NVIDIA GF100 & Friends speculation

[mczak]

It does.

Maybe he looked at the techreport result, where it failed to reach full speed performance (though still notably more than half speed): http://techreport.com/articles.x/19934/6
Maybe they used large textures which didn't fit into caches - techreport at least thinks it fails to reach its potential due to memory bandwidth and/or smaller L2 cache compared to GTX580.
(And I was long saying nvidia got it backwards, full speed fp16 isn't really needed for GF104 as this has higher tmu/alu ratio (and less bandwidth per tmu too) but GF100 should have had it. If this wasn't a bug on GF100 it looks like a major oversight to me.)

 

[Picao84]

It does.

Oh ye you are right again

Maybe he looked at the techreport result, where it failed to reach full speed performance (though still notably more than half speed): http://techreport.com/articles.x/19934/6

Exactly! It was this. Thanks!

 

[mczak]

Why would a GF104 or newer version of it go against a Cayman Pro, Cayman Pro and XT's are AMD's top end GPU's, they would go againt the top two nV GPU's.

Well if rumors of Cayman being ~360mm² are true, you could easily see why nvidia might want to push the ~330mm² GF104 (or make that GF114 if you want) against it instead of the ~520mm² GF110, at least against the Pro... So in their dreams they'd have a GTX580 sitting above Cayman XT, and some GF110 derivative against the XT, with full chip GF104 taking care of the Pro. But as said, I don't think it's going to work - I don't know if there's any credit to the Cayman die size rumors neither, obviously if it's going to be significantly larger it shouldn't be such a big problem.

 

[Razor1]

only if they go into a price war though, will that become a concern. Even last gen the price war really didn't happen.

Just be interesting to see the efficiency increases to Cayman. Has there been any possible bottlenecks in their current architecture that would have caused AMD to shift from a 5D to 4D ALU structure?

 

[CarstenS]

My guess: physical limits on more shrinkage goodness. And of course, if you have run enough simulations and gathered enough data points, you can make pretty educated guesses, what the trade off from 20% less peak MADD throughput will net you compared to the higher utilization rate in each core and the saved die area (thus your ability to increase the number of VLIW cores).

AMD seems to consider this a win, or they wouldn't have done it in the first place. And since they will be having the best source data available to run these simulations, I will agree until proved wrong.

 

[Razor1]

yeah that sounds more like what I had in mind, I don't think its going to cause "increased efficiency", its just a baby step to allow higher clocks. Shrinkage goodness, the amount of ALU's being cut don't really take up that much space .

 

[Alexko]

yeah that sounds more like what I had in mind, I don't think its going to cause "increased efficiency", its just a baby step to allow higher clocks. Shrinkage goodness, the amount of ALU's being cut don't really take up that much space .

It's about perf/mm², and perf/W. Taking out the T unit that was seldom used, at least not in conjunction with all the other units, and modifying the latter to enable them to gang up for transcendentals should save a few mm². It also means less transistors leaking power while doing nothing at any given time.

This should allow AMD to add 10~20% more SIMDs for a given area, and increase perf/W in top of what Barts already brings to the table.


PS: GF104 (GF114?) is 367mm², and I would imagine that Cayman is a lot closer to that than to GF100/110.

 

[Ancient]

Lower power consumption? Maybe in furmark because of the throttling. In Crysis Warhead it consumes ~10W more than HD5970 and performs significantly slower in every playable setting.

Now lets see Stalker CoP, the most important test :smile:

http://www.computerbase.de/artikel/...tx-580/16/#abschnitt_stalker__call_of_pripyat

 

[Man from Atlantis]

new tech demo up
aliens vs triangles
http://www.nvidia.com/object/cool_stuff.html#/demos/2508

 

[Razor1]

It's about perf/mm², and perf/W. Taking out the T unit that was seldom used, at least not in conjunction with all the other units, and modifying the latter to enable them to gang up for transcendentals should save a few mm². It also means less transistors leaking power while doing nothing at any given time.

This should allow AMD to add 10~20% more SIMDs for a given area, and increase perf/W in top of what Barts already brings to the table.


PS: GF104 (GF114?) is 367mm², and I would imagine that Cayman is a lot closer to that than to GF100/110.

At the end of the day, performance/mm only concerns the companies when going into a price war.

