AMD Radeon R9 300 Series MegaThread: FAQ and Resources

[renz496]

Just my speculation but the low end or even mid range AMD probably will rely on GF. The high end they probably go with TSMC. And it will be interesting to see if AMD will continue to use big die strategy for their gpu. Historically they prefer to have much smaller die with very tight transistor density. With Fiji they were force to go big or else they will have nothing to counter nvidia GM200 until their next gen arrive. But even with Fiji they barely maintain their matket share. This time around they probably not going to play the 'wait' game with nvidia but will be the one that strike first.

 

[Embra]

AMD's next-gen Polaris GPU architecture unveiled, arrives in mid-2016

http://www.tweaktown.com/articles/7496/amds-next-gen-polaris-gpu-architecture-unveiled-arrives-mid-2016/index.html

 

[jimmysmitty]

Do you think 14nm will give Polaris an edge over 16nm Pascal in efficiency? I'm asking that as a question so don't shun me.

Considering that the Apple A9 TSMC 16nm SoC has better power efficiency than their Samsung 14nm A9 SoC (slight) I am not sure there will be an advantage. I would think they are using it because it is probably cheaper to license it since Samsung and GloFlo are in a consortium to push for newer process nodes. It shows that not all processes are made equally.

As for the bandwidth, again I said that the bandwidth is not a bottleneck but the GPU itself is. This year we will see top end GPUs pushing probably 8GB of HBM VRAM with 1TB/s worth of memory bandwidth that will mostly go underutilized.

 

[Eggz]

Personally i just hope game developer optimize memory usage more. We should not need 6GB VRAM just to run ultra texture on 1080p

+15!

We have computers with so much system RAM that there should be an option or technology that can cache by priority. Often these high-VRAM games and will use only a fraction of the system RAM compared to the VRAM usage. Lower priority textures should be able to swap in and out to a certain extent without compromising performance.

 

[cdrkf]

Do you think 14nm will give Polaris an edge over 16nm Pascal in efficiency? I'm asking that as a question so don't shun me.

Considering that the Apple A9 TSMC 16nm SoC has better power efficiency than their Samsung 14nm A9 SoC (slight) I am not sure there will be an advantage. I would think they are using it because it is probably cheaper to license it since Samsung and GloFlo are in a consortium to push for newer process nodes. It shows that not all processes are made equally.

As for the bandwidth, again I said that the bandwidth is not a bottleneck but the GPU itself is. This year we will see top end GPUs pushing probably 8GB of HBM VRAM with 1TB/s worth of memory bandwidth that will mostly go underutilized.

I can't remember where but I'd read that the Samsung / GloFo process scaled differently to TSMC, and that the Glofo process was more efficient *at low clock frequencies* whereas TSMC scaled up better- this was in relation to phone soc's where the 'high frequency' parts were around 2.5ghz. If that's true the GloFo process might make a lot of sense given GPU's tend to sit in the 1 - 1.5ghz range...

 

[jimmysmitty]

Do you think 14nm will give Polaris an edge over 16nm Pascal in efficiency? I'm asking that as a question so don't shun me.

Considering that the Apple A9 TSMC 16nm SoC has better power efficiency than their Samsung 14nm A9 SoC (slight) I am not sure there will be an advantage. I would think they are using it because it is probably cheaper to license it since Samsung and GloFlo are in a consortium to push for newer process nodes. It shows that not all processes are made equally.

As for the bandwidth, again I said that the bandwidth is not a bottleneck but the GPU itself is. This year we will see top end GPUs pushing probably 8GB of HBM VRAM with 1TB/s worth of memory bandwidth that will mostly go underutilized.

I can't remember where but I'd read that the Samsung / GloFo process scaled differently to TSMC, and that the Glofo process was more efficient *at low clock frequencies* whereas TSMC scaled up better- this was in relation to phone soc's where the 'high frequency' parts were around 2.5ghz. If that's true the GloFo process might make a lot of sense given GPU's tend to sit in the 1 - 1.5ghz range...

True they are lower clock but they are higher power draw which might eliminate any advantage they have because remember these ARM chips not only run at that clock speed but also draw very low amounts of power, a couple of watts usually compares to a GPU that pulls something in the range of a couple hundred.

