Poll: GT300 VS. HD5870

[Tempered81]

We all know that Nvidia's GT200 was about 10-20% faster than ATi's RV770 last summer. What kind of performance do you expect from Nvidia's GT300?



Edit (9/29): I've just recently read that nvidia was supposed to have a competitor to the 5870 around aug/sept. however, after realizing that this part wouldnt compete with cypress or hemlock, it was scrapped and full attention diverted to GF100/fermi. the interim part was supposedly GT216/212/218 that had increased ram, bandwidth, shader count & speeds of the GT200 on a 40nm die. This evidently resulted in only a slightly better performing part compared to the already spectacular GTX285 gt200b. Now that full attention is focused on GF100 (GT300), they are rushed to bring it to the market by the end of the year. It IS a completely new radical part and no the same as 8800gtx - > gtx280 or hd4870 -> hd5870. This is why it's so far behind, it's sitting in the labs. No telling if all the charlie/faud rumors were true about yeilds and such, but the important part is that its a true redesign.

Now for the rumors. The chip IMO, is ~590mm, 3bln tran, with 512MIMD SP's, 40nm, 5ghz ddr5, 384bit bus, 700/1700/1250. Those are the most notable changes, and also that it has a huge focus on GPGPU computing, opencl, phsyx & CUDA. The rest of the particulars are unknown, but if these rumors are true it is going to be VERY FAST, AND VERY LATE.

So late to the party, but much faster. I originally voted "remain the same" but i would like to change it to "widen in favor of the GT300" because of these rumored specs, and that it is taking so long to produce. By the time it's out there will be competition from 5870X2 and a possible 5890.

 

[Barfo]

About the same as last generation.

 

[BTRY B 529th FA BN]

same

 

[alcoholbob]

About the same as 4870 vs GTX280 I would say. The thing is nvidia cards aren't inherently faster but have more market support. On purely 3rd party tech demos the difference is far less.

 

[crisium]

http://flimmr.passagen.se/movi...y_guy_aladdin_4.action

 

[Zap]

Not yet available versus out of stock?

 

[AzN]

Nvidia might be a tad slower or even.

 

[happy medium]

gt300 will avg 22% faster and cost 27% more

 

[RussianSensation]

I believe GT300 will be faster than a single-gpu 5870 because:

1) NV is making a larger die-GPU and will therefore have to sell it at higher prices in order to recoup the higher manufacturing costs of the larger die. To sell their GPU at higher prices, it will have to offer faster performance.

2) NV's strategy right now is to produce a large monolithic gpu to compete with ATI's dual medium sized gpus. Therefore, they will aim for 1 GT300 to be significantly faster than ATI's single card as that is their goal in the beginning (and downscale GT300 for lower end parts)

3) ATI's claim that its memory bandwidth is not a limiting component may be true for most resolutions. However, where you really need improved performance is 2560x1600 8AA. NV has an opportunity to give 220-250+ bandwidth vs. ATI's 155. After all, the memory bandwidth in the GTX 280 is equal to 140.8 already. I doubt NV is staying in the 150 range again!

4) This is a brand new architecture from NV. Last time this happened, 8800GTX easily doubled the performance of 7900 series. Since 5870 isn't doubling the performance of GTX 280, NV again will likely beat ATI's card.

5) NV's 8AA performance is inferior on the GTX285 right now compared to ATI's 4/5x series cards. Improve efficiency here and you are already 20% closer to 5870!

6) Since NV is not going to compete on price, strategicaly NV will have to be faster since they are coming out 2nd out of the gate, unless they want another 5900 series fiasco.

 

[uclaLabrat]

I think ATI already had cards that were closer to DX11 and have been for a while, hell, the 4800 series already had a tesselator and ATi has been experimenting with them for years. NV might be able to bring up a card that can compete with or even best 5870, but I'm starting to think they're going to hemorrhage cash in doing so.

 

[SlowSpyder]

I see more of the same as the GTX2x0 vs. 48x0 cards, but that's really just me taking a guess. We really have no idea how fast or slow the GT300 will be, Nvidia has been pretty tight-lipped.

 

[Idontcare]

Originally posted by: RussianSensation
I believe GT300 will be faster than a single-gpu 5870 because:

I voted same way for similar reasons. My thinking on GT300 probably more simple-minded though.

If Cypress is basically 2xRV770 then all Nvidia would have to do (at a superficial level) is simply double their GT200 in similar vein and they'd have a product that should hold the same performance delta over cypress as its half-sized sibling holds over RV770.

But Nvidia went and changed the architecture, that opens-up the performance opportunity to the upside (innovation does that) while minimizing the prospects of performance actually degrading (it can happen, but not with equal probability of innovation to the upside, progress is just that, on average).

