AMD Polaris 10 Samples work at 1.27 GHz

[RussianSensation]

Truth! Having a GPU sitting there unused when you've paid $$ for it sucks. Period.

Why are they unused then? Since April 12th, my 3 Hawaii chips have mined $494 USD as of 4:01pm May 20th. By the time GTX1070 drops June 10th, that will grow to ~$700 USD. How many hours a day do NV GPUs game? 3-5 hours, wasted for the other 19-21 hours per day? If you want to talk about unused hardware collecting dust....modern ASICs are General Purpose so why should be use them only for games in 2016? Imagine, imagine a world where a GPU plays games and does something else too. 6x Polaris 10 GPUs will generate $480 USD a month for only $400 higher price than GTX1080 SLI. Now how much is that by the time Big Pascal drops? Just saying.

Fury Air, I said. And it is 50%+ faster at 1080p and very close to that at 1440p (47%-49% from your 2 sources), which in the context of Polaris 10 I would imagine most would play games here. If Polaris 10 performance is here, based on speculative die sizes, I would be happy.

Why would you compare a $599-699 GTX1080 to Fury Air knowing AMD's Polaris 10 is a $349 and lower card? Makes no sense. Gamers don't just decide to spend double or $300+ more because the flagship card is 50% faster. That's now how the market works. If it did, almost no GTX960 cards would be sold against R9 290/390.

Why you'd pay $700 for 1080p... I'd never know.

1080p is a CPU limited resolution too. We have to be extremely careful when discussing it because modern sites use overclocked i7 6700K @ 4.5Ghz and many gamers have anemic i5 2500K/i5 3570K/i5 6400 at stock, etc.

In XCOM 2, i7 6700K @ 4.5Ghz was 65% faster than an i5 2500K with the same GTX1080.
http://www.computerbase.de/2016-05/...diagramm-xcom-2-1920-1080-intel-core-i5-2500k

But you already know how this forum goes:

- CPU bottlenecks are downplayed
- 1080p 60Hz resolution gaming on high-end cards is defended
- lower end stock Sandy/Ivy Bridge CPUs are defended
- Benchmarks of i7 smashing i5s are ignored
- 2GB cards are defended (but 4GB HBM1 are attacked ).

The average/mainstream PC gamer has no business buying GTX1070/1080/980Ti/Fury X level cards. I guarantee it their PC isn't fast enough to extract the maximum performance out of them in the first place. Even if they do have max overclocked i7 2600K or faster, then such cards will be rendering FPS far above 60, essentially wasting it the 60Hz screen.

It's a perfectly logical and rational view... assuming AMD were in a profit-making position and not *desperate* to regain share. A 25% share loss in two years in any industry is nothing short of an epic catastrophe and should result in very desperate measures. I still think you're probably right, but I think there's a higher probability than you that they actually show up with a 36 (near 1070) or 40 cu ( above 1070 ) part at the $300 point. Keep in mind also that for your view to be correct, it means SiSoft is misrepresenting the number of compute units.

Also, keep in mind that AMD has very much indicated they expect some aib to include g5x, which wouldn't make sense on a 32cu part.

You are right. I am just going off based on rumours. If P10 is a 2560-2816 CC part with 1.3-1.5Ghz clocks, that's one thing. If it's only a 2048-2304 shader part, that's totally different.

I think you are right, though. If "all" AMD wanted was to achieve 290x performance at a far better price, they would have just shrunk Hawaii, kept the frequence and had a bit of overclocking headroom. But they didn't. They added transistors and they improved on the architecture...

No, they wouldn't have done that. 14nm Die shrink is expensive and Hawaii was much larger than PS4's APU:

"This ‘die-shrink’ requires to re-develop the same chip again, with a cost measured in excess of a hundred million dollars (est. $120-220 million)."

