[BitsAndChips]390X ready for launch - AMD ironing out drivers - Computex launch

[crisium]

I think what Subyman was alluding to was that if AMD is able to make a dual die GPU interconnect function the same as a single die (i.e. without all the compromises of an XDMA/SLI interconnect), it would be huge. AMD would then be able to scale GPUs out much easier and we would see performance gains like we used to see each new generation.

I'm 99.9% sure AMD doesn't have that capability yet but I can dream.

I dream of this too. Imagine a 295X2 that actually always performed as a 5632 Shader and 128 ROPs card, and not just when the drivers were up to par.

 

[boozzer]

if amd can make this work, can it be use for cpu scalling also? no more quads or 8 cores, just gets treated as 1 huge, monster of a single core.

 

[ShintaiDK]

if amd can make this work, can it be use for cpu scalling also? no more quads or 8 cores, just gets treated as 1 huge, monster of a single core.

No.

 

[destrekor]

I think what Subyman was alluding to was that if AMD is able to make a dual die GPU interconnect function the same as a single die (i.e. without all the compromises of an XDMA/SLI interconnect), it would be huge. AMD would then be able to scale GPUs out much easier and we would see performance gains like we used to see each new generation.

I'm 99.9% sure AMD doesn't have that capability yet but I can dream.

What must be considered is that dual-die and dual-GPU as we know it are very much different. You won't see a dual-die package that has the same capability of a dual-GPU card, not if you are expecting a dual-die 295X2. The size of the GPU package would be enormous, and I don't think I've ever seen a processor package much larger than what can be found now, surely not something capable of hosting two dies the size of Hawaii XT. I don't know if there is an engineering limit to the package size and density of the BGA pinout, but I reckon there must be some reason we haven't seen monstrous packages.

And the heat of two monstrous workout dies on the same package would be extreme. At least with space between two separate GPU packages the cooler can be configured to deal with that heat, but if it's all on a single package, the cooler engineering marvels must be phenomenal to keep up.

Now, multi-die packages are sort of making a come back, but in a different way. I think AMD will capitalize in a major way by using Intel's multi-package route for mobile, by incorporating distinct CPU and GPU dies for APU packages. This plus integrated HBM could make for a terrific APU.

 

[S.H.O.D.A.N.]

I don't know if there is an engineering limit to the package size and density of the BGA pinout, but I reckon there must be some reason we haven't seen monstrous packages.

The reason is we have not see a single gpu that would make use of a package like that.

If AMD found the holy grail of packaging multiple gpus as a single entity, it would be bigger than HMB. Heck, that would be the biggest paradigm shift in design since the inception of modern gpu.

 

[xthetenth]

It'd be just in time too with the yield numbers that are getting mentioned.

 

[Kenmitch]

The reason is we have not see a single gpu that would make use of a package like that.

If AMD found the holy grail of packaging multiple gpus as a single entity, it would be bigger than HMB. Heck, that would be the biggest paradigm shift in design since the inception of modern gpu.

Would be nice! Somewhat of a kick in the balls to gameworks at the same time....Temporarily at least till some mysterious code is implemented.

 

[monstercameron]

I think what Subyman was alluding to was that if AMD is able to make a dual die GPU interconnect function the same as a single die (i.e. without all the compromises of an XDMA/SLI interconnect), it would be huge. AMD would then be able to scale GPUs out much easier and we would see performance gains like we used to see each new generation.

I'm 99.9% sure AMD doesn't have that capability yet but I can dream.

What about the interposer that they use to connect hbm to the dice? High bandwidth low latency.

 

[RussianSensation]

Dual Fiji XT around launch? Nope. And it seems ludicrous to name the dual-GPU configuration Fiji XT (or a dual-die, which again, ridiculous) when convention has seen that to simply represent the larger die size.

Fudzilla changed its tone, and now they claim there are two different Fijis, Fiji XT - a single-GPU for high-end gaming rigs, and Fiji VR - a dual-GPU designed for VR headsets. Naming conventions aside, let's look if it's even possible to fit 2x 550mm2 die on a single board. The answer is YES, rather easily.

1) NV fit 2x561mm2 28nm die on a single board with the Titan Z.

2) If you look at AMD's R9 295X2 design, there is a lot of extra space between VRMs/power circuitry which creates an opportunity for a more efficient design.

Asus' engineers have proved that if you start from scratch, you can design an R9 295X2 board that has WAY more power power/VRM circuitry and yet still manages to fit 3x8-pin connectors, in a more compact PCB. Impressive!

From an engineering point of view, nothing is stopping AMD from putting 2x500mm2+ 28nm GPUs with 250-300W of power usage each on a single PCB.