Of course saying it is faster is open-ended...makes no constraint on power-consumption scaling nor on price. And at the other side of the business equation, it makes no constraint on gross margins or R&D budget efficiency.

 

[MODEL3]

The same gap.
GT300=32ROPs.
GT300=512bit memory controller=double 5870 bandwidth. (1,2GHz GDDR5)
GT300=higher TU / ROP ratio & higher Flops/s / ROP ratio than GT200.

I think overall the same gap.

 

[Idontcare]

Originally posted by: MODEL3
The same gap.
GT300=32ROPs.
GT300=512bit memory controller=double 5870 bandwidth. (1,2GHz GDDR5)
GT300=higher TU / ROP ratio & higher Flops/s / ROP ratio than GT200.

I think overall the same gap.

Does that mean you are giving the new architecture no presumed performance benefits over GT200 architecture, or you think GT300 architecture is the same/similar as that of GT200 insomuch as Cypress architecture is the same/similar to that of RV770? (plus all the DX11 circuits to make it work of course)

 

[lavaheadache]

for all we know we haven't seen all the hd 5800 series has to offer. I'm sure by the time gt 300 ccomes out we will have an hd5890 to play with. =)

 

[zagood]

I believe it will be a rebrand of the 9800GTX+.

Some guy named Charlie told me.

 

[MODEL3]

Originally posted by: Idontcare

Originally posted by: MODEL3
The same gap.
GT300=32ROPs.
GT300=512bit memory controller=double 5870 bandwidth. (1,2GHz GDDR5)
GT300=higher TU / ROP ratio & higher Flops/s / ROP ratio than GT200.

I think overall the same gap.

Does that mean you are giving the new architecture no presumed performance benefits over GT200 architecture, or you think GT300 architecture is the same/similar as that of GT200 insomuch as Cypress architecture is the same/similar to that of RV770? (plus all the DX11 circuits to make it work of course)

I mean the same gap, like:

The gap between GT300 and 5870 will be the same,
as the gap between GT200 and RV770.

For me GT300 will have in relation with GT200:
1,1X pixel rate
1,75X texel rate
2,5X Flops rate
2,2X vertex rate

I wrote the day of the 5870 launch that the 5870 results are strange.
And that according to my perception, there are other reasons except bandwidth limitations and driver maturity, that the 850MHz 5870 hasn't doubled its performance in relation with a 850MHz 4890.

Usually when a GPU has 2X the specs of another GPU the performance gain is 2X (of cource i am not talking about games with engines that are CPU limited or engines that seems to scale badly or are poor coded for example)
There are many examples in the past that we had 2X performance gain with 2X the specs. (not in all the games, but in many games)

From the tests that i saw in Ryan's review and from my understanding of the 5870 architecture in general, i think there are 2 more reasons that 5870 performs like that.

The day of Ryan's review, i wrote to the forums the additional reasons that i think the 5870 performs like that, but nobody replied me.

I wrote that probably 5870 has:

1.Geometry/vertex performance issues (in the sense that the classic vertex unit cannot generate 2X geometry in relation with 4890) (my main assumption)

or/and

2.Geometry/vertex shading performance issues (in the sense that the geometry shader [GS] cannot achieve 2X speed in relation with 4890)(another possible assumption)

My assumption is that ATI though that many of the future DX11 games will use vertex techniques based on the tesselator unit.

I guess there are synthetic benchmarks that have tests like that (pure geometry speed, and pure geometry/vertex shader speed, in addition with the classic pixel shader speed tests) so someone that has a 5870 can see if my assumption is true.

 

[Idontcare]

Originally posted by: MODEL3
I wrote the day of the 5870 launch that the 5870 results are strange.
.snip.
The day of Ryan's review, i wrote to the forums the additional reasons that i think the 5870 performs like that, but nobody replied me.
.snip.

Don't take it personal when that happens, threads were zooming like mad and it is quite easy for a good post here and there to get buried quickly among lots of other equally interesting/compelling posts.

Originally posted by: MODEL3
I wrote the day of the 5870 launch that the 5870 results are strange.
And that according to my perception, there are other reasons except bandwidth limitations and driver maturity, that the 850MHz 5870 hasn't doubled its performance in relation with a 850MHz 4890.

Usually when a GPU has 2X the specs of another GPU the performance gain is 2X (of cource i am not talking about games with engines that are CPU limited or engines that seems to scale badly or are poor coded for example)
There are many examples in the past that we had 2X performance gain with 2X the specs. (not in all the games, but in many games)

From the tests that i saw in Ryan's review and from my understanding of the 5870 architecture in general, i think there are 2 more reasons that 5870 performs like that.