At this point, it's not cost effective to shrink Hawaii. Then you have to add the cost of incorporating new 4K UVD engine, HDMI 2.0a, and now a 100% required next gen memory compression tech since you are going down to a 256-bit bus with GDDR5/X. All of a sudden, this shrunken Hawaii R9 290 chip would cost you nearly as much as just designing a brand new 14nm chip from the ground-up. It's also more cost effective for AMD to sell 256-bit PCB with a lower pin count than a 512-bit Hawaii chip. Did you consider that if you were to die shrink the Hawaii chip, you wouldn't be able to fit all the necessary chip pins required for a 512-bit bus? It's not a simple as just randomly assigning 384-512 bit buses to any size chip.

BTW, AMD never said which Polaris chip would be at Hawaii-performance. They only stated that was one of the performance goals when designing the architecture.

That implication itself is a BIG deal. They are implying 290X/390X +/-10% level of performance at sub-$349 price levels with all the upgraded features, I/O, 8GB of memory and ~half the power usage. If 2048-2304 SP P10 is as fast as a Fury X, then a 4096 SP Vega that scales well over P10 starts to sound too good to be true.

 

[Arachnotronic]

Why are they unused then? Since April 12th, my 3 Hawaii chips have mined $494 USD as of 4:01pm May 20th. By the time GTX1070 drops June 10th, that will grow to ~$700 USD. How many hours a day do NV GPUs game? 3-5 hours, wasted for the other 19-21 hours per day? If you want to talk about unused hardware collecting dust....modern ASICs are General Purpose so why should be use them only for games in 2016? Imagine, imagine a world where a GPU plays games and does something else too. 6x Polaris 10 GPUs will generate $480 USD a month for only $400 higher price than GTX1080 SLI. Now how much is that by the time Big Pascal drops? Just saying.

Are you saying that if I were to invest $35k in 100 Polaris 10 chips, I could rake in ~$4800/month without much worry that this figure will plunge over the course of a year or two? Asking seriously here.

 

[ElFenix]

I double checked and GM206 underpins both the GTX950 and the GTX960, easily the highest volume dGPU laptop products. You are indeed correct about the top end, GM204 powering the GTX980M but this again will be countered by a Polaris 10 in mobile.

interesting. notebookcheck, TPU, and wikipedia all show GM107 powering the 950m and 960m. are you sure you weren't looking at desktop parts?

(frankly i was amazed to find that GM204 powers the 980m, when was the last time the same chip powered both the desktop and mobile version of a model number?)

 

[raghu78]

That implication itself is a BIG deal. They are implying 290X/390X +/-10% level of performance at sub-$349 price levels with all the upgraded features, I/O, 8GB of memory and ~half the power usage. If 2048-2304 SP P10 is as fast as a Fury X, then a 4096 SP Vega that scales well over P10 starts to sound too good to be true.

RS AMD have not implied exact performance of Polaris at any point of time. What we have now is rumours and guesses. We do not have a clue as to the architectural improvements in Polaris and how much they affect perf/sp and perf/CU. Actual specs like sp/CU/TMU count is still unknown. Die size is confirmed for small Polaris to be around 110 sq mm (anandtech rough guess after seeing the GPU at CES 2016) while for big Polaris die size is rumoured to be 230-240 sq mm. Add to that the GF 14LPP process and the actual clocks at launch. Nobody will know actual performance until the GPU launches and the reviews are released.

 

[Orvogg]

No, they wouldn't have done that. 14nm Die shrink is expensive and Hawaii was much larger than PS4's APU:

"This ‘die-shrink’ requires to re-develop the same chip again, with a cost measured in excess of a hundred million dollars (est. $120-220 million)."

At this point, it's not cost effective to shrink Hawaii. Then you have to add the cost of incorporating new 4K UVD engine, HDMI 2.0a, and now a 100% required next gen memory compression tech since you are going down to a 256-bit bus with GDDR5/X. All of a sudden, this shrunken Hawaii R9 290 chip would cost you nearly as much as just designing a brand new 14nm chip from the ground-up. It's also more cost effective for AMD to sell 256-bit PCB with a lower pin count than a 512-bit Hawaii chip. Did you consider that if you were to die shrink the Hawaii chip, you wouldn't be able to fit all the necessary chip pins required for a 512-bit bus? It's not a simple as just randomly assigning 384-512 bit buses to any size chip.

