AMD CPU speculation... and expert conjecture

[griptwister]

I would like to feel bad for them, I really would.

But I just can't stop laughing.

You too? :lol:

 

[JAYDEEJOHN]

http://hothardware.com/News/NVIDIA-Embraces-PlayStation-4-with-PhysX-and-APEX-Support/
I guess proprietary just aint what it used to be

 

[amdfangirl]

http://hothardware.com/News/NVIDIA-Embraces-PlayStation-4-with-PhysX-and-APEX-Support/
I guess proprietary just aint what it used to be

Nvidia just doesn't want game devs to abandon PhsyX so that it can continue to sell it as a feature in the PC space.

 

[mayankleoboy1]

I don't think so. Hair is notoriously hard to simulate well. Even in a simplified equation where each strand can only bend at one point in one direction, you have the effects of momentum, gravity, wind, and animations all comming into effect on thousands of individual stands, not to mention strand-on-strand action. On top of that, all these strands will individually deflect light, causing an extra surge in GPU load due to extra processing of the resulting shadows (especially if the engine isn't smart enough to treat hair as a flat object in this regard).

Hence why no ones bothered with a PhysX implementation of hair. Hair is hard.

sounds mildly innuendo-ish

 

[JAYDEEJOHN]

Nvidia just doesn't want game devs to abandon PhsyX so that it can continue to sell it as a feature in the PC space.

Oh, I understand this, of they want this.
Its just that the HW itll be running on will be the exact same HW found in other PCs where it wont be able to run

 

[anxiousinfusion]

Have a bone:
http://www.cpu-world.com/news_2013/2013030801_AMD_FX-4350_and_FX-6350_CPUs_are_available_for_pre-order.html

 

[mayankleoboy1]

Oh, I understand this, of they want this.
Its just that the HW itll be running on will be the exact same HW found in other PCs where it wont be able to run

Not to mention that almost the same exact hardware will be in the Xbox 720, where Physx will be able to run.

 

[de5_Roy]

amd relates news round up
http://semiaccurate.com/2013/03/07/opencl-performance-of-next-gen-amd-mid-range-gpus-hit-the-intertubes/
http://semiaccurate.com/2013/03/08/opencl-performance-of-next-gen-low-power-apus/

http://www.fudzilla.com/home/item/30710-amd-temash-tablet-looks-good
http://www.fudzilla.com/home/item/30711-three-new-kabini-parts-leaked
http://www.fudzilla.com/home/item/30709-amd-richland-notebooks-to-have-face-unlock
http://www.fudzilla.com/home/item/30708-temash-chip-pictured

 

[jdwii]

To be fair, last generation, EVERYONE was complaining that NVIDIA cards drew too much power and focused too much on compute performance, so NVIDIA went the other way, and made a more power efficient card that excelled in gaming.

Now everyone is complaining they suck at compute performance.

Secondly, on AMD cards, TressFX is showing a good 30FPS hit, so its not JUST NVIDIA here. TressFX kills performance on more or less every GPU out there.

The Tesselation bug is probably a bigger killer in all honesty.

Again it happens and why can't they do both compute and gaming performance you know like Amd.

Have a bone:
http://www.cpu-world.com/news_2013/2013030801_AMD_FX-4350_and_FX-6350_CPUs_are_available_for_pre-order.html

Kinda miss the old Amd days when they made a new phenom and put it at the same price of the current phenom and lowered the lesser chip by around 10$.

 

[Cazalan]

Speaking of GDDR5 I found JEDEC has a spec for a 256pin SODIMM from 2012.

http://www.jedec.org/standards-documents/docs/so-018a

Think of the marketing opportunities there if you have Intel DDR3 laptops sitting next to AMD GDDR5 laptops on the shelves. Every consumer can understand 5 > 3.

 

[mayankleoboy1]

Again it happens and why can't they do both compute and gaming performance you know like Amd.

If we believe S|A , Nvidia has issues with their design, which makes fabricating their GPU chips difficult, and lots of wastage.
So Nvidia is forced to cut down on the chip size. Which means removing the compute stuff. This also has the advantage that the power consumption gets reduced . But all this doesnt matter because Nvidia has a huge fanboi base, which will pay ridiculous prices for gimped cards.

 

[JAYDEEJOHN]

Understand tho, the HPC market demands low power usage, as thousands of these chips are used in one setup, and I believe Charlie was right with GF100, he really hasn't mentioned much about GK110 being this way.
And as the market matures, and NVidia or AMD for that matter pursue it, it has to fit within certain thermal designs.
Since this is AMDs first approach in this direction, they too will learn to get more bang per watt, and be capable of still giving decent GFX workloads.

 

[noob2222]

gamerk316 wrote :


To be fair, last generation, EVERYONE was complaining that NVIDIA cards drew too much power and focused too much on compute performance, so NVIDIA went the other way, and made a more power efficient card that excelled in gaming.

Now everyone is complaining they suck at compute performance.

Secondly, on AMD cards, TressFX is showing a good 30FPS hit, so its not JUST NVIDIA here. TressFX kills performance on more or less every GPU out there.

The Tesselation bug is probably a bigger killer in all honesty.

Again it happens and why can't they do both compute and gaming performance you know like Amd.

It was about money and the quadro sales, not the card design.

