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

[Despoiler]

I dont think you know this for sure. There will be nothing in the drivers that explains to the GPU on what to do with memory in this and that?
How does games improve with newer drivers then? Hardware does it all by itself according to you...
Why did Nvidia say they could improve GTX 970 VRAM issue with new drivers?

He is wrong. It's a well known fact that the performance of GCN was held back because the drivers were still using memory management built for the last architecture. There is very much is a component of the drivers that deals with memory.

 

[dacostafilipe]

A driver is a translation layer to the hardware. Nothing more.

As far as I know, nobody uses asm to talk to the GPUs ... so there's still a lot of optimisation potential on the driver with new hardware.

 

[hawtdawg]

my biggest fear is that the first 390's will only have 4 gigs of ram, and the 8 gig versions won't appear until the holidays. Whether or not there are 8 gig versions at launch will decide whether or not I get Titan's or 390x's

 

[ShintaiDK]

Wouldn't actually using HBM's bandwidth in existing game titles involve driver tweaking? I mean AMD and Nvidia have been known to completely rewrite shader models and the like for AAA titles in their drivers and similar code altering stuff.

You could optimize some parts if there was more bandwidth. But if its SPs goes from 2816 to 4096 and memory goes from 320 to 512GB/sec. You are pretty much off where you started.

And you be perfectly fine without. It would work as always. Something you could always add later. Certainly not something you would let your competitor beat you senseless over, while you wait countless months for...getting a bigger mammoth uphill challenge?

 

[Cloudfire777]

He is wrong. It's a well known fact that the performance of GCN was held back because the drivers were still using memory management built for the last architecture. There is very much is a component of the drivers that deals with memory.

Of course Shintai was wrong. Here is different code in the Nvidia drivers that deal with memory management. The same applies to AMD and their drivers as well.

http://developer.download.nvidia.com/compute/cuda/4_1/rel/toolkit/docs/online/group__CUDA__MEM.html

And I bet a whole deal of that code changes when a different type of memory like HBM is used. Which is why AMD needs a little more time with the drivers

 

[Vesku]

True or not the idea of AMD furiously working on 300 series drivers is a more pleasant thought for why nothing since Omega than the possibility half of their software people have been poached by Nvidia, Mediatek, and the like. Not many other reasons in between for them to be humming along improving their drivers through November last year and then not even have a public ready beta driver for their Freesync launch in mid March.

 

[ShintaiDK]

Of course Shintai was wrong. Here is different code in the Nvidia drivers that deal with memory management. The same applies to AMD and their drivers as well.

http://developer.download.nvidia.com/compute/cuda/4_1/rel/toolkit/docs/online/group__CUDA__MEM.html

And I bet a whole deal of that code changes when a different type of memory like HBM is used. Which is why AMD needs a little more time with the drivers

Windows also got memory management. But who do you think controls the memory? The CPU or Windows? For Windows memory is memory. Same with CUDA or anything else. It doesnt matter if its EDO, FP, SDR, DDR, GDDR, HBM, HMC, Wide I/O or whatever thats behind. Because the driver/system never operates with that. It just operates with memory addresses.

Your CUDA link says the exact same.

 

[senseamp]

Sounds like it's not ready to me. Driver being ready is a prerequisite for the GPU being ready.

 

[ginfest]

Sounds like it's not ready to me. Driver being ready is a prerequisite for the GPU being ready.

Yes, I agree. With the release of the TitanX adding to an already dominant product line, the AMD hype is in full speed ahead mode. Most of the stuff being released to AMD friendly websites is just propaganda at this point until the product is released and subject to real world testing by reputable sites.

 

[dacostafilipe]

Windows also got memory management. But who do you think controls the memory? The CPU or Windows? For Windows memory is memory. Same with CUDA or anything else. It doesnt matter if its EDO, FP, SDR, DDR, GDDR, HBM, HMC, Wide I/O or whatever thats behind. Because the driver/system never operates with that. It just operates with memory addresses.

VI for example supports a new type of "flat" memory access. (bypasses cache if I remember well)

This should only be used in specific cases and it is the drivers job to decide when todo so or not by reading the performance counters/registers.

So, yes, even drivers can improve memory access ...

 

[Paul98]

Both in the driver and in game code you can optimize by how you access memory, with the architecture change there very well may be plenty of areas where this can be optimized.

 

[Headfoot]

If the hardware's ready its probably a combination of additional software polish, and building stockpiles so they don't instantly sell out everywhere

 

[garagisti]

I dont see what HBM and drivers have in common. For the drivers its just plain VRAM.

390X ready for a long time and just waiting for drivers is a big pile of BS.

Actually no.

There are some games which don't yet have profiles for CF, thanks to some legal issues between developers and Nvidia (NDA, whereby they can't share stuff). May be, just may be AMD is trying to get CF and other things working for Game'barely'Works, and that's just it. Remember, [H] was actually beating down on AMD when a patch which should have come from Ubi never came till very late. I think this is only good, take away the stick with which reviewers can beat down on your card.

@Raghu
Should have scrolled down before i typed away.

 

[Stormflux]

AMD can't shake the "poor drivers" stigma they have with the consumers.
AMD choose to polish drivers for a launch of a new line up of cards.
Consumers cry foul saying they can't wait for polished drivers.
Consumers continue to blame AMD for poor drivers.

