The Fermi Thread - Part 3

[GoldenTiger]

Where did you saw clock's reference?

As GTX480 power draw was about 283W, it could be 512 Cuda Cores card with 700 mhz.

So the final GTX480 will have 480 shaders (one main unit disabled) and clocks upped to 675, so I'll pretty have the performance of the 512 shaders, 600MHz ES I tested, that being 3.6% faster on average then the 5870.
So judging by the price, power & heat - overall it's not worth the trouble.

One of the posts was already deleted which a user comments on in the thread... but he clearly said the clocks I mentioned before as what his ES came with. The one with that quote is top post on page 4 of the linked thread.

[Chickenfeed]

Virtually all 5870 (850Mhz) can max out overdrive at 900. Most cards do around 920-935 @1.15V.
Memory on 5870 (1200) overclocks to 1300. I doubt AMD will bother trying to source some rare uber high end GDDR5, since most of performance improvement is from core overclocking.

"5870 OC" is probably gonna be something like 950/1250 2GB

I agree with your estimations of a 5870OC / 5890 however where the hell did you pull this 920-935@1.15V number from I've used 4 different 5870s ( one of which is my own ) and not a single one claims such a feat ( I doubt they'd do 850 at 1.15 for that matter ) AMD have even upped the stock voltage from 1.168-1.172 on the reference bios to improve stability.

If anything they needed 1.2 (mine needs 1.237) best case to get 925. Now this isn't to say with improvements it couldn't be done but this just doesn't seem indictive of reference launch cards ( the cards I've used were all revision 1 with initial bios, all tested in 2009 ) Sure a 4 card sample isn't huge however they were all from different vendors and all consistent in how much voltage they seemed to play nice with (or not). Anyways just a pety thing to kick dust up at, just found this statement to be a tad up in the air is all.

I'd love to see a 5890 with at least 950/1300 clocks coupled with 2GB of memory. That should offer a 5-15% performance bump over a 5870. Going by the leaks so far, this would be more than enough to challenge the 480s. Given those supposive 512sp leaks were at 2560x1600, in some cases with 8x AA, im shocked the 5870 doesnt fare worse with its vram handicap. Should bode well for a future ATI refresh.

[Marios]

Is GeForce Fermi asking ppl to invest their money to support Nvidia's cloud server future?

[japamd]

One of the posts was already deleted which a user comments on in the thread... but he clearly said the clocks I mentioned before as what his ES came with. The one with that quote is top post on page 4 of the linked thread.

That explains it. Cheers

[SKYMTL]

All i know, reviewers like SKYMTL and Benchzowner have hinted that GF 100 cards will have AWESOME gaming performance compared to competitor, even on todays games, that GF 100 will stomp RV 870 to the thrash bin where it belongs, so i think i will put more weight on their cues rather than Charlie the satan himself for this launch occasion.

I believe BZ and I said it has the POTENTIAL to achieve this. Whether or not it will accomplish its goals is yet to be seen.

What we BZ and I have said again and again is that people should WAIT before jumping too far onto a bandwagon claiming anything extreme.

As for Charlie's Dirt 2 story, not many reviewers are benchmarking with the demo version of the game and with the retail version you can physically force DX11 through the config file.

[Marios]

No potential at all. I only see an inefficient premature design intended to be a scalable cloud server product, disguised into desktop GeForce Fermi.

[eric66]

I believe BZ and I said it has the POTENTIAL to achieve this. Whether or not it will accomplish its goals is yet to be seen.

What we BZ and I have said again and again is that people should WAIT before jumping too far onto a bandwagon claiming anything extreme.

As for Charlie's Dirt 2 story, not many reviewers are benchmarking with the demo version of the game and with the retail version you can physically force DX11 through the config file.

.........and here goes skymtl ruining your dreams

[LiquidReactor]

By pure math...

32 shaders would be about 6.5% of the shading, lost...

but you'd gain about 17% clock rate... on the cores alone... and a good chunk more memory bandwidth.

You'd also be speeding up the ROP's whereas more cores don't do that alone...