Interesting. I am the first to admit I know more or less nothing when it comes to the intricacies of chip design, but if I read what you are saying correctly, then having less SP would decrease the need of memory speed? Has the need of a wide memory bus more to do with the number of SPs/CCs, or more with the general performance? If the rumours are true, it would certainly seem that way with P10 having fewer SPs and narrower memory bus but the, presumably, same performance of Hawaii...

That implication itself is a BIG deal. They are implying 290X/390X +/-10% level of performance at sub-$349 price levels with all the upgraded features, I/O, 8GB of memory and ~half the power usage. If 2048-2304 SP P10 is as fast as a Fury X, then a 4096 SP Vega that scales well over P10 starts to sound too good to be true.

Only if they manage to come up with an architecture that will scale up to 600mm^2 this time. We saw what happened with Fiji, there were no more room to scale up after Hawaii. Besides, if 4096 SP Vega is to be competitive with GP104 they really need to hit Fury performance with Polaris, unless they have more architectural improvements in store with Vega (and not only power efficiency)

 

[Paul98]

Performance is really going to depend on how high this is going to be clocked. I am looking at this using data from the 380x and 390x. If it were on 28nm full polaris should be faster than 390x if they were both clocked the same.

 

[railven]

Are you saying that if I were to invest $35k in 100 Polaris 10 chips, I could rake in ~$4800/month without much worry that this figure will plunge over the course of a year or two? Asking seriously here.

Got to spend money to make money.

I'll give it to AMD, if their cards aren't being used for gaming, at least people are using them for mining.

 

[crisium]

Why would you compare a $599-699 GTX1080 to Fury Air knowing AMD's Polaris 10 is a $349 and lower card? Makes no sense. Gamers don't just decide to spend double or $300+ more because the flagship card is 50% faster. That's now how the market works. If it did, almost no GTX960 cards would be sold against R9 290/390.

Lol, please RS. Sometimes you try to get into too many conversations. swilli89 was wondering who thinks GTX 1080 will be 50% faster than full Polaris 10, since that's roughly the die size difference based on current P10 numbers. I said think this is very possible, and that would make Polaris 10 Fury Air speed. And actually, GTX 1080 is 314mm right? If Polaris 10 is actually 232mm, that's only a 35% size advantage for Nvidia. So if AMD can match Nvidia performance per mm it should be faster than Fury Air.

 

[Flapdrol1337]

Lol, please RS. Sometimes you try to get into too many conversations. swilli89 was wondering who thinks GTX 1080 will be 50% faster than full Polaris 10, since that's roughly the die size difference based on current P10 numbers. I said think this is very possible, and that would make Polaris 10 Fury Air speed. And actually, GTX 1080 is 314mm right? If Polaris 10 is actually 232mm, that's only a 35% size advantage for Nvidia. So if AMD can match Nvidia performance per mm it should be faster than Fury Air.

Being in a lower tier the polaris probably only has a gddr5 memory interface (no x), so that's extra die space, also the glofo node has a ~15% higher density, and amd usually has a higher density by design too (a reason why they generally clock lower?)

 

[littleg]

Has the ring of K8 vs Netburst to it. As Intel found out, higher clocks will only get you so far. Wonder if Nvidia have a Conroe waiting in the wings.....

 

[Arachnotronic]

Has the ring of K8 vs Netburst to it. As Intel found out, higher clocks will only get you so far. Wonder if Nvidia have a Conroe waiting in the wings.....

Uh huh. Is there anything from Pascal's results to suggest that it is a power guzzler and/or under performer like Netburst was?

As I recall, it is Hawaii, not Maxwell/Pascal that is the power guzzler.

 

[littleg]

Uh huh. Is there anything from Pascal's results to suggest that it is a power guzzler and/or under performer like Netburst was?

