AMD announces 32GB Fire Pro W9100

[Atmos]

3 hours ago, Fourmi Kill3r said:

Who will buy this knowing that we can't even use 12GB on a Titan X (gamer talking)

100% not a gamer card.

If you're putting it to use in a workstation or server render machine though, that could come in very handy. Hell, there are some entire projects I'd be able to load completely into the damn things vram, that would sure as shit cut down render times

[ChineseChef]

I completely understand this is a workstation card, but are there a lot of workloads that are currently being used where they don't have enough VRAM, yet also have sufficient GPU horsepower left to use more?  I know larger simulations and renderings likely need more VRAM, and scientists/engineers could likely use a near infinite amount of it, but will the GPU keep up?  I guess the greater general storage will be useful, but I would have thought there would be a delicate balance of usefulness between storage space vs computation speed.  At a certain point, having more of one doesn't help with the other.  And at if it becomes too unbalanced, it would likely simply not be worth the money. 

[dalekphalm]

1 minute ago, ChineseChef said:

I completely understand this is a workstation card, but are there a lot of workloads that are currently being used where they don't have enough VRAM, yet also have sufficient GPU horsepower left to use more?  I know larger simulations and renderings likely need more VRAM, and scientists/engineers could likely use a near infinite amount of it, but will the GPU keep up?  I guess the greater general storage will be useful, but I would have thought there would be a delicate balance of usefulness between storage space vs computation speed.  At a certain point, having more of one doesn't help with the other.  And at if it becomes too unbalanced, it would likely simply not be worth the money. 

Yes, there are limits of course where adding more VRAM will do nothing. But in the Compute sector, that ceiling is much higher than in games. There are plenty of scenarios where a W9100 will be able to fully utilize the 32GB VRAM without bottlenecking the GPU itself first.

[Doobeedoo]

That's quite of vram nice.

[dalekphalm]

2 minutes ago, Doobeedoo said:

That's quite of vram nice.

What? lol sounds like you just chewed a bunch of words and spit them out

[ChineseChef]

1 hour ago, dalekphalm said:

Yes, there are limits of course where adding more VRAM will do nothing. But in the Compute sector, that ceiling is much higher than in games. There are plenty of scenarios where a W9100 will be able to fully utilize the 32GB VRAM without bottlenecking the GPU itself first.

This would likely be the case in large scale simulations wouldn't it?  Things where the speed of processing is not the issue, but the amount of variables and resulting data can be cumbersome.  Especially multi-stage calculations, where you need to retain the intermediate results while you perform multiple iterations.  Things like 3d modeling and such.

[Doobeedoo]

2 minutes ago, dalekphalm said:

What? lol sounds like you just chewed a bunch of words and spit them out

Meh im on my phone... No internet for my pc for 2 days

[ChineseChef]

3 minutes ago, Doobeedoo said:

Meh im on my phone... No internet for my pc for 2 days

You poor poor man, my heart goes out to you.

[Sharkyx1]

I'd be impressed to see the PCB, probably looks like new york traffic or that unblock me game.

[Sakkura]

1 hour ago, Doobeedoo said:

Meh im on my phone... No internet for my pc for 2 days

Tethering? Saved me when my broadband connection was down for a while last year.

[manikyath]

7 hours ago, Fourmi Kill3r said:

Who will buy this knowing that we can't even use 12GB on a Titan X (gamer talking)

people who need a firepro are very different people to people who need a radeon/gtx

 

still tho, why is a card that literally doesnt change anything but the vram even newsworthy?

[Doobeedoo]

1 hour ago, Sakkura said:

Tethering? Saved me when my broadband connection was down for a while last year.

Changing ISP and wires in house. It's a mess.

[TonyTheNain]

2 hours ago, manikyath said:

still tho, why is a card that literally doesnt change anything but the vram even newsworthy?

I really don't know

[AlTech]

8 hours ago, Evanair said:

The random posts about a CPU chip on a GPU or using massive system ram as VRAM makes me facepalm.  Apparently no one has heard of ASIC before.

I have.

[Dabombinable]

Ok...but its still GCN 1.1? Aren't AMD up to GCN 1.3 on the 28nm node?

