AMD RX 400 series (Polaris) MegaThread! FAQ & Resources

[JordanMihailov]

I can't read the whole thread because it's huuge ) Wan't ask: WHEN IS VEGA COMING ? I really want new video card and i am wandering between 1070, 1080 and the upcoming AMD card. I am waiting for this unveiling 6 months now )

1H 2017 if i'm not wrong.

Wasn't it supposed to be released in the end of the last year, after that january and now 1h 🙁( Endless waiting...

The rumor of AMD releasing at the end of last year was from how vague the chart AMD released about their releases. AMD had said multiple times last year that they were hoping for 1Q17, however, that was amended to 1H17.

There could be several reasons for this, but there is no indication as to the cause. It could be HBM2 shortage, design correction, production delays, or maybe just giving themselves time to get a good amount of stock to avoid the ridiculousness that occurred with the GTX 1080, 1070, 1060, RX 480 and 470.

personally i think AMD are waiting for HBM2 to be fully ready and cheaper to implement. but it seems HBM in general did not really go well? hynix has been updating their chart for HBM2.

https://videocardz.com/65649/sk-hynix-updates-memory-product-catalog-hbm2-available-in-q1-2017

So as you can see even Hynix plans can change quite often. The current plan is to release 1.6 GHz HBM2 stack with 4GB memory, and since we already know Vega has two stacks, this gives us 8GB and 409,6 GB/s bandwidth. That’s obviously lower than Fiji, in fact, it’s even lower than Pascal GP102 (480 GB/s), so it’s unclear if AMD is really going to use Hynix memory, or like NVIDIA use Samsung instead.

I really hope that HBM2 will be in the new AMD card and really really hope that they will be with performance no less than nvidia rivals

 

[renz496]

I can't read the whole thread because it's huuge ) Wan't ask: WHEN IS VEGA COMING ? I really want new video card and i am wandering between 1070, 1080 and the upcoming AMD card. I am waiting for this unveiling 6 months now )

1H 2017 if i'm not wrong.

Wasn't it supposed to be released in the end of the last year, after that january and now 1h 🙁( Endless waiting...

The rumor of AMD releasing at the end of last year was from how vague the chart AMD released about their releases. AMD had said multiple times last year that they were hoping for 1Q17, however, that was amended to 1H17.

There could be several reasons for this, but there is no indication as to the cause. It could be HBM2 shortage, design correction, production delays, or maybe just giving themselves time to get a good amount of stock to avoid the ridiculousness that occurred with the GTX 1080, 1070, 1060, RX 480 and 470.

personally i think AMD are waiting for HBM2 to be fully ready and cheaper to implement. but it seems HBM in general did not really go well? hynix has been updating their chart for HBM2.

https://videocardz.com/65649/sk-hynix-updates-memory-product-catalog-hbm2-available-in-q1-2017

So as you can see even Hynix plans can change quite often. The current plan is to release 1.6 GHz HBM2 stack with 4GB memory, and since we already know Vega has two stacks, this gives us 8GB and 409,6 GB/s bandwidth. That’s obviously lower than Fiji, in fact, it’s even lower than Pascal GP102 (480 GB/s), so it’s unclear if AMD is really going to use Hynix memory, or like NVIDIA use Samsung instead.

I really hope that HBM2 will be in the new AMD card and really really hope that they will be with performance no less than nvidia rivals

personally i did not care if the card use HBM2 or GDDR5/GDDR5X because in the end what matter is still the final performance. though for AMD they probably really need that HBM for power saving purpose because their architecture in general are less efficient than nvidia architecture.

 

[TehPenguin]

I think HBM2 is more relevant than just the power savings it makes. If overclocking memory gives a performance boost then a faster memory should make a difference as well.

 

[Math Geek]

the Fury X made it very clear what the much faster ram could do. that's not really up for debate. so even faster HBM2 is not really a question on "if" it's faster but "by how much".

that card handled stuff it should not have been able to with the amount of ram it had. but the much faster speeds allowed it to do much more than it should have been able to. may not have been the best card on the market but it was certainly a proof of concept for HBM.

 

[renz496]

one of the problem that i see with HBM is it's complexity. it seems HBM maker have a hard time to hit their target goal on time. offering bigger capacity also becoming very expensive due to more capacity = more complex process and higher failure yield.

 

[AndrewJacksonZA]

personally i did not care if the card use HBM2 or GDDR5/GDDR5X because in the end what matter is still the final performance.

This.

that card handled stuff it should not have been able to with the amount of ram it had. but the much faster speeds allowed it to do much more than it should have been able to.

How much of that is due to lightning VRAM speeds, sheer grunt from the GPU and drivers that are being continually optimised as evidenced over at [H]ardOCP in their recent driver performance over time test?

