|  | | Nvidia GT300 Fermi Details Emerge |  Just like we disclosed in the first article "nVidia GT300 specifications revealed – it's a cGPU!", nVidia GT300 chip is a computational beast like you have never seen before. In fact, we would go as far out and state that this is as closest as GPU can be to a CPU in the whole history of graphics technology. Now, time will tell whatever GT300 was the much needed revolution.
Beside the regular NV70 and GT300 codenames [codename for the GPU], nVidia's insiders called the GPU architecture - Fermi. Enrico Fermi was an Italian physicist who is credited with the invention of nuclear reactor. That brings us to one of codenames we heard for one of the GT300 board itself - "reactor". When it comes to boards themselves, you can expect to see configurations with 1.5, 3.0 GB and 6GB of GDDR5 memory, but more on that a little bit later. Specifications
* 3.0 billion transistors
* 40nm TSMC
* 384-bit memory interface
* 512 shader cores [renamed into CUDA Cores]
* 32 CUDA cores per Shader Cluster
* 1MB L1 cache memory [divided into 16KB Cache - Shared Memory]
* 768KB L2 unified cache memory
* Up to 6GB GDDR5 memory
* Half Speed IEEE 754 Double Precision
As you can read for yourself, the GT300 packs three billion transistors of silicon real estate, packing 16 Streaming Multiprocessor [new name for former Shader Cluster] in a single chip. Each of these sixteen multiprocessors packs 32 cores and this part is very important - we already disclosed future plans in terms to this cluster in terms of future applications. What makes a single unit important is the fact that it can execute an integer or a floating point instruction per clock per thread.
Read the entire article in BSN.
Last edited by Regeneration; October 1st, 2009 at 10:19 AM..
| | | | 27 Comments | | | I hope they make better AA and AF. So far, enabling them gives a serious hit to performance when compared with ATI cards.
If they fixed that, I might buy my first nVidia card ever. | | | | AF has never made much of a performance hit.
AA's performance hit is angular dependant... | | | | | | As a non tech guy I'm little confused...as I understand it that Fermi uses multiple instruction multiple data MIMD as opposed to convetional graphics cards using single instruction multiple data SIMD....is that what makes it closer to a cpu and a "computational beast"
But I thought that direct compute and open cl was going to make all the DX11 cards computational beasts? So in practical terms what is MIMD and Fermi going to bring that I won't get with an ATi card when considering the average desktop user and not boffins running custom software on their tesslar gpu's? |
Last edited by technogiant; October 1st, 2009 at 09:53 AM..
| Quote | | | | | its not MIMD, but its very close to being a hybrid of SIMD and MIMD going by the design http://anandtech.com/video/showdoc.aspx?i=3651&p=1
the thread scheduler has been massively restructured so it can do A LOT more operations per clock, and to reduce the cycles of operations that typically require more then 2 to complete. | | | | Yeah I've read that review....long and short of it is saying that the Fermi is more designed with large business/reasearch requirements in mind rather than desktop users...was just wondering if Fermi will perhaps show its strengths for the home user when physics and AI on gpu become more utilised and the likes of real time raytracing come into being. Also do you think it will be the case that because of its difference in structure that we will see a lot of gpu accelerated desktop software that the ATi cards cannot utilize? Or perhaps both ATi and Nvidia's cards would be able to utilise them as they would probably use open cl but Fermi would probably be better at it because of its structural redesign? |
Last edited by technogiant; October 1st, 2009 at 12:57 PM..