That said, the Samsung 14nm has two versions, 14NM LPE and LPP. The former was their first iteration (the e means "early" of 14nm and is suited for low power designs. The second is the LPP, which GloFlo is using, which is another low power design but is optimized for performance parts as well.

I still don't think there will be an advantage for them TBH but we will have to wait and see.

 

[cdrkf]

Do you think 14nm will give Polaris an edge over 16nm Pascal in efficiency? I'm asking that as a question so don't shun me.

Considering that the Apple A9 TSMC 16nm SoC has better power efficiency than their Samsung 14nm A9 SoC (slight) I am not sure there will be an advantage. I would think they are using it because it is probably cheaper to license it since Samsung and GloFlo are in a consortium to push for newer process nodes. It shows that not all processes are made equally.

As for the bandwidth, again I said that the bandwidth is not a bottleneck but the GPU itself is. This year we will see top end GPUs pushing probably 8GB of HBM VRAM with 1TB/s worth of memory bandwidth that will mostly go underutilized.

I can't remember where but I'd read that the Samsung / GloFo process scaled differently to TSMC, and that the Glofo process was more efficient *at low clock frequencies* whereas TSMC scaled up better- this was in relation to phone soc's where the 'high frequency' parts were around 2.5ghz. If that's true the GloFo process might make a lot of sense given GPU's tend to sit in the 1 - 1.5ghz range...

True they are lower clock but they are higher power draw which might eliminate any advantage they have because remember these ARM chips not only run at that clock speed but also draw very low amounts of power, a couple of watts usually compares to a GPU that pulls something in the range of a couple hundred.

That said, the Samsung 14nm has two versions, 14NM LPE and LPP. The former was their first iteration (the e means "early" of 14nm and is suited for low power designs. The second is the LPP, which GloFlo is using, which is another low power design but is optimized for performance parts as well.

I still don't think there will be an advantage for them TBH but we will have to wait and see.

Well it looks likely that for the larger GPUs AMD will be with TSMC still. I was just speculating that maybe with the smaller part, and at suitably low clock frequencies, the GloFo process might have an edge (where it wouldn't with a higher power 'enthusiast' class gpu)? I guess only way to tell will be to look at the small Polaris GPU on GloFo 14nm and compare perf / w to higher performance parts built on TSMC and see if there is a marked difference in efficiency (although that would require two parts using comparable memory, no point comparing a GDDR5 equipped gpu vs one with HBM)...

 

[prtskg]

http://www.amd.com/en-us/press-releases/Pages/amd-demonstrates-2016jan04.aspx
"AMD’s Polaris architecture-based 14nm FinFET GPUs deliver a remarkable generational jump in power efficiency."
May be AMD isn't using TSMC at all.
And it'd be better if people stop thinking LPE means Low power early and LPP means low power plus. In Samsung's own words -
"14LPE (Early edition) targets the early technology leaders and time-to-market customers such as mobile application SoCs to meet the latest mobile gadgets’ aggressive schedule and improved performance/power requirements.

14LPP (Performance boosted edition) is the 2nd FinFET generation which the performance is enhanced up to 10%. 14LPP is the single platform for every application designs with the improved performance for computing/Network designs and the lowered power consumption for Mobile/Consumer designs."
Samsung itself doesn't use terms like low power early and low power plus.

 

[renz496]

isn't that AMD already mention that they will use both GF and TSMC for their next gpu?

 

[con635]

isn't that AMD already mention that they will use both GF and TSMC for their next gpu?

I haven't seen that? What I have noticed is amd being very clear in saying '14nm' finfet?
Another thing I've noticed is the stick used to beat Sammy's 14nm process is the 14nm iPhone chips having higher leakage than the same chips on 16nm tsmc, hence higher power consumption, now I'm not an engineer just an enthusiast/gamer/oc'er but I always thought higher leakage meant higher clocks, no? Also if a process is better is better than another at low power mobile chips it would still be better at higher power DT chips?

 

[renz496]

the one they show at CES is made using GF 14nm fintet. but for the whole line up they will be using both TSMC 16nm and GF 14nm. some people asking if they are going to do it the same way as apple did with their A9 soc and amd answer is no. most likely the smaller (or even mid range chip) will be made by GF while the bigger one will be made by TSMC.

