R9 390X Fiji XT - Extreme Memory Bandwidth & SPs

[geforceftw]

I wonder if the GPU will be a bottle neck. 4096 bit sounds like A LOT.

[dizmo]

Does anyone know how they plan to get past the limitations of PCIe?
Isn't this exactly why Nvidia was planning on introducing their own connector for Pascal? I thought PCIe didn't offer enough bandwidth to actually make use of the power the card has.

[Kukielka]

The entire R9 300 series is suppose to use HBM, so from the looks of it GDDR5 is done with on AMD based cards (good riddance).

 

Also keep in mind AMD isn't outsourcing their HBM, they developed it in partner with Hynix which gives them an advantage on low cost.

Protip: Stop arguing with him.

 

Does anyone know how they plan to get past the limitations of PCIe?

Isn't this exactly why Nvidia was planning on introducing their own connector for Pascal? I thought PCIe didn't offer enough bandwidth to actually make use of the power the card has.

Even PCIe 3.0 4x is far than enough for single core gpus.

[Opcode]

Does anyone know how they plan to get past the limitations of PCIe?

Isn't this exactly why Nvidia was planning on introducing their own connector for Pascal? I thought PCIe didn't offer enough bandwidth to actually make use of the power the card has.

PCIe doesn't necessarily have any limitations in general. People perceive Pascal as Nvidia pointing and saying PCIe has a problem when there's really no problem at all. Pascal is simply Nvidia's plan to get past the PCIe bandwidth limitations that will ultimately hinder their plans of unified memory in the long run. Keep in mind Nvidia is working to unify the GPU's memory with the CPU for easier programmable high end compute (essentially Nvidia's take on HSA except with a discrete GPU). PCIe is a limitation in that retrospect, tho as a general consumer who's going to be buying cards like these for primarily gaming PCIe isn't at fault. AMD has been working on unifying their discrete GPU's memory with the CPU as well ever since HSA was born. Tho these implementations are still far off and we won't have to worry about proprietary interfaces and all that garbage until 2016.

[Paragon_X]

Does anyone know how they plan to get past the limitations of PCIe?

Isn't this exactly why Nvidia was planning on introducing their own connector for Pascal? I thought PCIe didn't offer enough bandwidth to actually make use of the power the card has.

I am not into architecture and stuff but i think the Bandwidth we refer to here is between the VRAM and GPU,(Just like the ram and the CPU) so you can say "its inside the card". the bandwidth of PCI 3.0 x16 is enough for the communication from the card to the rest of the system and vice versa(I think its 16GB/s). If anyone is sure about this please provide info.

[MrSuperb]

just to shorten this pathetic conversation: i still didnt lose any word about the 390x and the reason why i click on people s profile is because i dont like having arguments with kids, but suprisingly this isnt any different in your case.. and this you can take personal dude.

Alright, so you didn't talk about the 390X, then I suppose I didn't talk about you either.

But you know what, this is directed to you: You might want to take a look in a mirror someday, because most of the stuff you right about other people, seems to be a perfect description of yourself.

[patrickjp93]

The entire R9 300 series is suppose to use HBM, so from the looks of it GDDR5 is done with on AMD based cards (good riddance).

 

Also keep in mind AMD isn't outsourcing their HBM, they developed it in partner with Hynix which gives them an advantage on low cost.

So far all I've heard using it are the 390/X. Also, I predict (damn autocorrect) a price jump for the AMD platform as well.

 

They didn't develop it with Hynix. Hynix developed it themselves. AMD just hopped on for the ride, the same way Nvidia hopped on Hybrid Memory Cube from Micron before Hynix's research looked more promising in the immediate term.

[bogus]

So far all I've heard using it are the 390/X. Also, I redact a price jump for the AMD platform as well.

They didn't develop it with Hynix. Hynix developed it themselves. AMD just hopped on for the ride, the same way Nvidia hopped on Hybrid Memory Cube from Micron before Hynix's research looked more promising in the immediate term.

That's quite a bold claim. Can you provide the source of this information?

[patrickjp93]

That's quite a bold claim. Can you provide the source of this information?

Sheer logic. Not only was the 20nm process delayed, it had low yields as well, and it's still at barely acceptable yields. Not to mention the R&D costs of HBM which Hynix will inevitably pass along to AMD as the first purchaser, who will have to pass at least some of it along to the consumer.

 

If you don't see a price jump coming, I question your sanity. Alternatively AMD could just drop the never settle bundle and reap pure profit.

[bogus]

Sheer logic. Not only was the 20nm process delayed, it had low yields as well, and it's still at barely acceptable yields. Not to mention the R&D costs of HBM which Hynix will inevitably pass along to AMD as the first purchaser, who will have to pass at least some of it along to the consumer.

 

If you don't see a price jump coming, I question your sanity. Alternatively AMD could just drop the never settle bundle and reap pure profit.

