FERMI Benchmarked?

[trinibwoy]

The architecture is indeed an unknown variable in this game called graphic-card-industry. Still, seeing and believing NVIDIA thinks the future is HPC I'm not all that optimistic about Fermi's gaming performance. Who knows how many transistors are there for solely for HPC and GPGPU?

That's key IMO. In GT200 it was pretty clear that the DP support and shared memory held little benefit for graphics. They might have stored a few interpolation and/or texturing inputs in shared memory but nothing that they couldnt have done with a bigger register file. In Fermi however things have been changed up so much that it's not clear what, if anything, could benefit graphics too. It could be nothing like GT200 or they could use the new cache hierarchy and larger shared memory pool to benefit graphics in interesting ways. And of course we still know nothing of the fixed function setup, texturing and blending stages.

[NapalmV5]

I didn't mean the shaders. But iIrc Fermi has quite massive cache as well. Is this needed for gaming and SP-performance? Just one example

But I'm afraid I have no clue about such things

come on if we are to know everything about a gpu inners and outters before it gets released then by the time its out its just a boring gpu.. no fun left

[zalbard]

come on if we are to know everything about a gpu inners and outters before it gets released then by the time its out its just a boring gpu.. no fun left

But you can OC it and bench / play games!
And voltmod!

[NapalmV5]

^

[m.fox]

THIS

[003]

http://www.tomshardware.co.uk/forum/...6100_15_0.html

it is the 8th post on the page. Now 003 please stop trolling

This is still in reference to those graphs on Tom's Hardware, not the ones this thread is talking about. I am asking very reasonable questions.

[Solus Corvus]

Not really very reasonable IMO. If you had an NDA source that showed you the performance then you should know if they are fake or not.

[Neuromancer]

Not really very reasonable IMO. If you had an NDA source that showed you the performance then you should know if they are fake or not.

ouch

[JohnZS]

THIS IS OFFICIAL AND WHAT WE KNOW NOW

IEEE 754 single & double precision floating point units
Achieve up to 600 Gigaflops of double precision performance for faster and more accurate results.

ECC Support
Offers protection of data in memory to enhance data integrity and reliability for applications. Register files, L1/L2 caches, shared memory, and DRAM all are ECC protected.

Up to 6GB GDDR5 compute memory
Allows faster access to larger data sets.

NVIDIA Parallel DataCache Technology
Accelerates algorithms such as physics solvers, ray-tracing, and sparse matrix multiplication where data addresses are not known beforehand.

NVIDIA GigaThread Engine
Maximizes throughput by faster context switching, concurrent kernel execution, and improved thread block scheduling.

Asynchronous transfer capability
Turbocharges system performance by overlapping data transfers with computation.

High Speed , PCI-Express Gen 2.0 Data Transfer
Fast and high-bandwidth communication between CPU and GPU.

CUDA programming environment with broad support of programming languages and APIs
Choose C, C++, OpenCL, DirectCompute, or Fortran to express application parallelism and take advantage of the “Fermi” GPU’s innovative architecture

Availability: in early Q2 2010, that is on my nVidia news letter...

[eric66]

FH spoke with PR guy from nvidia he said early q1 2010

[JohnZS]

My nVidia news letter today said that Fermi will come in early Q2 2010 and that the most powerful Fermi will come in Q3 2010.
Some specs too
Double Precision Floating Point Performance 520GFlops - 630 GFlops Mind you this is all about the Quado series products for professional workstations. IT IS however the first word on Fermi and rough indication of performance (ish) presuming the Quadro Clocks are similar to the consumer clocks etc... There are 2 RAM sizes 3GB and the later model 6GB although I can see this being 1.5GB and 3GB on the consumer GPU's
John

[LordEC911]

rv870 330 mm^2
Fermi something around 500 mm^2

~550mm2, around 65nm G200 size.

you guys talk like fermi is just a refresh.. let me remind you its not

dont be surprised how and why fermi is more than twice as fast than 280/285 and faster than the 5000 series



amds gonna need their new 6000 series (whatever theyre going to name) to combat dual fermi.. 5890/5990 refreshes to keep as close as possible

Only if they hit their target clocks, which doesn't seem to be the case.

[initialised]

nvidia does try and go beyond just gaming. if i had to use CUDA just a little, i might not care that i get 5FPS less compared to an equally priced ATI card, cause you are getting some bonus features. which may not be for everyone, but do exist and do have a purpose.