The day of Ryan's review, i wrote to the forums the additional reasons that i think the 5870 performs like that, but nobody replied me.

I wrote that probably 5870 has:

1.Geometry/vertex performance issues (in the sense that the classic vertex unit cannot generate 2X geometry in relation with 4890) (my main assumption)

or/and

2.Geometry/vertex shading performance issues (in the sense that the geometry shader [GS] cannot achieve 2X speed in relation with 4890)(another possible assumption)

My assumption is that ATI though that many of the future DX11 games will use vertex techniques based on the tesselator unit.

I guess there are synthetic benchmarks that have tests like that (pure geometry speed, and pure geometry/vertex shader speed, in addition with the classic pixel shader speed tests) so someone that has a 5870 can see if my assumption is true.

And see, the conspiracy side of me thinks "and what if 5870 performance was intentionally sandbagged by the drivers in hopes of pulling a one-two punch on NV after they release GT300 and set their prices and then AMD releases a driver update for 5870 that unleashes all that pent-up processing capability we know is under the hood given the specs and what we know RV770 could do?".

Possible but not probable, still though that would be the twist of 2009 everyone would talk about for years afterwards if AMD played that card.

But your right, somewhere there is some missing performance yet to be extracted from Cypress. It just doesn't make sense otherwise.

 

[ronnn]

looking for the gt300 will likely match the 5890 when it is readily available option.

 

[OCGuy]

Originally posted by: ronnn
looking for the gt300 will likely match the 5890 when it is readily available option.

So a completely new arch and adding SIMD and DDR5 will match a doubling of specs of the 4890?

Interesting prediction. Hopefully we get to see who is right before Xmas.

 

[MODEL3]

Originally posted by: Idontcare
Don't take it personal when that happens, threads were zooming like mad and it is quite easy for a good post here and there to get buried quickly among lots of other equally interesting/compelling posts.

Not at all.
I just meant that i didn't get reply and i didn't know what other people think about this possibility.
If i remember i replied to keys.
I was something like 5-6 pages behind the last post and after i saw a Keys post i replied to him, i check then the remaining pages and the conversation had moved to personal responses etc... (not the best time to reply...)

Originally posted by: Idontcare
And see, the conspiracy side of me thinks "and what if 5870 performance was intentionally sandbagged by the drivers in hopes of pulling a one-two punch on NV after they release GT300 and set their prices and then AMD releases a driver update for 5870 that unleashes all that pent-up processing capability we know is under the hood given the specs and what we know RV770 could do?".

Possible but not probable, still though that would be the twist of 2009 everyone would talk about for years afterwards if AMD played that card.

But your right, somewhere there is some missing performance yet to be extracted from Cypress. It just doesn't make sense otherwise.

I certainly hope this to be true (i have my own conspiracy book, lol), it would be good for guys that bought or are planning to buy a 58X0 card and it would be good also for guys that are planning to buy GT300 because NV will be forced to lower the price.

I'll check latter other tech site reviews for some geometry synthetic tests and i will post again if i can find any (not very optimistic...)

 

[Schmide]

I think it depends on how much power (wattage) nVida can design for. I find both companies about equal

complexity
-------------- = power usage
process

Cards of equal process perform about equal relative to power usage. One thing to back up Idontcare theory of future 5870s performance is its game power usage (110w) vs its artificial benchmark maximum (160w). There seems to be a lot of untapped power. Even pulling a page from Intel's book it is designed for thread level performance. Sans memory/cache miss-design there should be a fair amount of power left in the card if previous figures hold true.

 

[Extelleron]

I think GT300 will definitely be faster than the HD 5870, if it isn't then nVidia would be in big trouble. I don't know the exact numbers but the HD 5870 is only about 30-40% faster than the GTX 285 or sometimes even less. I highly doubt nVidia would be gunning for anything less than 50-60% or more improvement over GT200, meaning that unless GT300 is a complete flop it will be faster than HD 5870, perhaps by a large margin. I have a feeling that GT300 will be a far more revised architecture compared to GT200 than RV870 is to RV770.

Anyway I hope it is faster than the 5870. I'm really hoping to find a single card replacement that is as fast or faster than the SLI setup I have now. I thought 5870 might be that card but obviously not.

 

[sbuckler]

I think it'll be about the same gap for games. Where the GT300 will shine is gpu compute - this thing was designed to counter larrabee. It will fold at > 10 times the speed of a 5870.

 

[Qbah]

Same gap imo. nVidia was touting "optimized shaders" this round too and we all saw what happened. 2 8800GTs in SLI (112x2=224 shaders) were faster than a GTX280 (240 shaders). And that's not even twice as fast due to SLI scaling