HBM For Carrizo?

[Erenhardt]

Wouldn't HBM be better for system memory than DDR3/4?

HBM only system?

 

[monstercameron]

Wouldn't HBM be better for system memory than DDR3/4?

HBM only system?

yes, lower latency higher bandwidth.

 

[el etro]

The 512sps GCN card on the desktop is Radeon HD 7750 that is GCN1.0, consumes around 55-60W on gaming and performs around a GTX750m-760m.

GCN1.1 revision improves sightly the efficiency of GCN archs and GCN1.2 improves the performance/SP of the cards in a bunch of gaming situations.
7790 consumes ~20w more than 7750 and uses the same die area as 7750(that is a partially disabled 7770).

So, do AMD could put 768GCN1.2 sps on the next APU?

 

[NostaSeronx]

Wouldn't HBM be better for system memory than DDR3/4?

No.

HBM only system?

No.

What HBM will be good at is reducing mostly reads and some writes to/from DDR3/DDR4 system memory.

128-bit dual channel non-ECC // 2x64b --> 32 GB @ DDR3-1866/DDR4-2400
64-bit single channel ECC // 1x64b -> 16 GB @ DDR3-1866/DDR4-2400

(2 * 64-bit * 1.866(or)2.400) ÷ 8=> 29.856 GB/s for DDR3-1866 and 38.4 GB/s for DDR4-2400
(64-bit * 1.866(or)2.400) ÷ 8=> 14.928 GB/s for DDR3-1866 and 19.2 GB/s for DDR4-2400

versus
(2 * (8 * 128-bit) * 1) ÷ 8 => 256 GB/s for 2 GB HBM.

256 ÷ 38.4 => 6.67

Memory Bottleneck -> Removed
---

Anyone else notice that the GPU is 1 GHz. It also doesn't say it is integrated, but PCIe 3.00.

 

[Enigmoid]

The 512sps GCN card on the desktop is Radeon HD 7750 that is GCN1.0, consumes around 55-60W on gaming and performs around a GTX750m-760m.

GCN1.1 revision improves sightly the efficiency of GCN archs and GCN1.2 improves the performance/SP of the cards in a bunch of gaming situations.
7790 consumes ~20w more than 7750 and uses the same die area as 7750(that is a partially disabled 7770).

So, do AMD could put 768GCN1.2 sps on the next APU?

Do be careful. Lots of people are comparing performance per sp from tonga to taheti, which had the lowest perf/sp of all gcn 1.0 products.

http://www.hardware.fr/articles/926-24/tonga-vs-tahiti.html

Hell, that's <10% gain per sp from taheti to tonga (GCN 1.0 to 1.2). Not saying that this test isn't flawed but that tonga, at least in its neutered form needs a little more to make it competitive.

http://www.techpowerup.com/reviews/Sapphire/HD_7790_Dual-X/26.html

1200p
7790 - 95%
7870 - 141%
Same clocks, 1280 vs 896 sp. 43% more shaders on the 7870, 48% more performance (Pitcarin has more bandwidth: performance delta grows from 720p to 1440p). Bonaire does so with significantly greater efficiency though.

Bonaire is 160 mm^2. Cape Verde is 123 mm^2. Pitcarin is 212 mm^2.

7790 is significantly larger than a 7750.

More than 512 shaders only makes sense if HBM is a go. As well, 768 sp is too much for 35W mobile so AMD would have to either disable a bunch of shaders or produce two dies (512 barely fits in a 35W window).

 

[USER8000]

Cape Verde fully enabled has 640 shaders and is the basis of the HD7770. The HD7790 not only has 896 shaders but has audio DSPs onboard which take up more die sapce.

 

[krumme]

NV have clearly shown there is a huge market for discrete mobile gpu solutions.
I think AMD have severely underestimated the need for powerfull gpu here.

Now i think what makes the situation interesting is a combination of a beefed up carizo cpu that can handle games, and will not be a weak point, and then HBM that enables 3 times as fast gpu speed.

I think AMD is actually in a situation now to do something radically different on the market. What is the problem that a mobile solution only fits in 65watt? - it doesnt matter. There is sold tons of discrete solutions with that tdp and above. There is a market, and binning is zero problem this late in process node development if its 28nm.

If AMD did such a high perf apu solution they will be in a situation where there is far less competition than the positions below or above. But probably they will get in this no mans land where there is no need. Who needs 384/512 sp? - its useless. An eg. Intel 4400 gfx will do the same facebook gaming. And for higher perf. gaming its just a burden.

But probably we will get this 512 because from a technically and productions point its optimal. Typical AMD engineering thinking. Its like they design all their products the wrong way round.