Additionally the TMU's would be sped up as those are tied to the core clock and not shaders as far as I'm aware...

He also said the newer drivers added a couple % performance...

Also worth noting, this guy claims his Engineering Sample card had 512 shaders @ 600mhz, 1200mhz shader, and 2800 (GDDR5) RAM: full specs should be 480 shaders @ 700mhz, 1400mhz shader, and 3600-3700 (GDDR5) RAM.

So my guess is that with those #'s considered, it would be maybe 16-17% better performance for 700mhz @ 480 cores, vs. 600mhz @ 512 cores, taking into account the memory frequency as well. Couple that with a couple of extra percent from those tests adding 3% (he said a few) and we might see a graph looking a bit more favorable. Still, if this guy's accurate, it wouldn't be enough to make it what I would call a clean win for nV here.

If GDDR5 is at 2800Mhz that should be seriously crippling Fermi's performance.

[Marios]

GTX480 has 3 billion transistors doing what? 2.812 billion are enabled and supposedly working.
Though the originally designed capabilities are disabled on GeForce Fermi.

* NVIDIA OptiX engine for real-time ray tracing
* NVIDIA SceniX engine for managing 3D data and scenes
* NVIDIA CompleX engine for scaling performance across multiple GPUs
* NVIDIA PhysX 64-bit engine for real-time, hyper-realistic physical and environmental effects

What's left? The power consumption.

[jaredpace]

http://www.atomicmpc.com.au/Gallery/...-guide.aspx/15

[zerazax]

niec pics! finally a clear one of the ihs and pcb

[Mk]

damn from last leaked benchmarks i think Fermi is disappointing

[570091D]

GTX480 has 3 billion transistors doing what? 2.812 billion are enabled and supposedly working.
Though the originally designed capabilities are disabled on GeForce Fermi.

* NVIDIA OptiX engine for real-time ray tracing
* NVIDIA SceniX engine for managing 3D data and scenes
* NVIDIA CompleX engine for scaling performance across multiple GPUs
* NVIDIA PhysX 64-bit engine for real-time, hyper-realistic physical and environmental effects

What's left? The power consumption.

thank you for the enlightening post. btw, power consumption for the 480sp part should be around 250 watts.



does anyone know what this connector is for?

[Olivon]

No chip is designed to operate at 100C+.
You're physically killing the silicon in this case. There is no way around that.

Believe me or not but that the truth and W1zzard is not a noob too ;p

HD 4870 was really hot too especially with VRM's. Just have a look :

http://www.hardware.fr/articles/751-...n-hd-4870.html

Back to the subject, it seems that Fermi 1st generation isn't a great deal.

Hope ATI prices go down though, I don't want to pay a HD 5850 @ 250€, more 150-200€. Competition FTW

does anyone know what this connector is for?

nVidia greenlight logo if I remember well

[Mk]

thank you for the enlightening post. btw, power consumption for the 480sp part should be around 250 watts.



does anyone know what this connector is for?

Nvidia glowing Logo

[570091D]

nVidia greenlight logo if I remember well

Nvidia glowing Logo

ahhhh, thanks!

[Ramon Zarat]

Nice thread. These are my thoughts and speculations:

ATI:

The 5000 series 40nm process is now refined enough to produce a 1GHz chip that coupled with 2GB or RAM will stay within the PCIe electrical spec and deliver a catastrophic 1 - 2 punch to Nvidia. That would be a formidable marketing coup for ATI to be first to market a 1Ghz chip *AND* take over the 480.

ATI will do it. Simply for the market dominant position and bragging rights of having the fasted single GPU. No matter what's the manufacturing cost of that hypothetical card, they will be able to ask 500.00$+ for it anyway and push Nvidia prices down when they already have no margin left on these GF100 cards. If in the next couple of weeks it's technically feasible for ATI to jump at Nvidia's throat for the kill, why wouldn't they do it?


Nvidia:

Out of the chipset market, bump-gate scandal, wood-screw puppy, multiple re-bagged cards, miserable execution of the 6 month late Fermi, their general attitude problem and the fact they have pretty much alienated everyone in the industry confirm 1 thing: Nvidia is way too full of itself and need an ego / reality check, fast.

