How memory (frequency) starved is Polaris?

[ultima_trev]

Looking at the graphics scores (not overall) on Fire Strike it would seem the RX 480 seems to benefit a lot, if not more so from memory overclocking than core overclocking. Considering these are modest core clocks but 2250 MHz base clock for the GDDR5 chips:

Stock Clocks: http://www.3dmark.com/fs/9100800

1355 / 2250: http://www.3dmark.com/fs/9136460

1320 / 2250: http://www.3dmark.com/fs/9200319

Just a theory, I believe if AMD could have launched Polaris 10 with better power delivery, better cooling, +100 MHz core clock and 10 Gbps GDDR5X, the results of Polaris versus R9 Fury / Nano / GTX 980 would have been very different versus the present.

 

[Silverforce11]

It's bandwidth starved. I saw a dutch review site do core vs memory OC and it scales near perfect for memory OC.

 

[Irenicus]

Looking at the graphics scores (not overall) on Fire Strike it would seem the RX 480 seems to benefit a lot, if not more so from memory overclocking than core overclocking. Considering these are modest core clocks but 2250 MHz base clock for the GDDR5 chips:

Stock Clocks: http://www.3dmark.com/fs/9100800

1355 / 2250: http://www.3dmark.com/fs/9136460

1320 / 2250: http://www.3dmark.com/fs/9200319

Just a theory, I believe if AMD could have launched Polaris 10 with better power delivery, better cooling, +100 MHz core clock and 10 Gbps GDDR5X, the results of Polaris versus R9 Fury / Nano / GTX 980 would have been very different versus the present.

Speaking of which... can polaris even use gddr5x?

And if so, could a board partner just choose to it instead of gddr5?

What is the cost difference of 8GB of gddr5 vs gddr5x? would it just not be worth the cost difference?

 

[ultima_trev]

Something AMD should know the answers to. Perhaps there will be GDDR5X in a future RX 485 refresh?

 

[Dresdenboy]

Something AMD should know the answers to. Perhaps there will be GDDR5X in a future RX 485 refresh?

According to their naming scheme, a RX490 should have >256GB/s. And based on the hints given in this thread, this should be enough to improve on the same number of CUs. But AMD seems to use less dies for different purposes (Polaris 10, 11 for current gen GPUs and Zeppelin for everything Zen based from DT to 32C server via interposer). FinFET development and processing costs caused this. So Vega+HBM2 or Polaris+GDDR5X are the likely upgrades here.

 

[GodisanAtheist]

Speaking of which... can polaris even use gddr5x?

And if so, could a board partner just choose to it instead of gddr5?

What is the cost difference of 8GB of gddr5 vs gddr5x? would it just not be worth the cost difference?

- I believe Raja was quoted as saying the Polaris arch was fully compatible with gddr5x in some pre-release interview.

Wouldn't be surprising if the 490 was just a 480 with the bandwidth and power caps taken off.

 

[poofyhairguy]

Just a theory, I believe if AMD could have launched Polaris 10 with better power delivery, better cooling, +100 MHz core clock and 10 Gbps GDDR5X, the results of Polaris versus R9 Fury / Nano / GTX 980 would have been very different versus the present.

Sounds like a great AMD 580 card.

 

[boozzer]

Sounds like a great AMD 580 card.

gotta pace it for 3+ years. 14/16nm is not going away for 3+ years. my estimate of course.

 

[JDG1980]

Looking at the graphics scores (not overall) on Fire Strike it would seem the RX 480 seems to benefit a lot, if not more so from memory overclocking than core overclocking. Considering these are modest core clocks but 2250 MHz base clock for the GDDR5 chips:

Stock Clocks: http://www.3dmark.com/fs/9100800

1355 / 2250: http://www.3dmark.com/fs/9136460

1320 / 2250: http://www.3dmark.com/fs/9200319

Just a theory, I believe if AMD could have launched Polaris 10 with better power delivery, better cooling, +100 MHz core clock and 10 Gbps GDDR5X, the results of Polaris versus R9 Fury / Nano / GTX 980 would have been very different versus the present.

Those are very encouraging results. A modest core overclock combined with increasing the memory bandwidth by 12.5% yields a 3DMark FireStrike Graphics Score of 15241. That's nearly a 19% improvement over the stock score (12823). Going by VideoCardz's benchmark chart, it would fall between the vanilla Fury and the Fury X. That's closer to what a lot of us hoped to see out of Polaris 10.