Barts has nothing to do with anything Cayman. Barts is a cut down version of Cypress with increased clocks and mild modifications. It was a good trade off for a new mid range product where the power usage didn't exceed cypress while dropping the size of the chip and keeping the performance close to Cypress. Cayman is going to be a larger chip, AMD won't have the same increase in frequency levels without a consequence of increasing power usage. It will all come down to increasing efficiency of the new design and clock speed. Anyhow this should be in the AMD thread.

 

[Arty]

Reading my fifth review, the general consensus is "What GF100 should have been .."

20% over Cypress overall isnt much. I hope Cayman is more than a handful rather than just enough to pip this recovery chip.

 

[trinibwoy]

Amazingly, it's actually a decent product too. This feels more like the launch of the 8800GTX than Fermi.

That sorta depends on what Cayman does.

new tech demo up
aliens vs triangles
http://www.nvidia.com/object/cool_stuff.html#/demos/2508

Cool, will give that a whirl later.

 

[LordEC911]

Reading my fifth review, the general consensus is "What GF100 should have been .."

20% over Cypress overall isnt much. I hope Cayman is more than a handful rather than just enough to pip this recovery chip.

You mean ~20% over GTX480? Seems like ~15% in the majority of cases.
Which means a good ~25-35% over Cypress.

 

[TKK]

Reading my fifth review, the general consensus is "What GF100 should have been .."

20% over Cypress overall isnt much. I hope Cayman is more than a handful rather than just enough to pip this recovery chip.

Dunno which reviews you've read, but it's more like 30% over Cypress from what I've gathered.

But even then, I agree this isn't very much considering that the 580 achieves this with ~55% more die area and ~100W more power consumption under load.

 

[Arty]

Dunno which reviews you've read, but it's more like 30% over Cypress from what I've gathered.

But even then, I agree this isn't very much considering that the 580 achieves this with ~55% more die area and ~100W more power consumption under load.

I like Damien's reviews and from his numbers is anywhere from 15-25% so I averaged it to 20%.

 

[joker454]

Anyone know if the 580gtx supports pap, so that it can bitstream hd audio streams over hdmi when playing blu-ray movies?

 

[Rangers]

In the end, a $520 product just doesnt have a lot of relevance for most users though.

Though what dose these days, as nothing challenges any of these cards anyway.

But my point is it needs to be in the $200-$300 sweet spot to be relevant to most users, and closer to 200 frankly. Both AMD and Nvidia massive fail there.

maybe if cayman gets closer to 300 it will do better in that dept.

 

[ZerazaX]

In the end, a $520 product just doesnt have a lot of relevance for most users though.

Though what dose these days, as nothing challenges any of these cards anyway.

But my point is it needs to be in the $200-$300 sweet spot to be relevant to most users, and closer to 200 frankly. Both AMD and Nvidia massive fail there.

maybe if cayman gets closer to 300 it will do better in that dept.

If the 6950 is placed there and it's indeed faster than Cypress XT by a good chunk as some people are suggesting, then it could be the winner

 

[cal_guy]

Anyone know if the 580gtx supports pap, so that it can bitstream hd audio streams over hdmi when playing blu-ray movies?

Sadly no.

NVIDIA doesn't carry over the bitstreaming-over-HDMI (Dolby True HD and DTS-HD) goodness from GTX 460, however.

http://www.hexus.net/content/item.php?item=27307&page=3
http://www.anandtech.com/show/4008/nvidias-geforce-gtx-580/4

 

[Alexko]

At the end of the day, performance/mm only concerns the companies when going into a price war.

Yes and no. Obviously the first effect is on their margins, but at some point, if the chip is cheaper to make, the card will retail for less money, if there is indeed a price war.

And obviously, if you're going to make a huge chip, perf/mm² matters because you can't go over 600mm² on TSMC's process, so improving perf/mm² is the only way to improve performance when you're close to that limit.

Barts has nothing to do with anything Cayman. Barts is a cut down version of Cypress with increased clocks and mild modifications. It was a good trade off for a new mid range product where the power usage didn't exceed cypress while dropping the size of the chip and keeping the performance close to Cypress. Cayman is going to be a larger chip, AMD won't have the same increase in frequency levels without a consequence of increasing power usage. It will all come down to increasing efficiency of the new design and clock speed. Anyhow this should be in the AMD thread.

The point is that Barts enjoys higher perf/W than Cypress, so improvements were made in that area, and those improvements are likely to carry over to Cayman. But yes, wrong thread.