The gaming side of the Fermi equitation is an after thought at best. It's a GPGPU with a bolted-on gaming chunk. I really hope for their sake that the double precision floating point power, ECC RAM and CUDA SDK with C++ capability will be enough to save this monster of a chip. Putting 2 GF100 chip on a single card is out of the question, the clocks are already at their maximum. Short of introducing a 512sp version very quickly or solely rely on driver optimization, there is nowhere to go up for them.

What you'll see on the 26 is pretty much what you'll get for the next year or so. Nvidia won't be a force to be reckoned with in the gaming market until Fermi die shrink to 32nm. Problem is, they will have to face the HD 6000 way before that. Nvidia won't be able to absorb blow after blow like that for many more quarters. There's a limit to keep surfing on the wave of your dominant market position without sinking.


Conclusion:

It's all sad because based on the benchmark already available it's clear where we are going. I would have liked Nvidia to be not only competitive, but dominant in order to keep the prices low. I think they need to restructure, rebuild their bridges and corporate image. The first thing Nvidia needs to do is to fire Jen-Hsun Huang before it's too late.


Ramon

[kemo]

Believe me or not but that the truth and W1zzard is not a noob too ;p

HD 4870 was really hot too especially with VRM's. Just have a look :

http://www.hardware.fr/articles/751-...n-hd-4870.html

B

Yeap , my HD4850 went up to 110C Shader and 116C MemIO while running OCCT stress test for few minutes and still running till today with no problem at all

Is that is part of the Metal heatsink showing on the surface , if so then adding a 120mm will get some nice temps and no need to run the stock fan at high speed

[ToxAvenger]

.

Charlie is that you? Na, that's over the top even for him. I think we should all at least wait until we see legitimate tests before we start tossing dirt on Fermi. Edit - the small connector is for the fan most likely. Or the glow logo

[zalbard]

And this... GTX 480 sample with 512 SPs and possible 600Mhz (1200Mhz).

This looks about right, I guess... Not impressive, but decent enough.

No potential at all. I only see an inefficient premature design intended to be a scalable cloud server product, disguised into desktop GeForce Fermi.

Inefficient, yes. But it has potential. There could be quite a few interesting desktop applications, heavy tessellation environments being one of them.

GTX480 has 3 billion transistors doing what? 2.812 billion are enabled and supposedly working.
Though the originally designed capabilities are disabled on GeForce Fermi.

* NVIDIA OptiX engine for real-time ray tracing
* NVIDIA SceniX engine for managing 3D data and scenes
* NVIDIA CompleX engine for scaling performance across multiple GPUs
* NVIDIA PhysX 64-bit engine for real-time, hyper-realistic physical and environmental effects

What's left? The power consumption.

There is no real-time ray tracing.
The other stuff you mentioned is pure marketing...

Believe me or not but that the truth and W1zzard is not a noob too ;p

HD 4870 was really hot too especially with VRM's. Just have a look :

http://www.hardware.fr/articles/751-...n-hd-4870.html

I never said he was a noob.
And do not confuse VRMs with GPU chips.
VRM has a different structure and can operate at higher temperatures. Most GPU VRMs are rated for 120C. But the GPU chips are definitely not.

[DosDuoNo]

Days To Go

[LordEC911]

Lower binned GDDR5? It is clocked slower but I don't think it is a lower bin.

I can "think" its binned lower and is cheaper, same logic being used...

Until somebody posts what vram is used what basis of comparison is there without a spec sheet at the least.

We can both assume the other is wrong but that doesn't make either of us right.

I try to not be wrong... I don't assume things often.
Also your logic =| my logic. So no, the same logic isn't being used.
Is this good enough for you?





Samsung GDDR product sheet
K4G10325FE HC04
K - Samsung
4 - DRAM
G - GDDR5 SGRAM
10 - 1G 8k/32ms
32 - x32
5 - 8Banks
F - 7th Gen
E - 6th Gen

H - GDDR 170FBGA
C - Commercial Normal
04 - .4ns (5Gbps)

So the same memory ICs as 5870, so if the MCs aren't limiting frequency, ~250Gbps is possible w/ a memory clock of 5.2ghz.