3) What are the chances we will see R9 390/390X and an R9 395X2? Extremely likely. AMD had 4870X2, 5970, 6990, 7990 and R9 295X2 = every single generation AMD puts its top chip in a dual configuration. With AIO CLC, it should not be too difficult to keep a card with 2x300W GPUs cool and quiet. Worst case, they can always go 140mm radiator.

The naming leaks might be off but I don't have a lot of doubts that we will eventually see a dual-Fiji XT card to succeed R9 295X2.

Side-note: With AMD now announcing that they only expect next gen products by 2H of 2015, I think this strongly calls into question the re-branding theory. If all you are doing is re-branding R9 200 series, and AMD reduced shipments to OEMs starting as early as Q3 2014, why would AMD need another quarter+ to launch "re-brands"? Launching re-brands is the easiest thing to do and yet AMD has nothing to replace R9 270/270X/280/280X/285/290/290X.

I am not sure if AMD will manage to launch R9 395X2 (Bermuda XT?) at the same time as R9 390/390X Fiji XT cards, but I presume it will show up in the 6 months following their launch.

I think another reason AMD needs more time is to really make sure they have 8GB HBM version out.

"It looks like GeForce GTX 980 SLI is bottlenecking hard here with the limited framebuffer. GeForce GTX TITAN X is able to pull ahead in performance with much more consistent framerates. Again, the smoothest experience is GeForce GTX TITAN X.

What we have learned from these tests are that 4GB video cards are going to suffer at the highest in-game settings in this game due to the limited framebuffers. We are going to have to adjust settings to alleviate that bottleneck for these to truly shine. In our full-evaluation we will find out what settings are actually playable once we figure out which setting reduces the VRAM bottleneck on them. As it is right now, GTX TITAN X takes the cake as the best experience."

Looks like at 4K, 980 SLI and R9 295X2 really suffer due to lack of VRAM.

---

R7 360X accidental leak? Shows 2GB of GDDR5. I am not surprised as I don't expect HBM to make it to any cards besides R9 390, 390X and 395X2. It's too expensive and way too fast for some low-end card like the 360X to make sense in the current marketplace.

AMD's inventory clearance is in full force. PowerColor Devil 13 R9 290 XII is $600. Just a good reminder to all how quickly GPUs depreciate in value (R9 295X2 came out at $1,499 on April 8, 2014). An air-cooled 500W card is possible. At 2 ½” high, 12” long and weighing in at a sumo-like 5.2lbs.

 

[HurleyBird]

Asus' engineers have proved that if you start from scratch, you can design an R9 295X2 board that has WAY more power power/VRM circuitry and yet still manages to fit 3x8-pin connectors, in a more compact PCB. Impressive!

That PCB is very tall, not exactly smaller than the reference PCB in terms of surface area.

 

[Silverforce11]

I think another reason AMD needs more time is to really make sure they have 8GB HBM version out.

What we have learned from these tests are that 4GB video cards are going to suffer at the highest in-game settings in this game due to the limited framebuffers.

With the useful DSR/VSR, gamers are not limited to having 4K monitors to experience super sampling image quality. 1080/1440p with SS is the best form of AA. Thus, more than 4GB for the next flagship is a must.

If AMD launches the 390X only in 4GB config, they would be completely ridiculed and stomped on, not just from NV marketing, but from gamers themselves. We've seen many games in recent times that push 4GB, its a trend that's going to expand.

 

[3DVagabond]

With the useful DSR/VSR, gamers are not limited to having 4K monitors to experience super sampling image quality. 1080/1440p with SS is the best form of AA. Thus, more than 4GB for the next flagship is a must.

If AMD launches the 390X only in 4GB config, they would be completely ridiculed and stomped on, not just from NV marketing, but from gamers themselves. We've seen many games in recent times that push 4GB, its a trend that's going to expand.

Yes, they will be ridiculed and stomped on by the same people who drool all over the 970. Where are all of our 770 2gig cards are all you are going to need for years people now? They are the same ones who say 3.5gig 970's are OK.

Silverforce11, it's all marketing BS. People just give it too much credit.

I agree that AMD needs to make the 390X WCE an 8gig card. They can charge a premium for it. They also need to make a 4gig model (and a 4gig 390) though that is targeted at perf/$. That card will be the perfect card for 1440. They should not release a 8gig 390. Not right away. Make people who want the 8gig buy the top of the line premium model. When competition dictates, then release 8gig 390's.

 

[Silverforce11]

When competition dictates, then release 8gig 390's.

Competition already dictated, months ago. They need to release first with the 390X 8GB version, at the same time or not long after, the 390X 4GB and 390 4GB variants.

The first card is going to be the one used as "reference" on most review sites, it needs to be good else the stigma remains for the entire generation.

ps. I would NOT buy a card with Titan X or more in performance if it had 4GB vram.