 

[Erenhardt]

AMD can't shake the "poor drivers" stigma they have with the consumers.
AMD choose to polish drivers for a launch of a new line up of cards.
Consumers cry foul saying they can't wait for polished drivers.
Consumers continue to blame AMD for poor drivers.

Shin is not amd consumer.

Its quite a trolling to say drivers have nothing to do with new HBM memory after backing nv up on their gtx970 dedicated driver to fix segmented memory problem.

The tactics are simple. Clear the inventory, polish drivers for industry leading tech, release a good product and don't give ammo to the usuals.

 

[Stormflux]

I wasn't referencing Shin, just overall feeling from around the web.

I personally have no qualms with a Computex release. AMD have to hit this one out of the park top to bottom. If AMD knows they have a hit, we should be glad they're putting in the effort to execute. It's good timing too. What's coming down the line? nVidia have played their cards with the Titan. Next up is Pascal? Next year some time? Is there an expected refresh in the fall? AMD for the first time in a while can hold the performance crown while possibly offering reasonable pricing for an extended period.

 

[ShintaiDK]

Shin is not amd consumer.

Its quite a trolling to say drivers have nothing to do with new HBM memory after backing nv up on their gtx970 dedicated driver to fix segmented memory problem.

The tactics are simple. Clear the inventory, polish drivers for industry leading tech, release a good product and don't give ammo to the usuals.

How are those related? The driver doesnt know if its GDDR or HBM. Nor does it matter. Its all about addressing in the segment delivered by the memory controller.

 

[Cloudfire777]

How are those related? The driver doesnt know if its GDDR or HBM. Nor does it matter. Its all about addressing in the segment delivered by the memory controller.

Back that up with something. Anything.

We have several reasons to believe you are wrong based on:
7000 drivers having code regarding memory management from older VLIW architecture and causing trouble. Meaning HARDWARE have changed.
Nvidia saying they could fix 970 memory with improved drivers

I bet the API in the driver contain areas where 390 and 390X and all future HBM GPUs will use. AMD is making the foundation of it now and making sure to take the time to make it stable

 

[tviceman]

Back that up with something. Anything.
We have several reasons to believe you are wrong based on:
7000 drivers having code from older VLIW architecture and causing trouble. Meaning HARDWARE have changed.
Nvidia saying the could fix 970 memory with improved drivers

What Shintai is saying is that the driver doesn't discern between what type of memory it is (DDR3, GDDR5, HBM). And without exquisitely knowing, I have to agree with him. HBM still has all the same operating parts (controllers, ROPs, latency, etc.) and it's still doing all the same things (storing and accessing data). Memory access should still work exactly the same on a driver level. Partitioning memory (a la GTX 970) or working through a different GPU architecture (VLIW) are totally different animals and you should know that. If anything, having such massive bandwidth might reveal bottlenecks in other areas of the driver and GPU that can be addressed via software, so that may be what the rumor is referring to.

Regardless, this is a stupid rumor that means very little. The release date rumor was always Q2, and this rumor changes NOTHING in regards to that.

 

[ShintaiDK]

Back that up with something. Anything.

We have several reasons to believe you are wrong based on:
7000 drivers having code regarding memory management from older VLIW architecture and causing trouble. Meaning HARDWARE have changed.
Nvidia saying they could fix 970 memory with improved drivers

I bet the API in the driver contain areas where 390 and 390X and all future HBM GPUs will use. AMD is making the foundation of it now and making sure to take the time to make it stable

Tell me an example in history. Where a driver ever handled what type of memory a memory controller had connected. And not just querying the controller for any action relating to memory that the controller translates into an address space.

You are basicly saying that the memory controller is run by software and not hardware. Something that is completely wrong. Also if the driver wasnt loaded, how would you access it?

All the driver and applications see is simply a memory address from x to y. Then you can do all sorts of actions within that. However none of them have any kind of relation to what memory type is connected.

 

[Cloudfire777]

Im saying that the connection to HBM is different than GDDR5, and that is written in the code somewhere. With GDDR5 you have say 8x32bit spread out through 8 different chips. Each chip with their own memory controller. With HBM you have still chips spread out but they are stacked. Which needs a little different approach.

 

[kawi6rr]

How are those related? The driver doesnt know if its GDDR or HBM. Nor does it matter. Its all about addressing in the segment delivered by the memory controller.

I believe what he’s referring to was the fact that you forgave Nvidia and defended them even when they blatantly lied about the 970 memory but will troll every AMD thread to try to put anything AMD does in a bad light.

 

[therealnickdanger]

My completely baseless assumption is that AMD wants to hit the ground running with excellent drivers, CF profiles, stable hardware, and a plethora of FreeSync monitors all in one go so that when they actually launch the card, you and I can go out and buy the card(s) and monitor(s) while we're still salivating.

 

[Headfoot]

Im saying that the connection to HBM is different than GDDR5, and that is written in the code somewhere. With GDDR5 you have say 8x32bit spread out through 8 different chips. Each chip with their own memory controller. With HBM you have still chips spread out but they are stacked. Which needs a little different approach.

Probably ends up in firmware/microcode/vBios. Which may or may not still be in the optimization phase.

 

[kawi6rr]

My completely baseless assumption is that AMD wants to hit the ground running with excellent drivers, CF profiles, stable hardware, and a plethora of FreeSync monitors all in one go so that when they actually launch the card, you and I can go out and buy the card(s) and monitor(s) while we're still salivating.

This is what I gathered as well.