 

[Math Geek]

most all of it from what i saw is based on the super speed of the HBM. that's the point really. HBM2 will be even faster and allow for more vram as well. so why would you not expect it to be supremely faster than gddr5(x)?? because it will be on an amd gpu and nvidia does not have it yet?

what we saw from HBM on the fury X was pretty amazing for the 4 Gb it had on it. was easily hanging with 8+ GB cards from nvidia. as i said, proof of concept for what HBM can do. even paired with the not so great chip, the card did very well for itself. i fully expect HBM2 paired with a much more capable chip to be the stuff dreams are made of (ok maybe slight exaggeration for dramatic effect but you get the idea 😉 )

once nvidia unveils HBM in some form or another, then all of a sudden it'll be the greatest thing EVER i guess. we'll just have to wait until then for folks to finally see the (green??) light. red team gets no love 🙁

 

[RCFProd]

HBM2 will kill GDDR5X. Barely discussable.

 

[-Fran-]

Not in costs, no. I'll have to disagree there. Packaging HBM modules next to the GPU is still handled as an external operation and, from the general consensus I read everywhere, it's pretty darn hard and expensive to do.

HBM does have it's tangible and practical benefits, but from the "value" perspective, GDDR is still good and will continue to be with the next iterations of it.

Now, if the packaging part of the problem is "solved" (as in, gets cheaper and less complex), then it could be the same story as RDRAM and DDR-SDRAM.

Cheers!

 

[manleysteele]

most all of it from what i saw is based on the super speed of the HBM. that's the point really. HBM2 will be even faster and allow for more vram as well. so why would you not expect it to be supremely faster than gddr5(x)?? because it will be on an amd gpu and nvidia does not have it yet?

what we saw from HBM on the fury X was pretty amazing for the 4 Gb it had on it. was easily hanging with 8+ GB cards from nvidia. as i said, proof of concept for what HBM can do. even paired with the not so great chip, the card did very well for itself. i fully expect HBM2 paired with a much more capable chip to be the stuff dreams are made of (ok maybe slight exaggeration for dramatic effect but you get the idea 😉 )

once nvidia unveils HBM in some form or another, then all of a sudden it'll be the greatest thing EVER i guess. we'll just have to wait until then for folks to finally see the (green??) light. red team gets no love 🙁

The NVidia P-100 has HBM2. I don't know if it's the greatest thing ever, but NVidia wouldn't have gone to the expense of using it, if it didn't provide some easily perceived cost/performance advantage. As we say down heah, "That dawgs off the porch."

 

[renz496]

HBM2 will kill GDDR5X. Barely discussable.

lol it might be the other way around.

 

[renz496]

most all of it from what i saw is based on the super speed of the HBM. that's the point really. HBM2 will be even faster and allow for more vram as well. so why would you not expect it to be supremely faster than gddr5(x)?? because it will be on an amd gpu and nvidia does not have it yet?

what we saw from HBM on the fury X was pretty amazing for the 4 Gb it had on it. was easily hanging with 8+ GB cards from nvidia. as i said, proof of concept for what HBM can do. even paired with the not so great chip, the card did very well for itself. i fully expect HBM2 paired with a much more capable chip to be the stuff dreams are made of (ok maybe slight exaggeration for dramatic effect but you get the idea 😉 )

once nvidia unveils HBM in some form or another, then all of a sudden it'll be the greatest thing EVER i guess. we'll just have to wait until then for folks to finally see the (green??) light. red team gets no love 🙁

please don't bring red vs green into this. there is a valid concern to the future of HBM. HBM is a very complicated piece of tech. they easily superior to GDDR5/GDDR5X when it comes to bandwidth and power consumption. but at what cost? due to it's complexity HBM is very expensive and they are very hard to do right. the more stack you do the higher are the risk. this is one of the reason why the first HBM is limited to 4GB only. because they cannot guarantee adding more stack will work.

to be honest i'm not saying HBM is bad compared to GDDR5 tech. we know HBM can provide better bandwidth and have better power consumption as well. this is pure fact. but at the moment the cost is too high to implement. add to the fact that our current gpu are not really being held back by GDDR5/GDDR5X performance. so when it comes to financial perspective using GDDR5/GDDR5X is more logical decision to take.

 

[renz496]

Not in costs, no. I'll have to disagree there. Packaging HBM modules next to the GPU is still handled as an external operation and, from the general consensus I read everywhere, it's pretty darn hard and expensive to do.

HBM does have it's tangible and practical benefits, but from the "value" perspective, GDDR is still good and will continue to be with the next iterations of it.

