonably case friendly and hopefully not very expensive. There will be faster products at higher power draws obviously for those who want that, if not from AMD, then from nVidia. But a product such as the outlined one would fit my personal priorities nicely. Roughly 250mm2 should suffice, giving opportunities for a much larger flagship product later in the lithographic life cycle.

 

[Alexko]

Does anybody see any upside in telling the world your plans for the next of the year?

I can understand to a certain extent why they want to excite their base by talking about some relatively inconsequential upcoming display techniques, especially if they work on existing products.

But what does AMD gain by claiming that they'll only have 2 pieces of silicon? Or by disclosing how insanely great and power efficient their next low-end GPU will be, not tomorrow but something like 8 months from now.

They're giving away important competitive information and osborning themselves at the same time.

And finally: why only 2 pieces of silicon? You'd think they'd be up to something after releasing a grand total of 1 new silicon last year. But no.

AMD is full of surprises.

They might need to do it to prop up the stock, and perhaps dissuade people from buying Maxwell cards.

 

[hoom]

Yup. Quiet, performant, reasonably case friendly and hopefully not very expensive. There will be faster products at higher power draws obviously for those who want that

Exactly. I am unlikely to be going 4k anytime soon but could do with a bit more grunt than my current 7970 Ghz for my 30" and would love to have my room quieter/cooler (pretty hot today in our summer), cheaper power bills.

 

[Deleted member 13524]

Do you mean there is only 2 different cards? Logic would dictate at least 4 cards with 2 GPUs.

I don't know how it could have been any clearer :S
2 different GPUs. 2 different chips.

AMD is probably coming up with more than 4 cards with those 2 GPUs, probably closer to 8. Two desktop and two mobile graphics cards for each chip, at least.


If these latest rumors end up being correct, I'm expecting a full desktop line-up this year with the two new chips, plus Fiji and Tonga. Something like this:

- R7 460 - Polaris 10 "Pro" with 12 CU / 768 SP
- R7 460X - Polaris 10 "XT" with 16 CU / 1024 SP
- R7 470 - Tonga Pro with 28 CU / 1792 SP
- R7 470X - Tonga XT with 32 CU / 2048 SP
- R9 480 - Polaris 11 "Pro" with 42 CU / 2688 SP
- R9 480X - Polaris 11 "XT" with 48 CU / 3072 SP

- Fury rev.2 - Fiji Pro with 56 CU / 3584 SP (+ 8GB HBM and higher clocks like 1100MHz)
- Fury X rev.2 - Fiji XT with 64 CU / 4096 SP (+ 8GB HBM and higher clocks like 1200MHz)
- Fury X2 - Dual Fiji XT with 2*8GB and 1100MHz core clock.



Mobile parts would be a lot harder to predict since AMD went as far (and as low) as calling Pitcairn a R9 M390..
For all I know, they could be naming Polaris 10 as Fury Ultra MAXX 2 Mobile.

That's interesting, how successful was it?

See for yourself:
https://www.techinferno.com/index.p...ios-for-p377sm/&do=findComment&comment=127850
It's rarely done because flashing a new vBios means the laptops Bios will stop recognizing the graphics card, and there's no practical gain between using one Bios or the other.

And if it's the same chip as GCN1 pitcairn, why does it show up alongside GCN2 products in AMD's OpenCL2.0 conformance list?

Because AMD wanted to give the impression that something was new when they rebranded Pitcairn for the 99th time?

They're giving away important competitive information and osborning themselves at the same time.

So far, AMD only talked about the low-power chip, Polaris 10.
Osborning themselves would mean they were selling some chips within that range. They're not.
The discrete graphics for notebook market right now is probably 80% on GM107+GM108 and 10% on GM204. No notebook OEM is looking at AMD right now.

That said, they're disclosing this info in the hopes that people will wait for notebooks with Polaris GPUs instead of buying one right now with a Maxwell card.



EDIT: Am I the only one who's having lots of trouble with copy/paste when writing posts?

 

[Gubbi]

Being 14/16nm parts, Polaris would likely pack a lot more CUs than previous parts. Tonga is 366 mm² on 28nm.
I expect something like:
Polaris 10: 32 CUs (2K SPs), GDDR5/X memory, around 150mm² die size
Polaris 11: 96 CUs (6K SPs), HBM2 memory, around 350mm² die size

Cheers

 

[Deleted member 13524]

Being 14/16nm parts, Polaris would likely pack a lot more CUs than previous parts. Tonga is 366 mm² on 28nm.
I expect something like:
Polaris 10: 32 CUs (2K SPs), GDDR5/X memory, around 150mm² die size
Polaris 11: 96 CUs (6K SPs), HBM2 memory, around 350mm² die size

Cheers

Polaris 10 is definitely not a 32CU part, or it wouldn't be showcased against a Geforce GTX 950.
Simply packing more CUs wouldn't exactly be the best bet for AMD, since none of the previous GCN GPUs are exactly compute-limited.

 

[Gubbi]

Sorry missed the GTX 950 comparison.

That seems thoroughly unambitious.

Cheers

 

[I.S.T.]

I suspect the revealed Polaris is a low end part. Like, lower end than whatever you'd call the GTX 950's market segment. If that's true, using the GTX 950 as a benchmark is pretty damn smart.

If that's true. I hope so.

 

[Deleted member 13524]

Sorry missed the GTX 950 comparison.

That seems thoroughly unambitious.

Cheers

Offering GM206 performance at GM107/GM108 power consumption is anything but unambitious. Unless nVidia rushes in with a Pascal alternative, it would take the notebook market by storm.

 

[silent_guy]

So far, AMD only talked about the low-power chip, Polaris 10.
Osborning themselves would mean they were selling some chips within that range. They're not.
The discrete graphics for notebook market right now is probably 80% on GM107+GM108 and 10% on GM204. No notebook OEM is looking at AMD right now.

That said, they're disclosing this info in the hopes that people will wait for notebooks with Polaris GPUs instead of buying one right now with a Maxwell card.

Maybe...

I see a lot of excitement because of the Polaris announcement, and very little is related to laptops.

 

[Deleted member 13524]

I see a lot of excitement because of the Polaris announcement

Of course you'll see people being excited for the underdog. Enthusiasts usually enjoy competition.
AMD has this year to turn things around or everything goes to shit.

And to be honest, regardless of which IHV in question, I would always be much more excited about an affordable laptop GPU that brings PS4 performance to the masses than about something-neural-networks-deep-whatever-that-maybe-could-eventually-be-used-someday-maybe-in-vehicles-so-expensive-I-wouldn't-buy-even-if-I-could-afford.
So there's that.