It's usually the other way around, either the ATi card is a lot cheaper at a given performance point or it is faster for the same price point. You buy for CUDA then fair enough but if you're not going to use it why pay the CUDA tax for a GPU that is a generation behind and offers poorer value:

GT 220 £25 vs 4650 £29
GT 240 £51 vs 4670 £39
GTS 250 £70 vs 4850 £70
GTX 260 £140 vs 4890 £110
GTX 285 £212 vs 5850 £192
GTX 295 £400 vs 5870 £309
2x GTX 295 £800 vs 5970 £500

[Chumbucket843]

Only if they hit their target clocks, which doesn't seem to be the case.

so you got some A3 ES's ehh? stock clocks are not what napalm cares about if you havent seen his shader clocks! we still have no info on graphics performance.

[Neuromancer]

Yah I was trying to figure out the math earlier but my brain is a little fried today from what I figured

Fermi at 750MHz is projected to be 40% faster than Cypress (no comment on WHICH cypress though average it too 800MHZ and give nVidia the benefit of the doubt). However, they are currently capable of producing 500 MHz Fermi. Which is 33% slower Than projected fermi. So 500 MHz Fermi should be about equal to a 5850.

Which is pretty damn good, I do not know why people are saying that Fermi Absolutely must defeat ATi with best card.

In the overused Car analogy, Lamborghini and Ferrari should give up making cars because the Bugati Veyron is faster? Even though All of them are capable of going fast enough to have you thrown in jail?

[safan80]

My nVidia news letter today said that Fermi will come in early Q2 2010 and that the most powerful Fermi will come in Q3 2010.
Some specs too

John

So 6 months for Fermi to show up and least +1 month for actual availability.

the most powerful Fermi I take is another dual pcb that will show up in 7-9 months.



Looks like the new number is closer to 7 months of ATI rule. Looks like Nvidia needs to get back into chipsets or something.

[Solus Corvus]

Is it fiscal Q2 or calendar Q2?

Either way, a disappointment and already contradicting what 003 said.

[Vit^pr0n]

Is it fiscal Q2 or calendar Q2?

Either way, a disappointment and already contradicting what 003 said.

B-b-b-b-b-b-but he knows someone under NDA! How can he be wrong?!?!

[003]

Is it fiscal Q2 or calendar Q2?

Either way, a disappointment and already contradicting what 003 said.

That is for the Quadro and Tesla variants of Fermi and this has been known for a while now. GeForce is still on track for Q1 and will have higher clocks than Tesla and Quadro, and have ECC disabled, which trades off performance for precision, only needed in HPC applications.

B-b-b-b-b-b-but he knows someone under NDA! How can he be wrong?!?!

I'm not. It's been known for a while now that the Quadro and Tesla cards will be launching after GeForce. That isn't under NDA and was stated by Jen-Hsun Huang in Nvidia's most recent financial conference call.

[m.fox]

Slightly OT but I am wondering how 5970 with over 4TFLOPS of power can perform worse than a card with 1/3 of the crunching power.

You can't compare differect archs. ATI uses more "complex" SPs which need to be fed properly by the compilers. And there's the texturing, setup rate differences etc.

[Solus Corvus]

How about responding to the real question asked several times in this thread: Are the numbers in those charts close to the real ones? You can say that without giving any information about your "source".

[003]

How about responding to the real question asked several times in this thread: Are the numbers in those charts close to the real ones? You can say that without giving any information about your "source".

The charts are most likely faked. Price/performance of the high end GeForce Fermi will be very good. If the 5970 remains priced at what it is currently after Fermi is released, it will be quite a waste of money.

[Solus Corvus]

Nice dodge. So I take it that the numbers you have supposedly seen don't match what is in these graphs. A simple no would have sufficed.

[m.fox]

If the numbers aren't close, then why would the price of R800 have go down, as 003 says?

[003]

Nice dodge. So I take it that the numbers you have supposedly seen don't match what is in these graphs. A simple no would have sufficed.

Don't take my word for any of this. Wait for Q1 and see for yourself

If the numbers aren't close, then why would the price of R800 have go down, as 003 says?

As an example to answer your question. The original MSRP of the GTX280 was very high, quite a bit higher than the 4870. Although it was faster than the 4870, most people were not willing to pay the significant premium and Nvidia was at one point forced to lower prices to a level at which they were actually losing money with each GTX280 sale. That is why they quickly responded with the GTX285 which slightly bumped up performance and significantly lowered cost of manufacturing.