NVIDIA Fermi: Architecture discussion

[XMAN26]

What kind of baloney is that? I'm hoping Fermi knocks it out of the park and would love to see some benchmarks.

Paranoid much?

Let's not make this thread about ATI vs. NV (again) but about cool new tech.

Have you been reading PSUs posts? Almost all of them are, numbers or Fermi doesn't perform or something along those lines. Personally, I'm sick of seeing it, Nvidia has no need or reason to actually show numbers as the launch is about 1.5m away and they'll let the numbers speak for the card then.

 

[Mize]

Have you been reading PSUs posts? Almost all of them are, numbers or Fermi doesn't perform or something along those lines. Personally, I'm sick of seeing it, Nvidia has no need or reason to actually show numbers as the launch is about 1.5m away and they'll let the numbers speak for the card then.

I just try to ignore those posts. I was only taking exception with your "everyone demanding benchmarks hates nvidia" comment. I'm dying for benchies and couldn't care less which vendor's board I'm using (I currently have two of each in my four home machines).

 

[digitalwanderer]

Why is it most people posting in this thread display such love for ATI or hate for Nvidia?

Because that is how it should be.

(Sorry, couldn't resist!)

 

[seahawk]

64 more SPs, higher clocks... Besides, memory chips don't draw a lot of power.

For over 50% more consumption? I doubt that. I would believe that they could reach something around 230W in worst case scenarios and actually require a 6+8 configuration to run their top version, but not that the card actually uses 300W.

 

[digitalwanderer]

Unless Nvidia made some unknown quantum leap in performance/sq mm in their design, the SIZE of the chip ROUGHLY equates to the power it uses if operating AT THE SAME PERFORMANCE LEVEL.

Or are you inferring Nvidia can design a chip the same size and performance as Cypress at the same process node, with 2/3 the tdp?

Maybe he was implying that nVidia can design a chip that is the same size and performance as Cypress with 2x the tdp...

 

[aaronspink]

I'm not saying that that's what the x2 will have. I did say that a TDP as high as 300w is expected for the X2 card and not the single chip version, as seems to be implied by the french site. Single chip version should be on the same level of GT200 in terms of TDP: 236w
If GT200 @ 65 nm, 576 mm2 and 1.4 billion transistors had a TDP of 236w, how can 3 billion transistors @ 40 nm, < 576 mm2 have a TDP of 300w ?

Because lets see...

2x the transistors
roughly the same voltages - ~10%

So we start out at roughly 382w just doing the 2x transistors and ~10% V reduction, assuming C and F stay the same, for the gross calculation.

One thing that people need to realize is that power scaling from process generations is rather weak at this point as the big hammer that is V is becoming less effective as we push close to the threshold voltages. In addition, C is scaling much weaker as well without fundamental process changes, which TSMC didn't make in order to push to 40nm as fast as they could.

 

[spigzone]

Performance-wise, you would need to isolate factors that can contribute to performance in a wide variety of scenarios. In ALU-limited situations where RV770 could put its math advantage to work, GT200 would have needed a lot more power to get a 45% performance lead.

You would need to cite the source of your numbers. I am reluctant to say that Fermi's additional 50% in transistor count is devoted solely to GPGPU. I would say a lot of its enhancements serve multiple purposes, and it has devoted quite a bit of room to texturing and ROPs in keeping with its higher transistor count.

Why a need to isolate factors? The madshrimp's link averages out the gaming performance of the various cards from across an array of tech site reviews.

GT200 was know to be at a cost disadvantage against the RV770. Nvidia always struggled with their profit margin on those cards. When the 4870 released it was a sucker punch to Nvidia's profit kidneys, ergo their huge rebates to the GT200 early adopters.

Every factor I am aware of increases that profit squeeze against Nvidia and in AMD's favor in the newest generation. If you are aware of any that don't, I'm all ears.

If Nvidia's chip is 50% larger than AMD's and has 50% more transistors and Nvidia finds it necessary to clock it as high as they can, which seems a near certainty at this point, how can it be that the power usage isn't roughly 50% higher? On precisely the same process node from the same fabrication company?

It all comes down to gaming performance. If Nvidia finds it necessary to push Fermi's clocks to compete with Cypress on cost/performance, then 50% more transistors on the same fabrication process = ~ 50% more power draw. Unless Nvidia has a design that provides vastly better performance/transitor than AMD. I have read of no such magic wand, have you?

 

[Alexko]

For over 50% more consumption? I doubt that. I would believe that they could reach something around 230W in worst case scenarios and actually require a 6+8 configuration to run their top version, but not that the card actually uses 300W.

Over 50% more? 300W is 33.33% more than 225W.

It can easily be explained by voltage and clocks alone, assuming a 10% bump on each:

225 * 1.1² (voltage) * 1.1 (frequency) = 299.475, and that doesn't even take the disabled SMs under account, but then again it doesn't take the number of memory chips under account either.

 

[spigzone]

Maybe he was implying that nVidia can design a chip that is the same size and performance as Cypress with 2x the tdp...

Thanks for the correction, implying it is.

 

[FrameBuffer]

Nah, only the french site is saying that. At least based on the translation provided.

There's no way that a single Fermi chip will suck that much power.

Hope you like hats .. or do you prefer crow ??