If the memory controller of Polaris 10 supports GDDR5X (and I believe someone from AMD once stated that it does), then that means we could be seeing a much better P10 card later this year. Using 10Gbps GDDR5X (the same as on the GTX 1080) on P10 would increase memory bandwidth by 25% (from 256GB/sec to 320GB/sec) and simultaneously allow the memory clock to be decreased by 37.5%, from 2000 MHz to 1250 MHz. That's huge. AMD's statement on the power usage issue implicitly blamed the "unprecedented 8 Gbps" memory clock. Add to that the fact that GloFo's 14LPP process is still immature and will almost certainly improve throughout the year, and it's quite possible that AMD could bring out a new "RX 485" in October-November that switches to GDDR5X, increases the core clocks a bit, and keeps power consumption to reasonable levels, providing much better performance and perf/watt than the RX 480.

 

[Arachnotronic]

gotta pace it for 3+ years. 14/16nm is not going away for 3+ years. my estimate of course.

We'll see. Wouldn't be surprised if Navi/Volta were 10nm, but also wouldn't be surprised if they were 14/16nm.

 

[nurturedhate]

We'll see. Wouldn't be surprised if Navi/Volta were 10nm, but also wouldn't be surprised if they were 14/16nm.

Thought the current rumors were both are skipping 10nm and going to 7nm.

 

[Arachnotronic]

Thought the current rumors were both are skipping 10nm and going to 7nm.

Nope, NVIDIA at least is going to 10nm.

 

[tviceman]

It's bandwidth starved. I saw a dutch review site do core vs memory OC and it scales near perfect for memory OC.

Yet the 7ghz vram 480's are only ~2-3% slower. Makes no sense at all.

 

[JDG1980]

Yet the 7ghz vram 480's are only ~2-3% slower. Makes no sense at all.

The launch day reviews were done with special cards that had a BIOS switch to switch between 4GB and 8GB. Are we sure that this switch affected the RAM clocks as well? Or did the reviewers wind up with a configuration (4GB @ 2000 MHz clock) that doesn't directly correspond to retail products?

 

[boozzer]

We'll see. Wouldn't be surprised if Navi/Volta were 10nm, but also wouldn't be surprised if they were 14/16nm.

I would be very happy if I am wrong :thumbsup:

 

[Wall Street]

The launch day reviews were done with special cards that had a BIOS switch to switch between 4GB and 8GB. Are we sure that this switch affected the RAM clocks as well? Or did the reviewers wind up with a configuration (4GB @ 2000 MHz clock) that doesn't directly correspond to retail products?

It is more complicated than that. Most of the launch day 4 GB models have 8 Gbps RAM but some may be shipped with 7 Gbps RAM.

 

[SPBHM]

The launch day reviews were done with special cards that had a BIOS switch to switch between 4GB and 8GB. Are we sure that this switch affected the RAM clocks as well? Or did the reviewers wind up with a configuration (4GB @ 2000 MHz clock) that doesn't directly correspond to retail products?

it would make no sense at all for the bios to reduce the card to 4GB and not reduce the clock to the correct spec (7GHz), also as the performance difference shows on the test with both (8GB and 4GB), that's the case, if you had at 8GHz on both you would get the same performance

the retail 4GB card also comes with 8GHz memory clocked at 7GHz

 

[tviceman]

The launch day reviews were done with special cards that had a BIOS switch to switch between 4GB and 8GB. Are we sure that this switch affected the RAM clocks as well? Or did the reviewers wind up with a configuration (4GB @ 2000 MHz clock) that doesn't directly correspond to retail products?

If that was the case, then there should have been no measurable difference in performance, but several sites ran the 4gb card through benches and it came out consistently 2-3% slower. So the reviewed "4gb" cards were running at 7ghz on the vram and when increasing the vram speed by 14% from 7 to 8ghz yields only a 2% increase in performance the RX 480 is not bandwidth starved.

http://www.anandtech.com/show/10446/the-amd-radeon-rx-480-preview/4
http://www.anandtech.com/show/10446/the-amd-radeon-rx-480-preview/5
http://www.gamersnexus.net/guides/2503-amd-rx-480-4gb-vs-8gb-benchmark-is-it-worth-it

Only in games where 4gb of vram is insufficient is there a discernible performance difference. It is not particularly bottle necked at 7ghz, let alone 8ghz.

 

[Dresdenboy]

We'll see. Wouldn't be surprised if Navi/Volta were 10nm, but also wouldn't be surprised if they were 14/16nm.

I'm agreeing with you. While Vega can still go the higher TDP, updated uarch, more shaders at more sweet spot like V/F routes, Navi and possibly also Volta might already use 3D stacked memory, thus they need to dissipate even less power per mm².

Nope, NVIDIA at least is going to 10nm.

That's quite interesting!

 

[Thala]

So the reviewed "4gb" cards were running at 7ghz on the vram and when increasing the vram speed by 14% from 7 to 8ghz yields only a 2% increase in performance the RX 480 is not bandwidth starved.

