Nvidia to demo Fermi-based Tesla GPUs November 17

[lightman]

Well to put it into perspective,the HD5870 does 2.7 terraflops in single precision math,and dual precision math is 1/5th that value,so it comes in at about 550 gigaflops,which isn't exactly a slouch,and still way faster than current teslas as well.

Rumors suggest Fermi cranks that up to about 700 Gigaflops,better than the HD5870,but not a night and day difference either,and that's ignoring the delay it's suffering,and that a refresh version of the HD5870,probably clocked much closer to the 1 Ghz mark,would likely close into that 700 Gigaflop double precision floating point math of fermi quite easily.

All true, but you're not considering the software part of the equation. Right now, there are quite a number (most, probabily, but I don't have a big enough of a statistical population to back the claim with solid data, so take it for what it is, a guesstimate ;)) of scientific applications, most in-house, ours included, that are CUDA-based. Rewriting them to accomodate for a change in architecture is most of the times not an option.

ATI right now doesn't offer a comparable alternative language/environment to CUDA. Last time I checked, they didn't even have an OpenCL driver, especially under linux (which, like it or not, is widely used within the scientific community, and ATI drivers under linux have always been problematic, or at least way worse than NVIDIA's).

OpenCL might change the name of the game, but right now, CUDA is the answer to our computational needs.

[shadow001]

All true, but you're not considering the software part of the equation. Right now, there are quite a number (most, probabily, but I don't have a big enough of a statistical population to back the claim with solid data, so take it for what it is, a guesstimate ;)) of scientific applications, most in-house, ours included, that are CUDA-based. Rewriting them to accomodate for a change in architecture is most of the times not an option.

ATI right now doesn't offer a comparable alternative language/environment to CUDA. Last time I checked, they didn't even have an OpenCL driver, especially under linux (which, like it or not, is widely used within the scientific community, and ATI drivers under linux have always been problematic, or at least way worse than NVIDIA's).

OpenCL might change the name of the game, but right now, CUDA is the answer to our computational needs.


Indeed,it's mostly on the software front with cuda where the biggest advantage lies,but Let's see how OpenCL turns out since it has the huge advantage that it's an open standard and works with both Nvidia and ATI cards,and if developers are interested to develop software that runs on both brands,that's the only answer right now.


Btw,They do have OpenCL drivers now,but you have to register on their developer site to get them.

[josiahsuarez]

Seems that according to ATI,they've sold close to 100 000 HD5800 cards so far,so at 300 to 400$ a pop(depending which version purchased),so that's already between 30 and 40 million in sales and it's only been 7 weeks since it was released,and the christmas shopping season is coming,and the Bad yeilds from TSMC won't continue right up until the month of march/April of next year,when the gaming version of fermi is released either,so the pressure on Nvidia is very much on,no matter how much damage control they try to do.

if worse comes to worse, GTX 380 should be available at liquidation prices when Nvidia goes out of business soon.

[Airbrushkid]

I do not see Nvidia going out of business any time in the next 5 - 10 years. Video cards are not the only they make.

[shadow001]

I do not see Nvidia going out of business any time in the next 5 - 10 years. Video cards are not the only they make.


I can imagine them becoming more a software company as time goes on if you think about it,since Intel won't give Nvidia a licence for making chipsets for current and future Intel chips using built in memory controlers,then add that both Intel and AMD will be introducing CPU's with built in GPU's,either in the same die,or at least in the same cpu packaging,so that pretty much eliminates the lower end and maybe even the midrange market for GPU's,PC's simply won't need them,at least when it comes to the OEM market.


Same goes for the mobile market,since CPU's with built in GPU's on die will also end up there as well,and again,Nvidia also won't have a licence to make chipsets for mobile platforms using CPU's with built in memory controlers,so they're out of that market as well.


That leaves the high end gamer market,quadro workstation and tesla,but even there,it'll become a 3 way fight soon enough since ATI will continue to compete there,and Intel will eventually join the fight with larrabee,so it'll be interesting to see what happends in the next few years.

[Muppet]

I can imagine them becoming more a software company as time goes on if you think about it,since Intel won't give Nvidia a license for making chipsets for current and future Intel chips using built in memory controllers,then add that both Intel and AMD will be introducing CPU's with built in GPUs,either in the same die,or at least in the same CPU packaging,so that pretty much eliminates the lower end and maybe even the midrange market for GPU's,PC's simply won't need them,at least when it comes to the OEM market.