As I recall, it is Hawaii, not Maxwell/Pascal that is the power guzzler.

I was commenting on the philosophy rather than the implementation. Since you mention it anyway, 83 degrees is quite hot for this calibre of GPU and the TDP of the 1080 is a little higher than I had expected from the node shrink.

Netburst didn't necessarily underperform anyway, it just couldn't clock as high as they intended and that screwed the game for them.

Once you run out of headroom in that kind of design you're boned, there's nowhere you can go. Bringing up your IPC on the other hand means you can always go for higher clocks as the process matures.

 

[IEC]

Being in a lower tier the polaris probably only has a gddr5 memory interface (no x), so that's extra die space, also the glofo node has a ~15% higher density, and amd usually has a higher density by design too (a reason why they generally clock lower?)

Polaris 10 is rumored to have support for GDDR5X. We'll have to wait for 5/31 for more details (hopefully).

 

[Arachnotronic]

Once you run out of headroom in that kind of design you're boned, there's nowhere you can go. Bringing up your IPC on the other hand means you can always go for higher clocks as the process matures.

Please tell me that you're joking?

 

[Flapdrol1337]

Once you run out of headroom in that kind of design you're boned, there's nowhere you can go. Bringing up your IPC on the other hand means you can always go for higher clocks as the process matures.

I thought if you had a low theoretical max clock you only burn much more power trying to get a little closer to it.

A shrink would get you higher clocks though I think?

Anyway, I don't know too much about this stuff.

 

[JDG1980]

Polaris 10 is rumored to have support for GDDR5X.

I can't find the link now, but someone said this was confirmed by AMD on the 5/18 conference call.

 

[swilli89]

Lol, please RS. Sometimes you try to get into too many conversations. swilli89 was wondering who thinks GTX 1080 will be 50% faster than full Polaris 10, since that's roughly the die size difference based on current P10 numbers. I said think this is very possible, and that would make Polaris 10 Fury Air speed. And actually, GTX 1080 is 314mm right? If Polaris 10 is actually 232mm, that's only a 35% size advantage for Nvidia. So if AMD can match Nvidia performance per mm it should be faster than Fury Air.

Ah you're right. That's that I get for rounding in my head. We have seen cases from both companies depending on the generation.. I'm really just basing my estimate of Polaris higher perf/mm on the fact that:

1) GF/Samsung's process may be 10-15% denser
2) nVidia probably had to add transistors to bridge the gap in DX12
3) AMD was probably able to lean out some components like FP64 getting paired down
4) Polaris seems to be a mobile-first design (promising in my book as legendary chips like AMD's Barton Athlon-XP, and Intel's Conroe were both born from this design philosophy!)
5) AMD historically makes denser/lower clocking chips

^^So AMD was saddled with a slightly over-engineered card in Tahiti/Hawaii (a great bonus for those like my that are keeping 290X gen cards for going on 2 years now) and nVidia went with the lean and mean approach for Kepler/Maxwell. This gen it looks like each firm went the other way which is why Pascal didn't quite live up to what some were expecting for performance/watt and performance/mm. Again though, we have no other Finfet cards to compare Pascal to so its metrics may prove to be best in class yet, we just have to wait to find out.

At the end of the day does perf/mm matter? In my eyes not at all! I'm just merely discussing Polaris' potential performance compared to Pascal based on this metric, once we know for sure where the chips fall in terms of performance I couldn't care less about this.

 

[Olecki]

1080p is a CPU limited resolution too. We have to be extremely careful when discussing it because modern sites use overclocked i7 6700K @ 4.5Ghz and many gamers have anemic i5 2500K/i5 3570K/i5 6400 at stock, etc.