Yet every review side which overclocked the memory got significant performance increases.
This makes me rather believe, there was something wrong with AMDs "underclocking" of the 4/8Gbyte hybrid samples otherwise the big performance increases are not explainable. Optionally the 7Ghz BIOS could set the memory to more agressive timings, which could improve effective bandwidth. Your mistake is, that you compare cards with different BIOS and completely unknown memory controller settings.

In any case the 8GByte model is bandwidth starved as all the test show in unison with the very same BIOS.

 

[tviceman]

Yet every review side which overclocked the memory got significant performance increases.

Synthetic benchmarks are NOT real world examples of how overclocking works. That said, how many reviews did you actually read that included overclocking? You apparently missed these two, which posted abysmal results.
https://www.techpowerup.com/reviews/AMD/RX_480/27.html
http://www.guru3d.com/articles_pages/amd_radeon_r9_rx_480_8gb_review,35.html

In fact, searching the web for legitimate RX 480 reviews that included OCing, the ONLY favorable one I found was eurogamer and they didn't account for fan speed when they overclocked vs. stock. Most sites didn't or haven't yet done OCing scores while some sites say Wattman just crashed left and right.

 

[Thala]

Synthetic benchmarks are NOT real world examples of how overclocking works. That said, how many reviews did you actually read that included overclocking? You apparently missed these two, which posted abysmal results.
https://www.techpowerup.com/reviews/AMD/RX_480/27.html
http://www.guru3d.com/articles_pages/amd_radeon_r9_rx_480_8gb_review,35.html

In fact, searching the web for legitimate RX 480 reviews that included OCing, the ONLY favorable one I found was eurogamer and they didn't account for fan speed when they overclocked vs. stock. Most sites didn't or haven't yet done OCing scores while some sites say Wattman just crashed left and right.

Oh dear, we are not trying to answer the question of how well a RX480 overclocks, if Wattman is crashing or what fan speed settings are required but if the RX480 is memory starved. Seems you want to make an argument, which fits your narrative but you are going way off-topic here.
Anyway, computerbase.de investigated how performance scales only with memory overclock, leaving the GPU at 1266MHz.

https://www.computerbase.de/2016-06/radeon-rx-480-test/12/

They concluded that the gap to R9390X and GTX980 is closed by just overclocking memory.

 

[tviceman]

Oh dear, we are not trying to answer the question of how well a RX480 overclocks, if Wattman is crashing or what fan speed settings are required but if the RX480 is memory starved. Seems you want to make an argument, which fits your narrative but you are going way off-topic here.
Anyway, computerbase.de investigated how performance scales only with memory overclock, leaving the GPU at 1266MHz.

https://www.computerbase.de/2016-06/radeon-rx-480-test/12/

They concluded that the gap to R9390X and GTX980 is closed by just overclocking memory.

To which I linked two websites to your one that overclocked both the memory and core but the RX 480 neither scaled well not caught up to a gtx 980. AMD crapped out another technical turd and despite all the prehype that was inevitably let down, people are still trying to spit shine this thing. The one and only redeeming quality is the $199 price tag for the 4gb.

 

[Thala]

To which I linked two websites to your one that overclocked both the memory and core but the RX 480 neither scaled well not caught up to a gtx 980. AMD crapped out another technical turd and despite all the prehype that was inevitably let down, people are still trying to spit shine this thing. The one and only redeeming quality is the $199 price tag for the 4gb.

For the second time you going off topic here and turning this thread into a FUD contest. I close my case, not worth discussing the OPs topic with you as you are more interested in throwing FUD around.

 

[guskline]

Looking at the graphics scores (not overall) on Fire Strike it would seem the RX 480 seems to benefit a lot, if not more so from memory overclocking than core overclocking. Considering these are modest core clocks but 2250 MHz base clock for the GDDR5 chips:

Stock Clocks: http://www.3dmark.com/fs/9100800

1355 / 2250: http://www.3dmark.com/fs/9136460

1320 / 2250: http://www.3dmark.com/fs/9200319

Just a theory, I believe if AMD could have launched Polaris 10 with better power delivery, better cooling, +100 MHz core clock and 10 Gbps GDDR5X, the results of Polaris versus R9 Fury / Nano / GTX 980 would have been very different versus the present.

Perhaps, but at a $199 4g/$239 8g price point neither AMD nor Nvidia could or would use GDDR5X memory. Too expensive and in limited quantity.
Same explanation about cooling.

Personally, despite the PCI-E power draw problem that is now effectively corrected, AMD RTG division has done a great job of delivering this product at the price point it has considering the performance.