 

[Mize]

Oooh, it's just like before the Texas vs. Crimson Tide game! Taunts and all!

 

[Subtlesnake]

If Nvidia's chip is 50% larger than AMD's and has 50% more transistors and Nvidia finds it necessary to clock it as high as they can, which seems a near certainty at this point, how can it be that the power usage isn't roughly 50% higher? On precisely the same process node from the same fabrication company?

You were just given example of where Nvidia was able to be more power efficient on a per-transistor/die area basis (GT200b), so I don't see why it's guaranteed that Fermi will consume at least 50% more power than Cypress.

What Nvidia's efficiency in terms of performance per watt will be, is obviously going to depend on final performance, but if it performs like its specs say it should, (i.e a doubled up 285), then you'd expect it to be 50% faster as well, which could give Nvidia a performance per watt lead.

 

[trinibwoy]

Hope you like hats .. or do you prefer crow ??

Gotta agree with XMAN in one respect though, there are some who seem to be consistently hoping for / predicting bad news for Fermi. To the benefit of whom is the baffling question.

 

[Silus]

Because lets see...

2x the transistors
roughly the same voltages - ~10%

So we start out at roughly 382w just doing the 2x transistors and ~10% V reduction, assuming C and F stay the same, for the gross calculation.

One thing that people need to realize is that power scaling from process generations is rather weak at this point as the big hammer that is V is becoming less effective as we push close to the threshold voltages. In addition, C is scaling much weaker as well without fundamental process changes, which TSMC didn't make in order to push to 40nm as fast as they could.

Are the voltages and frequencies in Cypress much different than in RV770 ?

Because from RV770 (150w) to Cypress (190w), there's also two times the transistors. So you are either assuming NVIDIA is incompetent in what they've been doing since...ever...or they simply won't get that far in terms of power consumption.

 

[Ninjaprime]

You know, someone skilled of photochoppery and video-editing wizardry could just cut the video from CES and play it side by side with a video of a 5780 running the same benchmark as the same time, and you could see which one was smoother...

Of course, if they are running it on the tri-SLI setup there, you'd have to get a tri-crossfire setup to properly compare. 5970@ 5870 clocks + a 5870 in crossfire? That would be the best bet, I think?

 

[spigzone]

Good grief man, 4870 is on par with the GTX260 216 and slower than the GTX285 was is about 25-30% faster. The 4890 was redone and offers performance on par with the GTX285. But again, 45% more die space and yet as the other person pointed out, only 28% more TDP.

Why is it most people posting in this thread display such love for ATI or hate for Nvidia?

There's a difference arguing a point from sound physics and logical reasoning vs. arguing by waving a magic wand negating the laws of physics and logic. There is flatly no logical, physics based cost/performance advantage Fermi is known to have over Cypress. There are a number of logical, physics based cost/performance advantages Cypress is known to have or can be logically inferred to have over Fermi.

It's primarily a matter of reality based argument vs. a magic wand based argument, not a hated of Nvidia. My personal attitude toward Nvidia is subordinate to the logic and physics of my arguments ... therefore irrelevant.

 

[Silus]

You were just given example of where Nvidia was able to be more power efficient on a per-transistor/die area basis (GT200b), so I don't see why it's guaranteed that Fermi will consume at least 50% more power than Cypress.

What Nvidia's efficiency in terms of performance per watt will be, is obviously going to depend on final performance, but if it performs like its specs say it should, (i.e a doubled up 285), then you'd expect it to be 50% faster as well, which could give Nvidia a performance per watt lead.

Heh, the point being made by some, is that if it's bad it must be true and there are all sorts of explanations and if it's good, it's from an unreliable source and probably just BS. This, obviously happens on the other side of the fence too, which is expected.

But it's still quite funny to see that given the specs, being quite a bit faster than Cypress is a "day dream" or "not going to happen", but consuming vastly more power is "perfectly normal"

 

[trinibwoy]

Of course, if they are running it on the tri-SLI setup there....

The Maingear demo machine? It only had a single card.

There is flatly no logical, physics based cost/performance advantage Fermi is known to have over Cypress.

Hmmm, can you tell us again how sound physics and logical reasoning lead you to believe that Fermi will only be 30-40% faster than GT200? That's the comparison people seem to be shying away from. See I would be disappointed if Fermi at 600/1200 couldn't approach 2xGTX 285. Guess my standards are too high?

But it's still quite funny to see that given the specs, being quite a bit faster than Cypress is a "day dream" or "not going to happen", but consuming vastly more power is "perfectly normal"

I'm in the camp that expects pretty high power consumption just based on the sheer size of the chip and TSMC's issues with 40nm. But I also expect it to be much faster than Cypress as well for one simple reason: if it's not then it's less efficient than GT200 and that's hard to assume given what we know of the architecture so far.

 

[Ninjaprime]

The Maingear demo machine? It only had a single card.

Well, that makes it easy then! Someone have at it!

 

[Mize]

You know, someone skilled of photochoppery and video-editing wizardry could just cut the video from CES and play it side by side with a video of a 5780 running the same benchmark as the same time, and you could see which one was smoother...

Of course, if they are running it on the tri-SLI setup there, you'd have to get a tri-crossfire setup to properly compare. 5970@ 5870 clocks + a 5870 in crossfire? That would be the best bet, I think?

A lot of work if you don't know the settings they used.