As for RTG’s FinFET manufacturing plans, the fact that RTG only mentions “FinFET” and not a specific FinFET process (e.g. TSMC 16nm) is intentional. The group has confirmed that they will be utilizing both traditional partner TSMC’s 16nm process and AMD fab spin-off (and Samsung licensee) GlobalFoundries’ 14nm process, making this the first time that AMD’s graphics group has used more than a single fab. To be clear here there’s no expectation that RTG will be dual-sourcing – having both fabs produce the same GPU – but rather the implication is that designs will be split between the two fabs. To that end we know that the small Polaris GPU that RTG previewed will be produced by GlobalFoundries on their 14nm process, meanwhile it remains to be seen how the rest of RTG’s Polaris GPUs will be split between the fabs.

http://www.anandtech.com/show/9886/amd-reveals-polaris-gpu-architecture/3

 

[de5_Roy]

AMD Confirms, Polaris Will Feature HBM2 And GDDR5 Memory Technologies
http://wccftech.com/amd-confirms-polaris-will-feature-hbm2-and-gddr5/
not sure if it's been posted already.

 

[renz496]

I thought they already said that during CES?

 

[renz496]

I know anyone ate getting hyped with polaris (or GCN 4th generation or what ever you want to call it). But is there people still interested about getting Fury X2 when they come out in Q2 this year?

 

[cdrkf]

I know anyone ate getting hyped with polaris (or GCN 4th generation or what ever you want to call it). But is there people still interested about getting Fury X2 when they come out in Q2 this year?

Well it'll be interesting (and very powerful). I think the main issue with it is it's still limited to 4gb of vram (per gpu) which is the big drawback of Fury for high end.

What I'm also interested in, they have only listed *two* Polaris gpus, one entry (aka HD 7000 series replacement + laptop part) and one 'Fury Level' enthusiast gpu. What about the mid range?

Maybe the original Fury + HBM 1 will be re-purposed as a mid tier gpu? I can't think they're going to be able to fill a full product line up with only 2 gpus. The only other thing is perhaps the 'enthusiast' gpu they have mentioned is actually more of a upper mid range (e.g. same configuration as Fiji but on 14nm would be half the size) and that they have a bigger gpu in the works for a future release?

 

[renz496]

The problem of Fury X2 will be obviously that 4GB VRAM....unless somehow they can fit the card with HBM2 for much higher capacity of VRAM. Though the way AMD already hyping for polaris it is not impossible if they end up ditch the plan to release Fury X2 when Q2 is here.

As for polaris i'd expect one low end and one high end (mid range turned into high end? Much like nvidia Gxx04 chip). They might keep Fury as their new R9 x90 line up. It would be funny to see if their new high end end up consume much less power than their slower card. But the most interesting to see was AMD actually use more and more similar strategy that nvidia did.

 

[cdrkf]

The problem of Fury X2 will be obviously that 4GB VRAM....unless somehow they can fit the card with HBM2 for much higher capacity of VRAM. Though the way AMD already hyping for polaris it is not impossible if they end up ditch the plan to release Fury X2 when Q2 is here.

As for polaris i'd expect one low end and one high end (mid range turned into high end? Much like nvidia Gxx04 chip). They might keep Fury as their new R9 x90 line up. It would be funny to see if their new high end end up consume much less power than their slower card. But the most interesting to see was AMD actually use more and more similar strategy that nvidia did.

Actually thinking about it, they are pushing Fury X2 as a VR solution. In that context the resolution (per eye) isn't that high (it's 900 x 1080p on the occulus?). That would mean the 4gb vram buffer is plenty, and that the high frame rates to sustain 90+fps is more important....

 

[renz496]

The thing is the time release of Fury X2. AMD mention Q2 this year. Then also said polaris will come out in time before back to school season. So which one will you're going to pick for VR? Fury X2 or wait for at least another quarter for polaris in CF setup? The later have potential to use HBM2 or even with GDDR5 still most likely going to have more VRAM than Fury X2. Fury X2 end up faster but with amd announcing their plans for polaris; even go as far as telling the time period when polaris should be on sale isn't that somehow going to affect Fury X2 sales? If polaris end up faster.....no even just 90 percent of Fury X performance some might actually favoring to get two polaris based card instead of buying Fury X2 because of 2 polaris will going to be less stressing on the PSU front. And we still not talking about new feature that only available on polaris. Of cource this is just my speculation but i could not help to think that the appeal of Fury X2 will become much lesser over time the closer we are to polaris release.