 

[3DVagabond]

Competition already dictated, months ago. They need to release first with the 390X 8GB version, at the same time or not long after, the 390X 4GB and 390 4GB variants.

The first card is going to be the one used as "reference" on most review sites, it needs to be good else the stigma remains for the entire generation.

ps. I would NOT buy a card with Titan X or more in performance if it had 4GB vram.

I said 390 not 390X. Make people buy the 390X to get 8gigs

 

[Silverforce11]

Yes, they will be ridiculed and stomped on by the same people who drool all over the 970. Where are all of our 770 2gig cards are all you are going to need for years people now? They are the same ones who say 3.5gig 970's are OK.

They were wrong, obviously.

Also I agree, 8GB 390X should be there to command a premium.

 

[dacostafilipe]

AMD just released their AMDGPU open source linux driver.

http://www.phoronix.com/scan.php?page=news_item&px=AMD-AMDGPU-Released

There's an iceland_dpm file that mentions "Topaz", and there's a tonga_dpm file. This could indicate that Topaz is not a Tonga rebrand as it would not require a new dpm for the same architecture version. "Iceland", "Carrizo" and "Tonga" are all from the VI (Volcanic Islands) family.

Also, Topaz has a lower device ID then Tonga.

 

[ShintaiDK]

AMD just released their AMDGPU open source linux driver.

http://www.phoronix.com/scan.php?page=news_item&px=AMD-AMDGPU-Released

There's an iceland_dpm file that mentions "Topaz", and there's a tonga_dpm file. This could indicate that Topaz is not a Tonga rebrand as it would not require a new dpm for the same architecture version. "Iceland", "Carrizo" and "Tonga" are all from the VI (Volcanic Islands) family.

Also, Topaz has a lower device ID then Tonga.

Topaz?

http://www.techpowerup.com/gpudb/2572/radeon-r7-m265.html

While the desktop versions are code-named "Sea Islands", the mobile GPUs of the AMD Radeon Rx 200 Series are code-named "Crystal Series": "Vesuvius", "Topaz" and "Amethyst".

Nothing new.

 

[dacostafilipe]

Topaz?

http://www.techpowerup.com/gpudb/2572/radeon-r7-m265.html

Yes, Topaz: http://cgit.freedesktop.org/~agd5f/...u&id=6377c9da05427ef167c74a0f613f4b2bd368ffb8

...
+#include "amdgpu.h"
+#include "iceland_smumgr.h"
+
+MODULE_FIRMWARE("radeon/topaz_smc.bin");
+
+static void iceland_dpm_set_funcs(struct amdgpu_device *adev);
...

 

[dacostafilipe]

While the desktop versions are code-named "Sea Islands", the mobile GPUs of the AMD Radeon Rx 200 Series are code-named "Crystal Series": "Vesuvius", "Topaz" and "Amethyst".

Nothing new.

Yeah, I've also found that page at the beginning, but it makes no sense to me.

It's to similar to Opal except clocks and memory an released just 2 month after. Seems like an TPU error.

Once thing is certain, we will see some new Topaz GPUs in the future.

 

[Cloudfire777]

This is pretty interesting. Tweaktown writes:

we've been hearing some rumblings from within the industry. Our sources have said that AMD will have two Radeon R9 390X cards to launch: one of them will be the Radeon R9 395X2 (a dual-GPU version) with 8GB of VRAM. The second card will be a 4GB version that won't beat the GeForce GTX Titan X.

Our industry sources also tell us that there are some hurdles with yields on HBM, which will see a very limited supply of Radeon R9 390X and Radeon R9 395X2 cards right through to the end of the year.

 

[AtenRa]

Asus' engineers have proved that if you start from scratch, you can design an R9 295X2 board that has WAY more power power/VRM circuitry and yet still manages to fit 3x8-pin connectors, in a more compact PCB. Impressive!

Actually the PCB area is almost the same as the reference 295, but the ASUS custom design PCB its smaller in length but bigger in height.

 

[Head1985]

This is pretty interesting. Tweaktown writes:

I hope this is not true.
If they need dual card to beat titanX this is just wrong.Not after they are so late.
Like really this looks like worst case scenario for AMD.
They have worst financial results in history and lost most market share and on top of that they have problems with HBM and have very limited numbers to the end of they year?
If rest GPUs are rebrands... i dont want even think about that.

 

[Keysplayr]

So expectations of an 8GB single GPU 390X are waning then? Somebody said 8GB for 390X was "confirmed". I guess it was, but for the Dual GPU 390X (395).

 

[showb1z]

Imagine what kind of rumors they will come up with as the wait for the 390 goes on!

 

[Silverforce11]

Imagine what kind of rumors they will come up as the wait for the 390 goes on!

People who are still waiting would be those not willing to folk $999 for Titan X.. so as soon as NV launches 980ti with 6GB vram for around $699, not many would remain waiting.