And lower clocked DDR5 is cheaper, obviously.... and I don't know why they would use fast DDR5 and downclock it (if that is what you are implying).

Maybe they purchased the high binned GDDR5 before they knew they messed up the MC(Charlie) or that they couldn't use the full clocks due to a power/TDP wall
and decide to leave it on there to allow users/AIBs to overclock them.

[Ramon Zarat]

Charlie is that you? Na, that's over the top even for him. I think we should all at least wait until we see legitimate tests before we start tossing dirt on Fermi. Edit - the small connector is for the fan most likely. Or the glow logo

LOL.... No, it's not "Charlie"!

I'll admit, I *HATE* corporate arrogance. From that perspective, Nvidia is very easy to hate and it shows in my "editorial" position. Another example: Followed Apple business practice lately? I just can't stand it. In Apple's case, on top of the ego-maniacal, self appointed emperor Steve Jobs's tyranny, you get overpriced and somewhat palatable products. I'm not saying ATI/AMD is all white. It's just hard to deny that they're just not in the same league when it come to business practice and corporate attitude.

But this thread is about the new and upcoming Fermi, isn't it? Then lets put my logic to the test with arguments that disprove my theory:

FACT: Out of the chipset market
FACT: Bump-gate scandal
FACT: Wood-screw puppy
FACT: Multiple re-bagged cards
FACT: Miserable execution of the 6 month late Fermi
FACT: General attitude problem
FACT: Alienated everyone in the industry
FACT: Putting 2 GF100 chip on a single card is technically impossible at 40nm
FACT: At 700Mhz, the clocks are already at their maximum because of heat.
FACT: Driver optimization and a 512sp version are the only way to gain speed on Fermi. What you see on the 26, except a 14% bump (7% sp 7% drivers), is what you'll get until 32nm.
FACT: Unless a VERY dramatic turn of even happen, ATI will launch the HD 6000 series in Q2 2010, WAY before Nvidia can shrink it's Fermi to 32nm.

Do you deny any of those previous statement? If so, please explain.

My opinion: Fermi gaming capability is an after thought.
My opinion: Fermi is first and foremost a GPGPU with a bolted-on gaming section.
My opinion: DPFP, ECC RAM, C++ and superior SDK are the real differentiator that will make or kill Fermi in the GPGPU market.
My opinion: If Fermi fail to sell significantly as a GPGPU and to a lesser extent as a gaming card, that can only bad news for Nvidia and all of us.
My opinion: Nvidia can't afford to miss the boat like that too many time in a row without seriously impacting it's long term viability.

I would like to hear you opinion about those specifics point too. I'm fully aware that no official benchmark are available and announcing Fermi's death right now would be premature. But you have to admit, based on what we already know, it's gona be FAR from the 5870 killer it was supposed to be. In fact, it's clearly going to be an attainable target for a 5870 refresh.

Lets put it that way: A cheaper 5870 that consume 62 Watt less under load, and even much less at idle, run cooler in your casing, with less noise from the fan and run your game at 85-90% of the speed of the 480 OR a 5890 that run same/faster, with more RAM than 480 for the same price as a 480. Which one you choose? 5870, 5890 or the 480? I would take the 5870 and overclock the hell out of it if and when I need it. That is, unless Fermi end up 40% faster than the 5870 instead of just 10-15% *AND* if they fire Jen-Hsun Huang. Then I'll buy TWO 480 and put them in SLI. LOL.


Ramon

[[XC] Oj101]

It's OK because it's ATI, not Nvidia. It's also OK that eventually, the VRMs on that card will pop when running Furmark. Again, because it isn't nvidia.

QFT. 99% of NVidia haters aren't even sure why they hate the company.

[eric66]

http://www.atomicmpc.com.au/Gallery/...-guide.aspx/15

disassembling ? is that what reviewers do all day with fermi ? wonder what they want to find inside the card gold or something