[f22luke]

14 hours ago, ShadowCaptain said:

Eventually they will put a CPU on it and the GPU will become a PC

wait, this is circular  I guess what we want is faster RAM that is accessible by both teh CPU and GPU, so that a GPU no longer needs VRAM but can use like a 200GB super fast RAM that is shared

 

HSA...?

[dalekphalm]

24 minutes ago, Dabombinable said:

Ok...but its still GCN 1.1? Aren't AMD up to GCN 1.3 on the 28nm node?

Correction: There is no GCN 1.3. GCN 1.2 was the latest 28nm arch.

 

There does not exist any GCN 1.2 Compute GPU's. GCN 1.2 was only released in the form of Tonga and Fiji. Tonga is too small to be of use in the high end Compute market, and Fiji is limited by VRAM (4GB HBM max) - plus the only Fiji variants we saw (Fury series) has 1/16th FP64 (Double Precision) compute power - which is awful. The w9100 (Hawaii based) has 1/2 FP64 compute power. (This is the ratio of FP32 to FP64 compute power)

[Dabombinable]

16 minutes ago, dalekphalm said:

Correction: There is no GCN 1.3. GCN 1.2 was the latest 28nm arch.

 

There does not exist any GCN 1.2 Compute GPU's. GCN 1.2 was only released in the form of Tonga and Fiji. Tonga is too small to be of use in the high end Compute market, and Fiji is limited by VRAM (4GB HBM max) - plus the only Fiji variants we saw (Fury series) has 1/16th FP64 (Double Precision) compute power - which is awful. The w9100 (Hawaii based) has 1/2 FP64 compute power. (This is the ratio of FP32 to FP64 compute power)

Ah right. That makes sense.

[Sakkura]

5 hours ago, dalekphalm said:

Correction: There is no GCN 1.3. GCN 1.2 was the latest 28nm arch.

 

There does not exist any GCN 1.2 Compute GPU's. GCN 1.2 was only released in the form of Tonga and Fiji. Tonga is too small to be of use in the high end Compute market, and Fiji is limited by VRAM (4GB HBM max) - plus the only Fiji variants we saw (Fury series) has 1/16th FP64 (Double Precision) compute power - which is awful. The w9100 (Hawaii based) has 1/2 FP64 compute power. (This is the ratio of FP32 to FP64 compute power)

If you want to be entirely correct, there is no GCN 1.2 either.

 

There's GCN gen1, gen2, and gen3. The 1.0, 1.1, and 1.2 names were just made up by the press back when AMD didn't have any official codenames for the architecture revisions.

 

1/16 FP64 performance in Fiji is pretty awful, but it's actually better than what Nvidia offers. GM200 cuts all the way back to 1/32.

[patrickjp93]

On 4/15/2016 at 10:20 AM, Moonzy said:

maybe one day we'll not just have ram disk

but vram disk for our game textures, so no more loading time

That already exists with modern titles, and it's stupid. Your game should just intelligently stream in data to the GPU when you are getting close to a map boundary so you're not wasting GPU memory (and heat) by just having it sit there full-time.

[Moonzy]

1 minute ago, patrickjp93 said:

That already exists with modern titles, and it's stupid. Your game should just intelligently stream in data to the GPU when you are getting close to a map boundary so you're not wasting GPU memory (and heat) by just having it sit there full-time.

i'll trade a couple more watts (less than 5% of total system watts anyway) for fast load time

[patrickjp93]

Just now, Moonzy said:

i'll trade a couple more watts (less than 5% of total system watts anyway) for fast load time

HBM2 uses 60W at full speed, so take that how you will.

[Moonzy]

Just now, patrickjp93 said:

HBM2 uses 60W at full speed, so take that how you will.

well... sht `-`

 

GDDR5X ftw? D:

[patrickjp93]

1 minute ago, Moonzy said:

well... sht `-`

 

GDDR5X ftw? D:

HMC or 3DXPoint eventually.

[AlexGoesHigh]

16 minutes ago, patrickjp93 said:

HMC or 3DXPoint eventually.

Wait, isn't 3DXPoint flash for SSD when did Intel/Micron said it was volatile memory? Or am i missing something.