Now, if the packaging part of the problem is "solved" (as in, gets cheaper and less complex), then it could be the same story as RDRAM and DDR-SDRAM.

Cheers!

this. +1

 

[renz496]

most all of it from what i saw is based on the super speed of the HBM. that's the point really. HBM2 will be even faster and allow for more vram as well. so why would you not expect it to be supremely faster than gddr5(x)?? because it will be on an amd gpu and nvidia does not have it yet?

what we saw from HBM on the fury X was pretty amazing for the 4 Gb it had on it. was easily hanging with 8+ GB cards from nvidia. as i said, proof of concept for what HBM can do. even paired with the not so great chip, the card did very well for itself. i fully expect HBM2 paired with a much more capable chip to be the stuff dreams are made of (ok maybe slight exaggeration for dramatic effect but you get the idea 😉 )

once nvidia unveils HBM in some form or another, then all of a sudden it'll be the greatest thing EVER i guess. we'll just have to wait until then for folks to finally see the (green??) light. red team gets no love 🙁

The NVidia P-100 has HBM2. I don't know if it's the greatest thing ever, but NVidia wouldn't have gone to the expense of using it, if it didn't provide some easily perceived cost/performance advantage. As we say down heah, "That dawgs off the porch."

nvidia use HBM2 on those tesla because those tesla will really benefit from that massive bandwidth. they will also will let nvidia to pack more performance because the power saving they get from HBM can be used by the gpu to increase performance. but nvidia definitely not picking HBM due to it's cost.

 

[Martell1977]

Have to remember that AMD has "Navi" up next after Vega and according to the road map:
http://www.extremetech.com/gaming/224773-amd-unveils-2016-gpu-roadmap-polaris-offers-2-5x-performance-per-watt-may-utilize-gddr5
It will bring "Scalability and Nextgen Memory".

AMD has discussed a little about the scalability, but it seems that there is another memory on the horizon and HBM2 will likely have as a short a life as HBM did. I've seen nothing to imply it could be HBM3 or something totally new. It could be more bandwidth or a refinement of HBM2 to make it less complex and more cost efficient. But, if it should offer more bandwidth, as discussed here, will the new GPU's really benefit from it?

Considering how little we know about Vega, even this close to launch, I have a feeling that info on Navi will be just as scarce until 1h2018. AMD seems very willing to test new memory systems. HBM was experimental/proof of concept, HBM2 is refined and capacity increased, next gen??? It's like the GPU memory version of Intel's Tic-Tock-Tweak cadence.

 

[manleysteele]

most all of it from what i saw is based on the super speed of the HBM. that's the point really. HBM2 will be even faster and allow for more vram as well. so why would you not expect it to be supremely faster than gddr5(x)?? because it will be on an amd gpu and nvidia does not have it yet?

what we saw from HBM on the fury X was pretty amazing for the 4 Gb it had on it. was easily hanging with 8+ GB cards from nvidia. as i said, proof of concept for what HBM can do. even paired with the not so great chip, the card did very well for itself. i fully expect HBM2 paired with a much more capable chip to be the stuff dreams are made of (ok maybe slight exaggeration for dramatic effect but you get the idea 😉 )

once nvidia unveils HBM in some form or another, then all of a sudden it'll be the greatest thing EVER i guess. we'll just have to wait until then for folks to finally see the (green??) light. red team gets no love 🙁

The NVidia P-100 has HBM2. I don't know if it's the greatest thing ever, but NVidia wouldn't have gone to the expense of using it, if it didn't provide some easily perceived cost/performance advantage. As we say down heah, "That dawgs off the porch."

nvidia use HBM2 on those tesla because those tesla will really benefit from that massive bandwidth. they will also will let nvidia to pack more performance because the power saving they get from HBM can be used by the gpu to increase performance. but nvidia definitely not picking HBM due to it's cost.

Agree. Plus NVidia has already announced that Volta will use GDDR6. No idea how the divisions will fall out at NVidia. It's yet another case of we'll have to wait and see.

 

[TehPenguin]

add to the fact that our current gpu are not really being held back by GDDR5/GDDR5X performance. so when it comes to financial perspective using GDDR5/GDDR5X is more logical decision to take.

Wrong. If it would be so there'd be no point in overclocking the memory but OCing your VRAM is a thing. It's also the reason why GDDR6 is in works.

 

[renz496]

add to the fact that our current gpu are not really being held back by GDDR5/GDDR5X performance. so when it comes to financial perspective using GDDR5/GDDR5X is more logical decision to take.