| Quote | | | | | | | Quote:
Originally Posted by mkey "Emberge"? | Ahh, so someone *does* pay attention? How nice XD | | | | ObliVioN_ - in DP maybe, in SP I don't think so... and since AMD can do about 40% more chips from waffer (perhaps even more) than NV with Fermi...and since NV uses 384 bit memory interface which is more complex and...costly thn 256bit of Cypress...I guess new GF won't be a cheap one. | | | | We'll see how this actually performs, sofar all ive seen is a bunch of dubass fanboy hype and nvidia fueled bs 'previews'
As with any new hardware, before its released its the most amazing then ever, once it's released it never lives up to expectations | | | | There are some facts...ok? So...512 Cuda Cores is a fact. 256 of which are DP capable is a fact. So...if GT300 has DP capabilities 8x larger than GT200 I can assume this means some ~700GFLOPS DP. So I guess full 512 cuda cores with FP32 gives about 1.5TFLOPS peak SP power. Hmm...any comment for that? | | | | just one, it usually looks better on paper then it works in reality. | | | | Quote:
Originally Posted by Promilus There are some facts...ok? So...512 Cuda Cores is a fact. 256 of which are DP capable is a fact. So...if GT300 has DP capabilities 8x larger than GT200 I can assume this means some ~700GFLOPS DP. So I guess full 512 cuda cores with FP32 gives about 1.5TFLOPS peak SP power. Hmm...any comment for that? | Here are some facts for ya Nvidia Fakes Fermi Boards In GPU Technology Conference | | | | mkey - never denied that fermi board presented by nv was operational. So...what's the meaning of your comment? Anything about specs? No? | | | | No, nothing about specs, it's only about the integrity of the company. | | | | The manager in charge of Geforce knows price is a concern and has conceded that not only has nvidia not taken enough risks (swinging to far towards conservative RnD since NV30) but also that the G200 was priced wrong from the start.
Expect more aggressive pricing this time round. | | | | I dunno squall, I guess we'll see but I have a feeling your gonna need to get a second mortgage on your house to buy one of these things | | | | IMO its going to be at least 2x of the price of a 5870. | | | | 2x? Nah...chip itself is more than 40% larger and ram bus is 50% wider...however they'll have to price it right...We don't know how much AMD can reduce price of their GPUs...calculating that they've got some 60% working chips from one waffer (~220 Cypresses) which means they do ~130 chips from one waffer. If one 300mm waffer costs about 6k $ with production, design, development etc. etc. this gives 6000/132 = ~45$ per chip. This might not be the most accurate...however near real chip value. | | | | 280 was about two times as expensive as 4780 (at least here). Also, Nvidia will have to rush with this product and I'm guessing we'll see many chips released with the same card model. That will cause even more confusion on the market. | | | | For goodness sake all the news is Nvidia Nvidia Nvidia....instead of trying to rain on ATi's parade why don't they just shut up until they have got something to show other than a fake card.....so that will probably be Q2 2010....and then it will be mega expensive. | | | | well we all know how dose cards ends when nvidia rush
g92= 8800GT-GTS, 9800 , and also complete 100 series that is rename of one chip LOL
and i dont trust that this will be a great card |
Last edited by gen.Rage1991hrv; October 3rd, 2009 at 12:00 PM..
| Quote | | | | | Because ATI are busy bitching about batman not having the aa option show when using an ati card because they were too complacent and lazy to get off their fat bastard arses and work with the Dev team to check in a compatibility patch. | | | | lol those assholes, if anything, are lazy. | | | | You can enable MSAA for BAA in the CCC anyways...wow thats the most abbreviation Ive used in a sentence I think | | | | For the Double Precision Floating Point (DP) i have to say that maybe the ATi Radeon HD5870 OC Edition or the ATi Radeon HD5890 will be comparable with theSame GT300 Fermi, the reason I'm saying that is because the ATi Radeon HD5870 @850MHz has 2.72TFLOP/s of Single Precision Floating Point and for 1000MHz GPU Core frequency is 3.2TFLOP/s so if the ATi introduce the Same Core or a Refresh of the Core at 1000MHz then we have 3.2TFLOP/5=640GFLOP of IEEE 754 Double Precision Floating Point (DP). | | | | @Theodoros - wrong. Fermi DP performance index is based on FMA instructions. Radeon HD5k has it too! It doubles peak DP power as well...so Radeon 5870 with 2.72TFLOPS SP will have peak of 1088GFLOPS DP. It is still more than GF with ~700GFLOPS DP. Not much...but still. | | |