Nvidia argues it’d rather sell you a workstation-oriented Quadro card or dedicated Tesla-based board.
http://www.tomshardware.com/reviews/geforce-gtx-680-review-benchmark,3161-15.html

Kepler's actual compute power is pretty impressive gain over fermi.
http://www.xbitlabs.com/articles/graphics/display/nvidia-quadro-k5000_4.html#sect1

The gtx 680 was purposely compute gimped so they could sell the $2250 cards.

 

[jdwii]

If we believe S|A , Nvidia has issues with their design, which makes fabricating their GPU chips difficult, and lots of wastage.
So Nvidia is forced to cut down on the chip size. Which means removing the compute stuff. This also has the advantage that the power consumption gets reduced . But all this doesnt matter because Nvidia has a huge fanboi base, which will pay ridiculous prices for gimped cards.

Agreed but that is with every brand hell i was like that i even almost bought Bulldozer after the benchmarks i guess i eventually learned and did not.

 

[mayankleoboy1]

Kepler's actual compute power is pretty impressive gain over fermi.
http://www.xbitlabs.com/articles/graphics/display/nvidia-quadro-k5000_4.html#sect1

Though one has to wonder if that perf/watt win is because of the design, or because of the switch to 28nm, that allows moar transistors to be added.

Looking at the power usage of HD79xx cards, and the relative lesser power use of GTX680/670 , it appears that the HPC/compute part of the chip uses quite a lot of power, and die size.
With AMD getting a easier design to fabricate for TSMC, they could afford to design bigger chips. Plus, AMD being the underdog, they had to make a master-of-all chip . The biggest criticism of HD79xx is the power usage.

The gtx 680 was purposely compute gimped so they could sell the $2250 cards.

That depends on if you are a gamer, or a HPC engineer. Gamers should be happy that the GTX680 uses lesser powr than HD79xx, for equal per in games. (larger/smaller depending on the game an drivers)
As Nvidia has ~90% of the GPU HPC/Pro market, they can afford to sell the "full" chips at that ridiculous price. AMD is underdog in HPC too, so it cant price their chips at the same price.

Same story as Intel and AMD. Intel can sell insanely gimped chips at high prices, because they have almost a monopoly. Its just how economics work.

Never ever believe that AMD/ATI wont do the same thing if they had a bigger share of the market.

 

[mayankleoboy1]

Agreed but that is with every brand hell i was like that i even almost bought Bulldozer after the benchmarks i guess i eventually learned and did not.

Dont ever call BD as "not worth my money". Its the best proc ever.
Its just that the software dont like it.... :lol:

 

[de5_Roy]

^^ yeah. cerdible unbiased hardware enthusiast sites who are not paid off by amd said that it's worth gold awards. that's you, eteknix and amdzone. windows blue's bd patch will solve cpu performance issues when it comes out.

 

[sarinaide]

I don't think so. Hair is notoriously hard to simulate well. Even in a simplified equation where each strand can only bend at one point in one direction, you have the effects of momentum, gravity, wind, and animations all comming into effect on thousands of individual stands, not to mention strand-on-strand action. On top of that, all these strands will individually deflect light, causing an extra surge in GPU load due to extra processing of the resulting shadows (especially if the engine isn't smart enough to treat hair as a flat object in this regard).

Hence why no ones bothered with a PhysX implementation of hair. Hair is hard.

But TressFX is a AMD patent so you would expect Radeons to excel at this more than Nvidia regardless. So as AMD and its new partners develop games with tressfx support so it will improve.

Nvidia just doesn't want game devs to abandon PhsyX so that it can continue to sell it as a feature in the PC space.

CUDA and Physx will loose its product support because it is mutually exclusive and creates a closed market. DC and OpenCL etc encourage more software support, nvidia have a bit of a problem.

 

[amdfangirl]

CUDA and Physx will loose its product support because it is mutually exclusive and creates a closed market. DC and OpenCL etc encourage more software support, nvidia have a bit of a problem.

Do we have an open standard for a physics API?

 

[JAYDEEJOHN]

I would also add, as its often forgotten here, as well as elsewhere, the node shrinkage was a full node, not the half we are used to seeing, and is why nVidias power/perf rate looks so good compared to last gen, and why AMD could afford the first rough GNC/compute solution as well

 

[truegenius]

^bullet ! ?

 

[JAYDEEJOHN]

http://bulletphysics.org/wordpress/

 

[sarinaide]

I wonder if AMD will actually stick 8 cores into their 8-core processors or if they're going to lie yet again. 8 threads never has and never will be 8 cores.

We have done this to death, so to avoid the TL;DR, AMD do use 2,4,6 and 8 cores across the varying product lines. So the FX 8000 family is a octo-core regardless of what you have been reading from the internet.

 

[JAYDEEJOHN]

And if rumors are true, and the frontend gets its overhaul, itll be even octomoarish

 

[mayankleoboy1]

CUDA and Physx will loose its product support because it is mutually exclusive and creates a closed market. DC and OpenCL etc encourage more software support, nvidia have a bit of a problem.

CUDA is unlikely to die sonn.
For the same reason that PowerPC arch is stil alive : Its easier to buy newer hardware than to convert and optimise existing code to a new architecture.
Basically, code thats already on CUDA, will remain on CUDA for a long time. Newer code will like to remain x86 (Xeon Phi, anyone ?)