Wrong. If it would be so there'd be no point in overclocking the memory but OCing your VRAM is a thing. It's also the reason why GDDR6 is in works.

i said current gpu. it doesn't mean future gpu doesn't need it. just look at 1080 itself. it is faster than 980ti but it have much less bandwidth available to it. (320 Gb/s vs 336 Gb/s). and yet overclocking the memory did not more yield performance for the card.

https://www.youtube.com/watch?v=pbOS9XQUQXA

 

[TehPenguin]

i said current gpu. it doesn't mean future gpu doesn't need it. just look at 1080 itself. it is faster than 980ti but it have much less bandwidth available to it. (320 Gb/s vs 336 Gb/s). and yet overclocking the memory did not more yield performance for the card.

https://www.youtube.com/watch?v=pbOS9XQUQXA

Neither of the 3 games they tested are memory intensive.
It's the same with GPU OC, it doesn't boost every game by the same amount.

Take a look at this:
http://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-1080-and-1070-founders-edition-review/31

 

[renz496]

i said current gpu. it doesn't mean future gpu doesn't need it. just look at 1080 itself. it is faster than 980ti but it have much less bandwidth available to it. (320 Gb/s vs 336 Gb/s). and yet overclocking the memory did not more yield performance for the card.

https://www.youtube.com/watch?v=pbOS9XQUQXA

Neither of the 3 games they tested are memory intensive.
It's the same with GPU OC, it doesn't boost every game by the same amount.

Take a look at this:
http://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-1080-and-1070-founders-edition-review/31

anandtech test 5 games. and look from anandtech table OCing memory only add 3% more performance at best which is 1 fps increase in the witcher 3. if anything this prove GDDR5X did not holding back GTX1080 performance at all even when the gpu being overclocked.

 

[Robert Cook]

i said current gpu. it doesn't mean future gpu doesn't need it. just look at 1080 itself. it is faster than 980ti but it have much less bandwidth available to it. (320 Gb/s vs 336 Gb/s). and yet overclocking the memory did not more yield performance for the card.

https://www.youtube.com/watch?v=pbOS9XQUQXA

Neither of the 3 games they tested are memory intensive.
It's the same with GPU OC, it doesn't boost every game by the same amount.

Take a look at this:
http://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-1080-and-1070-founders-edition-review/31

anandtech test 5 games. and look from anandtech table OCing memory only add 3% more performance at best which is 1 fps increase in the witcher 3. if anything this prove GDDR5X did not holding back GTX1080 performance at all even when the gpu being overclocked.

To be fair Nvidia has a fair amount of memory compression, and I have always found myself gaining more from a memory overclock on AMD cards. Just because the 1080 does not use more bandwidth, does not mean more is not needed or more could be taken advantage of.

 

[TehPenguin]

You showed me a bench of 3 games before and it didn't bother you then ^^
Yes, 10% increase in memory speed does give as little as 3% in some games, but if you'd read the article you'd see that 12% increase in GPU clock only gives 4% more performance in some games which is also 1fps at 4k so what is your point?

 

[renz496]

You showed me a bench of 3 games before and it didn't bother you then ^^
Yes, 10% increase in memory speed does give as little as 3% in some games, but if you'd read the article you'd see that 12% increase in GPU clock only gives 4% more performance in some games which is also 1fps at 4k so what is your point?

i'm not really talking about performance increase with overclocking. my point is more about if current GDDR5/GDDR5X holding back our gpu performance or not. if the gpu are severely limited by it's bandwidth then overclocking the memory should uplift it's performance quite noticeably.

 

[renz496]

i said current gpu. it doesn't mean future gpu doesn't need it. just look at 1080 itself. it is faster than 980ti but it have much less bandwidth available to it. (320 Gb/s vs 336 Gb/s). and yet overclocking the memory did not more yield performance for the card.

https://www.youtube.com/watch?v=pbOS9XQUQXA

Neither of the 3 games they tested are memory intensive.
It's the same with GPU OC, it doesn't boost every game by the same amount.

Take a look at this:
http://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-1080-and-1070-founders-edition-review/31

anandtech test 5 games. and look from anandtech table OCing memory only add 3% more performance at best which is 1 fps increase in the witcher 3. if anything this prove GDDR5X did not holding back GTX1080 performance at all even when the gpu being overclocked.

To be fair Nvidia has a fair amount of memory compression, and I have always found myself gaining more from a memory overclock on AMD cards. Just because the 1080 does not use more bandwidth, does not mean more is not needed or more could be taken advantage of.

very true. with pascal nvidia also improve their color compression two times better than maxwell. but did AMD card will gain more performance if their gpu clock are running at stock clock?

 

[RCFProd]

Are there any rumored or confirmed clockspeeds on the non-reference RX480s?

They've been out for 6-7 months. You can find the clock speed for each non-reference RX 480 that has been released already. Or was it a typo and did you mean upcoming AMD VEGA graphics cards?