5970 has more transistors than GF100: 4,200 millions vs. 3,100 milliones. So, i guess the 5970 has a lot of useless transistors, too?
NVIDIA GF100 & Friends speculation
- Thread starter Arty
- Start date
5970 has more transistors than GF100: 4,200 millions vs. 3,100 milliones. So, i guess the 5970 has a lot of useless transistors, too?
There is really nothing there, you are trying to make something out of nothing
, they used the heaven demo because that is what AMD has been using to show off tesselation so far in a game engine (externally it is easy to compare for public to look into it). And nV has actually alot more flexibility when it comes to increasing the tesselation amounts.MSAA how is it impacted with increased polygons on the screen?
After your edit, triagle set up is also frequency bound
thats why its always stated triangles per clock! But again, we are looking at shifting bottlenecks.
Last edited by a moderator:
Yes. an X2 makes for a redundant mess.
I think what we are looking at is this -
The benchmarked card is clearly faster than the 5870, so much so that a gpu refresh will not give ATI the single fastest card back this series.
Nvidia will need their own dual gpu to beat the 5970, however it is highly likely that ATI will be able to beat it on dual gpu terms with a refresh. That is assuming neither abandons the 300w limit.
While Nvidia has improved more on performance compared to last series, it is still not enough to beat ATI on price-performance.
Smart move by Nvidia this however, they have given very few answers and left more questions being asked.
The benchmarked card is clearly faster than the 5870, so much so that a gpu refresh will not give ATI the single fastest card back this series.
Nvidia will need their own dual gpu to beat the 5970, however it is highly likely that ATI will be able to beat it on dual gpu terms with a refresh. That is assuming neither abandons the 300w limit.
While Nvidia has improved more on performance compared to last series, it is still not enough to beat ATI on price-performance.
Smart move by Nvidia this however, they have given very few answers and left more questions being asked.
DavidGraham
Veteran
And you think Nvidia wasted huge area on something that would give them only 2% of performance?!
Razor1: Yes, that's it exactly. Doubling of triangle-rate is probably the best thing on AFR. And ATi used it as alternative solution. Cypress isn't direct competitor of Fermi, so I see no reason why its triangle-rate should be comparable to Fermi. The competitor is Hemlock.
Are you sure Hemlock isn't faster in another games/benchmarks? Anyway, this isn't about transistors, it's manufacturing costs / performance (two 300mm² GPUs aren't more expensive then a single 450mm² one).
chavvdarrr
Veteran
Not to mention blasphemy called AFR
PSU-failure
Newcomer
That just show you never analyzed the benchmark results...
MSAA shows a huge drop with grass, to the point Cypress "heavy tessellation" framerate is not that different than "no tessellation and much grass" framerate.
And as I said in my previous edit, Heaven is far from being 100% setup bound, something unrelated to it nor ALU throughput, nor memory BW, nor texturing, nor blending comes into play on the whole benchmark when comparing Cypress to Juniper with almost the same specs except setup rate.
Holly molly are you serious about playing down Fermi single GPU becouse it doesn't spank a 5970 which is dual-GPU but "just" comes close!?
I'm asking you how MSAA is affected..... I'm looking for your analysis, just stating it doesn't mean much.
Juniper has less in every department so bottlenecks will automatically shift between the GPU's.
Hmm well Fermi will compete in price range somewhere near or just under Hemlock, so that gives you a good idea of what nV is going for. Market slides are marketing slides, just have to wait till independent reviewers will have to show what you are looking for.
NathansFortune
Regular
So I just read the summaries and what-not of Fermi. I'm not the most technical person on the GPU side, more a CPU person, but from looking at the slides Nvidia have build a quad-core GPU no, with each core having 128CUDA cores and a rasterizing engine.
This diagram looks eerily similar to early dual/quad core CPU designs from AMD and later Intel. So this chip is going to be massively parallel right. Is this the best path for Nvidia to take, from experience with parallelism in the CPU space it has taken 6-7 years for some of the more basic elements to take advantage of having >1 core. Then again I suppose the same questions were raised when AMD went multi-core with Athlon.
Anyway, that is my limited understanding of what Nvidia are trying to achieve, it's probably misguided and wrong, so please correct me if this is the case. If I'm right the design paves the way for very easy scalability, I mean they can basically cut the chip in half and get 50% performance with a 50% die size or increase the number of 'cores' in the next iteration (32nm) to 6 along with other general architectural improvements and efficiency gains and finally stop increasing the die size but get massive performance increases similar to what Intel/AMD do in the CPU space.
What is the real thinking, am I completely wrong?
This diagram looks eerily similar to early dual/quad core CPU designs from AMD and later Intel. So this chip is going to be massively parallel right. Is this the best path for Nvidia to take, from experience with parallelism in the CPU space it has taken 6-7 years for some of the more basic elements to take advantage of having >1 core. Then again I suppose the same questions were raised when AMD went multi-core with Athlon.
Anyway, that is my limited understanding of what Nvidia are trying to achieve, it's probably misguided and wrong, so please correct me if this is the case. If I'm right the design paves the way for very easy scalability, I mean they can basically cut the chip in half and get 50% performance with a 50% die size or increase the number of 'cores' in the next iteration (32nm) to 6 along with other general architectural improvements and efficiency gains and finally stop increasing the die size but get massive performance increases similar to what Intel/AMD do in the CPU space.
What is the real thinking, am I completely wrong?
PSU-failure
Newcomer
I quite agree with that.
As for the "multicore" issue, there's nothing wrong as graphics is an inherently multi-thread task.
Why do people insist on comparing a dual GPU card to a single GPU card? I'm sorry, it may work on a $ to $ basis, but after that, it holds no water. Most people with a decent IQ will compare it to Cypress, not Hemlock.
Hemlock was designed as direct competitor of nVidia's single-GPU high end. It's already 3rd generation product using this approach. So it's completely correct to compare these products. And as we see, Fermi is "just" competitive in some cherry-picked benchmarks (in fact cherry-picked scenes of cherry-picked benchmarks), which reflects its strongest point - triangle processing.
XMAN26: So you're saying it's perfectly legitimate to compare 6 months old product targeting price segment A to a new product targeting price segment B?
Last edited by a moderator:
Most people with a decent IQ will realise the whole story too, transistors, die size etc etc. Unfortunately for Nvidia, most people who don't know the whole story will only see gf100 as being second quickest card.
Nvidia has a real problem here, because they are going to release a very late graphics card that is not the fastest available. Last time that happened was when?
Similar threads
- Deleted member 2197
- Architecture and Products