Same goes for the mobile market,since CPU's with built in GPU's on die will also end up there as well,and again,Nvidia also won't have a license to make chipsets for mobile platforms using CPU's with built in memory controllers,so they're out of that market as well.


That leaves the high end gamer market,quadro workstation and tesla,but even there,it'll become a 3 way fight soon enough since ATI will continue to compete there,and Intel will eventually join the fight with larrabee,so it'll be interesting to see what happens in the next few years.

I think in the future, we will have a single chip solution for both CPU and GPU. I see that Intel is moving towards this with Larrabee and Nvidia with Fermi. It will still be a little ways off yet. But i'm sure that is the future. And if that is the case Nvidia will be in a strong position. Even more so that AMD, unless they are working on something similar, which they well could be.

[Viral]

Yes, AMD were first to unveil that plan with fusion. I think the first fusion (Bulldozer) will be MCM (two seperate dies on one package) linked via HTT.

Idealy they are the company most likely to do it right, having capable CPU's and GPU's already as well direct connection with a foundary capable of producing both in the same place.

Nvidia don't have a CPU and Intel don't have a GPU worth mentioning.

[Muppet]

Yes, AMD were first to unveil that plan with fusion. I think the first fusion (Bulldozer) will be MCM (two separate dies on one package) linked via HTT.

Ideally they are the company most likely to do it right, having capable CPU's and GPU's already as well direct connection with a foundry capable of producing both in the same place.

Nvidia don't have a CPU and Intel don't have a GPU worth mentioning.

I didn't know about Fusion. I will have to do some research.

As for Nvidia don't have a CPU, no they don't. But I think that Fermi is the start of there integrated CPU/GPU. It has to start somewhere and I think this is a stepping stone on the way to that mark.

[Heinz68]

On what reasoning is it a fail. Nobody but Nvidia knows that atm. It could be a real winner and blow the competition away. Only when we get some real info will we know for sure. Either way it's to early to call it a failed GPU.
I agree. I doubt the Fermi is going to be fail. NVIDIA almost always come up with some great product.

The problem is, I need upgrade now and not who knows when, so the only fail we know right now is being late on the market.

I was always happy with the NVIDIA cards I had, can't say the same about the NVIDIA tactics lately or the NVIDIA shills and top management.

PS
Getting the GPU too big to push the GPGPU on the gaming card might also prove as fail due to higher price, but we need to wait for the card first to find out.

[shadow001]

I didn't know about Fusion. I will have to do some research.

As for Nvidia don't have a CPU, no they don't. But I think that Fermi is the start of there integrated CPU/GPU. It has to start somewhere and I think this is a stepping stone on the way to that mark.


Honestly,i don't think that a single chip can do both CPU related workloads and GPU related workloads all in the same design,as the requirements for both in terms of architecture features are vastly different for both types....CPU's are specialised in out of order workloads where even today,multithreading is still used lightly in most software.


While GPU's are massively parallel designs,which to get the best performance out of that design,CPU related software has to be split between hundreds of execution units simultaniously within that GPU architecture,such as the 512 cuda units within Fermi for instance,and the maximum theoretical that fermi could handle is over 10 000 threads in flight,but there's always software that can't be split up into multiple threads to that extreme degree,if at all.


Not to mention the legal issues,as Nvidia might need an X86 licence,and it's intel that owns it,or they might need to buy out a company that has one already like VIA,which bought out Cyrix a few years ago.

[Muppet]

I agree. I doubt the Fermi is going to be fail. NVIDIA almost always come up with some great product.

The problem is, I need upgrade now and not who knows when, so the only fail we know right now is being late on the market.

I was always happy with the NVIDIA cards I had, can't say the same about the NVIDIA tactics lately or the NVIDIA shills and top management.

PS
Getting the GPU too big to push the GPGPU on the gaming card might also prove as fail due to higher price, but we need to wait for the card first to find out.

I know just what you mean. I really wanted a new Graphics card before Christmas and not next year. You are also more likely to be right about the price. I'm sure it will be expensive compared to ATI's latest offering. But in the past that has generally been the case, so I wouldn't expect it to change. Although it would be nice.