In XCOM 2, i7 6700K @ 4.5Ghz was 65% faster than an i5 2500K with the same GTX1080.
http://www.computerbase.de/2016-05/...diagramm-xcom-2-1920-1080-intel-core-i5-2500k

But you already know how this forum goes:

- CPU bottlenecks are downplayed
- 1080p 60Hz resolution gaming on high-end cards is defended
- lower end stock Sandy/Ivy Bridge CPUs are defended
- Benchmarks of i7 smashing i5s are ignored
- 2GB cards are defended (but 4GB HBM1 are attacked ).

The average/mainstream PC gamer has no business buying GTX1070/1080/980Ti/Fury X level cards. I guarantee it their PC isn't fast enough to extract the maximum performance out of them in the first place. Even if they do have max overclocked i7 2600K or faster, then such cards will be rendering FPS far above 60, essentially wasting it the 60Hz screen.

Hmmm... I though that whole hype with DX12 is possibility to move some of the compute workload from bottlenecked processor to much faster in compute tasks GPU- basically changing GPU into co-processor. If I'm right - the slower your procesor is you should get bigger boost if you buy powerful GPU. So if you have i5 6500 stock and you buy GTX1070/1080/980Ti/Fury X you should get nice results in both CPU and GPU heavy games.

 

[maddie]

Hmmm... I though that whole hype with DX12 is possibility to move some of the compute workload from bottlenecked processor to much faster in compute tasks GPU- basically changing GPU into co-processor. If I'm right - the slower your procesor is you should get bigger boost if you buy powerful GPU. So if you have i5 6500 stock and you buy GTX1070/1080/980Ti/Fury X you should get nice results in both CPU and GPU heavy games.

You're right. DX12 reduces CPU bottlenecks. Not really due to transferring the compute load however.

One thing about over reliance on historical models. You get lost when the paradigm changes, and we're definitely in one of those periods.

 

[wahdangun]

Uh huh. Is there anything from Pascal's results to suggest that it is a power guzzler and/or under performer like Netburst was?

As I recall, it is Hawaii, not Maxwell/Pascal that is the power guzzler.

yes it is, if u see the review it was quite clear that pascal sacrifice IPC to get higher clock, just like netburst.

 

[Headfoot]

Every generation of graphics card has had more and more dynamic clocking ability and turboing features. I don't trust pre-release software utility clock speed predictions AT ALL. Is that base clock? Is it a random point in a dynamic range? Is it a mobile chip? Is it TDP limited? What's the TDP used in that test? Is PowerTune or similar feature turned up, down, off? Way too many variables now that clock is no longer a set and forget target.

 

[Flapdrol1337]

yes it is, if u see the review it was quite clear that pascal sacrifice IPC to get higher clock, just like netburst.

Not sure if they sacrificed all that much. 1080 has 50% more tflops/s than 980Ti, and is 37% faster according to techpowerup.

I'd attribute most of that difference to not increasing the memorybandwidth at all.

 

[Arachnotronic]

yes it is, if u see the review it was quite clear that pascal sacrifice IPC to get higher clock, just like netburst.

Netburst's problem was that it made this trade off and was a power hog/underperformer for it. NVIDIA seems to have driven frequencies up significantly at roughly similar perf/clock while improving power efficiency.

Netburst this is not.

 

[ShintaiDK]

"Speedracer" designs work until you reach the physical limit so to say.

 

[wahdangun]

Not sure if they sacrificed all that much. 1080 has 50% more tflops/s than 980Ti, and is 37% faster according to techpowerup.

I'd attribute most of that difference to not increasing the memorybandwidth at all.

but how much ??

GTX 1080

GTX 980 TI @ 1,4 Ghz


now u can see that GTX 980 TI @ 1,4 Ghz can match GTX 1080 @ 1,8 Ghz.

so GTX 1080 need 400 Mhz more to get same performance.

Netburst's problem was that it made this trade off and was a power hog/underperformer for it. NVIDIA seems to have driven frequencies up significantly at roughly similar perf/clock while improving power efficiency.

Netburst this is not.

it was still not clear if it was power hog/underperformer until we can test AMD equivalent, but it was still displaying regression in IPC like P3 to P4 or star architecture to bulldozer.