 

[jimmysmitty]

The biggest problem I see with Fury X2 is that by the time t comes out we will see Polaris and Arctic Islands and normally with a new generation the top end single GPU is as powerful (or close to) as two of the previous generations. For example a 980Ti is about as powerful as a 295X2. I would expect to see this especially with a die shrink that will give them more thermal headroom and die space for more SPUs/TMUs/ROPs etc.

Which would you rather buy? A single GPU or a power hunbgry dual GPU that might have bugs and issues in games yet both perform the same?

 

[turkey3_scratch]

According to WCCftech, they showed off two Polaris GPUs codenamed Polaris 11 and Polaris 10. They said one of them (I forget which) is the "successor to Fury X". So I think we're still going to see a Fury X2, and for all we know they may get 8GB VRAM in there.

I'd rather just get a Polaris though than a Fury X2. Or a Pascal. Nvidia's been very quiet about Pascal while AMD has been marketing Polaris out the wazoo. I don't see it as any indication of trouble for Nvidia, if anything they're being patient and letting AMD show off perhaps so they come in for the kill later?

 

[fudgecakes99]

Is fury x2 the successor to the fury or the name to that one 2 fury x's one pcb thing.

 

[turkey3_scratch]

Not sure, that wasn't specified.

 

[17seconds]

It's the usual pattern of pre-release marketing. AMD tends to over-hype, leading to release day disappointment. Nvidia tends to rely more upon leaks, and then often releases a GPU that exceeds expectations.

The classic example was the release of the 6970, which was highly anticipated, but was completely upstaged by an out of nowhere release of the GTX 580.
https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_580/1.html
https://www.techpowerup.com/reviews/HIS/Radeon_HD_6970/

 

[cdrkf]

The biggest problem I see with Fury X2 is that by the time t comes out we will see Polaris and Arctic Islands and normally with a new generation the top end single GPU is as powerful (or close to) as two of the previous generations. For example a 980Ti is about as powerful as a 295X2. I would expect to see this especially with a die shrink that will give them more thermal headroom and die space for more SPUs/TMUs/ROPs etc.

Which would you rather buy? A single GPU or a power hunbgry dual GPU that might have bugs and issues in games yet both perform the same?

Uh a 980TI *isn't* faster than a 295x2 unless you're talking specific cases where crossfire doesn't work, or games have a *heavy* NV bias (gamesworks titles for example). Where the 295X2 is working as intended it's still the fastest single card on the market. That's also it's weakness as it's hit and miss whereas a 980ti, whilst overall slower, is consistent and free of the issues of dual gpus.

 

[Eggz]

The biggest problem I see with Fury X2 is that by the time t comes out we will see Polaris and Arctic Islands and normally with a new generation the top end single GPU is as powerful (or close to) as two of the previous generations. For example a 980Ti is about as powerful as a 295X2. I would expect to see this especially with a die shrink that will give them more thermal headroom and die space for more SPUs/TMUs/ROPs etc.

Which would you rather buy? A single GPU or a power hunbgry dual GPU that might have bugs and issues in games yet both perform the same?

Uh a 980TI *isn't* faster than a 295x2 unless you're talking specific cases where crossfire doesn't work, or games have a *heavy* NV bias (gamesworks titles for example). Where the 295X2 is working as intended it's still the fastest single card on the market. That's also it's weakness as it's hit and miss whereas a 980ti, whilst overall slower, is consistent and free of the issues of dual gpus.

I don't think jimmysmitty ever said 980 ti is faster than 295x2. You quoted his post saying it's "about as powerful" as the 295x2. That's accurate. The 295x2 is faster, as you point out, but not by leaps and bounds. Take a look below. There's sometimes wider gaps between the two cards, but in Crysis 3 on low at 4K, the difference is only about 5 fps (click image for original article).