I'm talking about you claiming that AMD only hoped in the development of HBM like they had a irrelevant part in the development (in fact no development at all, according to you). I assume it's the same for GDDR?

The price and process is pure speculation so far... 

[patrickjp93]

I'm talking about you claiming that AMD only hoped in the development of HBM like they had a irrelevant part in the development (in fact no development at all, according to you). I assume it's the same for GDDR?

The price and process is pure speculation so far...

Price is speculation, but you can be guaranteed at the outset it's higher than GDDR. Millions if not $1 billion + in development costs will show up in the costs for early adopters, just like any other product.

As per AMD jumping in towards the end, this is widely known except by diehard AMD fans.

For instance, Hynix already was shipping stacked DRAM and FPGAs in professional and server devices. All AMD did was say GPUs needed some loving. Most of the work had already been done: http://wccftech.com/amd-working-hynix-development-highbandwidth-3d-stacked-memory/

I chose WCCF because it actually presents the whole truth, not just AMD's announcement of "co-development". It's an article sourced from a number of reputable media outlets, but the whole picture had not been put together by anyone else at the time.

[bogus]

As per AMD jumping in towards the end, this is widely known except by diehard AMD fans.

For instance, Hynix already was shipping stacked DRAM and FPGAs in professional and server devices. All AMD did was say GPUs needed some loving. Most of the work had already been done: http://wccftech.com/amd-working-hynix-development-highbandwidth-3d-stacked-memory/

I chose WCCF because it actually presents the whole truth, not just AMD's announcement of "co-development". It's an article sourced from a number of reputable media outlets, but the whole picture had not been put together by anyone else at the time.

From what I understand from the articles, die stacking was only being use in FPGAs and image sensors - from that to HBM for GPUs and APUs, it's more then saying "GPUs need some loving". That's why in WCCF tech and the articles you mention that AMD and Hynix developed HBM.

You are the only one claiming AMD did nothing. Even your sources contradict your claims.

[tedjani]

[patrickjp93]

From what I understand from the articles, die stacking was only being use in FPGAs and image sensors - from that to HBM for GPUs and APUs, it's more then saying "GPUs need some loving". That's why in WCCF tech and the articles you mention that AMD and Hynix developed HBM.

You are the only one claiming AMD did nothing. Even your sources contradict your claims.

Read it again. Stacked memory was developed for those sensors because if you want a 60 Megapixel photo, you need to move all that data extremely fast over very high bandwidth at low power. The bulk of the work had already been done.

AMD only came in and asked it be brought to GPus as well. The difficult engineering work (through-silicon Vias being used in memory and computational chips) had already been done. The only thing new was getting a set of interconnects for a GPU PCB and rearranging a reference card for multi-channel RAM, which isn't much work.

[The_Evenger]

it's going to be a repeat of last time, they release the 390x, and NVidia will release the 980ti.

[Briggsy]

I am not into architecture and stuff but i think the Bandwidth we refer to here is between the VRAM and GPU,(Just like the ram and the CPU) so you can say "its inside the card". the bandwidth of PCI 3.0 x16 is enough for the communication from the card to the rest of the system and vice versa(I think its 16GB/s). If anyone is sure about this please provide info.

 

you are correct, the bandwidth of GPU memory is internal, between the gpu and the vram. the more cores/processors a gpu has, the greater the bandwidth needed (i believe, compression helps with that too). the pci-e lane is to transfer game assets that are taken from storage, placed into system ram, and them transferred to the graphic card. the communication between the gpu and cpu to regulate timing etc. along the pci-e lanes doesn't require a lot of bandwidth, just the game assets that the gpu needs for rendering. if anyone knows more, they can correct me.

[PortentousLad]

you just judge on things I didn't say, not good. judging things u don't understand in general, very bad. also I don't care what you study, incase you think this "personal route" was any serious. posting silly things like this shows to me how much perusal you made in your studies.

what i am interested in, is the 390. @PortentousLad ,on the other hand, seems to get my point.

 

Of course I understand, I'm very smart you know. 

See I even have my egg-headphones on, no wait that's Linus. Although to be fair I do look a bit like a tall, bearded version of Linus oh and with glasses.

 

cooles Profilbild BTW

[Opcode]

So far all I've heard using it are the 390/X. Also, I predict (damn autocorrect) a price jump for the AMD platform as well.

 

They didn't develop it with Hynix. Hynix developed it themselves. AMD just hopped on for the ride, the same way Nvidia hopped on Hybrid Memory Cube from Micron before Hynix's research looked more promising in the immediate term.

There's a possibility that lower end cards will still use GDDR5 but seeing how Volcanic Islands is already setup to use HBM I don't see why AMD wouldn't just use HBM. They can stack 2 GB of HBM 2048-bit @ 360 GB/s onto the card instead. Or even 3 GB of HBM 3072-bit @ 480 GB/s. So HBM even on the flagship mainstream cards still seem like a viable option. Sure AMD won't be able to squeeze high densities out of the cards this time around but the second generation of HBM will change that.

 

AMD did indeed develop it in partner with Hynix. AMD went to Hynix and said "we need a faster memory solution" and Hynix said "sure we can make that happen, just give us the money". This is why Hynix branded HBM will more than likely be exclusive to just AMD based hardware. Otherwise Nvidia would more than likely be trying to outsource it which I doubt AMD wants to happen.

[patrickjp93]

There's a possibility that lower end cards will still use GDDR5 but seeing how Volcanic Islands is already setup to use HBM I don't see why AMD wouldn't just use HBM. They can stack 2 GB of HBM 2048-bit @ 360 GB/s onto the card instead. Or even 3 GB of HBM 3072-bit @ 480 GB/s. So HBM even on the flagship mainstream cards still seem like a viable option. Sure AMD won't be able to squeeze high densities out of the cards this time around but the second generation of HBM will change that.

AMD did indeed develop it in partner with Hynix. AMD went to Hynix and said "we need a faster memory solution" and Hynix said "sure we can make that happen, just give us the money". This is why Hynix branded HBM will more than likely be exclusive to just AMD based hardware. Otherwise Nvidia would more than likely be trying to outsource it which I doubt AMD wants to happen.

You do know Nvidia's Pascal architecture is coming with HBM onboard, right? Nvidia originally was going to use the Hybrid Memory Cube, but research progressed too slowly. HBM is not AMD-exclusive by any means. AMD is just the first adopter for GPU makers.

[Opcode]

You do know Nvidia's Pascal architecture is coming with HBM onboard, right? Nvidia originally was going to use the Hybrid Memory Cube, but research progressed too slowly. HBM is not AMD-exclusive by any means. AMD is just the first adopter for GPU makers.

Pascal isn't until 2016 so quite irrelevant at this time (has some bad concepts imo, but we won't go into discussing them). No one said HBM is an AMD exclusive, you do understand there is a JEDEC standard for HBM right? If you're going to troll do a better job at it as I am still correcting you every step of the way.

[patrickjp93]

Pascal isn't until 2016 so quite irrelevant at this time (has some bad concepts imo, but we won't go into discussing them). No one said HBM is an AMD exclusive, you do understand there is a JEDEC standard for HBM right? If you're going to troll do a better job at it as I am still correcting you every step of the way.

"This is why Hynix brand HBM will more than likely be exclusive to AMD hardware." Get on your knees and kiss my feet. Most blatant lie of the day Monsignor.

Also, speculation on the launch of Pascal.

[Opcode]

"This is why Hynix brand HBM will more than likely be exclusive to AMD hardware." Get on your knees and kiss my feet. Most blatant lie of the day Monsignor.

Also, speculation on the launch of Pascal.

Someone clearly needs to either A ) learn how to read, B ) learn English, C ) both A and B.

 

I'm going to stop responding to your posts, you're clearly out to troll and nothing more. You've lost all credibility to most of the forum members anyways (including me).

 

ProTip: If you wanna be a "professional learner" don't automatically assume you know everything.

[patrickjp93]

Someone clearly needs to either A ) learn how to read, B ) learn English, C ) both A and B.

I'm going to stop responding to your posts, you're clearly out to troll and nothing more. You've lost all credibility to most of the forum members anyways.

If you have any left I question humanity's ability to evolve and think rationally. I also have As in every English class I've taken at the university level, all 4 of them from the intro sequence everyone has to take, to technical writing, to legalese. What you said in court would amount to declaring Hynix's own patented technology would be exclusive to AMD, which is demonstrably false.

[Opcode]

If you have any left I question humanity's ability to evolve and think rationally. I also have As in every English class I've taken at the university level, all 4 of them from the intro sequence everyone has to take, to technical writing, to legalese. What you said in court would amount to declaring Hynix's own patented technology would be exclusive to AMD, which is demonstrably false.

Thanks for validating my statement.

1. Extremely poor English.
2. No one even talked about patenting.
3. You're wrong 99% of the time.

You should honestly stick to posting less, because it seems each time that you do you tend to make a fool out of yourself all the more.

[patrickjp93]

Thanks for validating my statement.

• Extremely poor English.
• No one even talked about patenting.
• You're wrong 99% of the time.

You should honestly stick to posting less, because it seems each time that you do you tend to make a fool out of yourself all the more.

There is exactly one punctuation flaw in the opening sentence. The rest is perfect. I put work into my phrasing. It's a more archaic style, but it's completely valid.

You can't have an exclusive tech without a patent, and since Hynix does have parents on HBM, it stands to reason Nvidia can't just outsource it. You're wrong four times for every one I am, as this thread demonstrates.