videocardzAMD Radeon R9 290X confirmed to feature 64 ROPs

[BallaTheFeared]

Clearly can't be as effective as Tahiti ROPs if there are 64 of them with improved/more shaders on a larger bus and it's only 30% faster.

Can they?

 

[GaiaHunter]

Clearly can't be as effective as Tahiti ROPs if there are 64 of them with improved/more shaders on a larger bus and it's only 30% faster.

Can they?

Look at GTX680->Titan.
50% more ROPs, 75% more shaders, 50% more bandwidth and 75% more TMUs didn't deliver 50% more performance (I see 30% in TPU).

Why should 100% more ROPs, 11% more bandwidth, 37.5% more shaders and TMUs deliver more than 30%?

 

[blackened23]

Look at GTX680->Titan.
50% more ROPs, 75% more shaders, 50% more bandwidth and 75% more TMUs didn't deliver 50% more performance (I see 30% in TPU).

Why should 100% more ROPs, 11% more bandwidth, 37.5% more shaders and TMUs deliver more than 30%?

Performance has never increased linearly with specifications, but with that said - this is all part of the "who the EFF cares" territory. Performance and noise output are the two most important metrics I care about, and I suspect that most consumers feel that way.

 

[toyota]

The PHYs on Hawaii are less complex than the ones on Tahiti, and being less complex makes Hawaii's 512-bit memory controller smaller Tahiti's 384 bus.
This allows AMD to go with cheaper lower clocked memory and to get the desired bandwidth by going with a wider bus
There may also be potential power savings by doing it this way. [Beyond3D source of this info]

all I said was that it's nonsense to think they went slower memory for faster timings

 

[RussianSensation]

I think it was designed for 20nm and rushed to 28nm.

How can you rush a new GPU 2 years after Tahiti was released and way behind GTX780? Your theory is odd to say the least considering AMD's rep publicly acknowledged that they stuck to 28nm on purpose to get higher clocks. Rory Read said before that AMD is going to stick to the same nodes longer rather than adopt newer nodes quicker as was done in the past. Your entire statement is in completely contradiction to AMD's current strategy of realizing that moving to 20nm right now is too costly, the yields are too low and the cost/benefit is not yet there.

The reason AMD probably couldn't release Hawaii faster is lack of financial resources compared to NV that had GK110 finalized a lot longer because NV needs GK110 for Tesla/Quadro cards while AMD is specifically designing Hawaii for gaming first, not HPC. It makes total sense that NV is far interested in launching flagship cards like GK110 to service their HPC markets.

64 ROPs and 512 bus for Titan is terrible.

You should apply for a job at AMD/NV if you think 64 ROPs/512-bit bus is "terrible." I guess you know more than all of AMD's engineers combined.

AMD doubled ROPs and went to 512 bus, seems like a trivial gain compared to the amount the hardware increased.

Please post some benchmarks of your R9 290X card that proves the move to 64 ROPs/512-bit bus provides a trivial gain over 7970GE/R9 280X. As has already been stated the move to 512-bit bus reduced the size of the memory controller. 5Ghz GDDR5 likely operates at lower voltages than 6.7Ghz GDDR5. The move to double the ROPs is key too as AMD has been far behind NV's best cards in pixel fill-rate for several generations. This now reverses that role and sets AMD nicely for 20nm having learned how to double the ROPs in such an efficient way from a transistor density point of view.

Seems like a poorly balanced card, or that they've reduced the effectiveness of the ROPs so greatly it's like a pre Kepler core vs a Kepler core.

Saving this for later when a sub-440mm2 R9 290/X is trading blows with a 561mm2 780/Titan and beating 770 4GB/780/Titan on price/performance.

 

[Slomo4shO]

higher than any other video card ever made in history, don't see how it isnt' good enough for u

Considering that there are factory clocked R9 280X running at 307.2 GB/s and many models are capable of overclocking to 336GB/s on a 384 bit bus, achieving 320GB/s bandwidth on a 512 bit bus isn't very impressive. In fact, anything less than 352GB/s on a 512 bit bus is rather pathetic for current generation of vram.

 

[AtenRa]

I'm looking at it from a R280x to R290x standpoint, I don't care what Nvidia has only that for 20-30% more performance AMD doubled ROPs and went to 512 bus, seems like a trivial gain compared to the amount the hardware increased.

Seems like a poorly balanced card, or that they've reduced the effectiveness of the ROPs so greatly it's like a pre Kepler core vs a Kepler core.

I strongly believe there will be more than 20-30% performance increase in high resolutions (above 2560x1600) and high AA filters to Tahiti.

Also as i have said before, although they have installed double the ROPs, i dont believe they have doubled the compute performance. That means, Tahiti to Hawaii ROPs is not 1 to 1 ratio.

Can anyone confirm if they have 16x RBEs (Render Back Ends) ???

 

[toyota]

Russian, to be fair AMD did have 512 bit before and you know it was not a well implemented decision. You'll be surprised how some things are behind closed doors.I know several Automotive Engineers and its painful to listen to the stupidity and ignorance that they have about the very cars they work on and design.they have made some very odd and bad decisions for the companies they work for.

 

[BallaTheFeared]

How can you rush a new GPU 2 years after Tahiti was released and way behind GTX780? Your theory is odd to say the least considering AMD's rep publicly acknowledged that they stuck to 28nm on purpose to get higher clocks. Rory Read said before that AMD is going to stick to the same nodes longer rather than adopt newer nodes quicker as was done in the past. Your entire statement is in completely contradiction to AMD's current strategy of realizing that moving to 20nm right now is too costly, the yields are too low and the cost/benefit is not yet there.

They also said 7xxx series throughout 2013, and that stutter was caused by game engines not their drivers, who really cares what AMD says?

You should apply for a job at AMD/NV if you think 64 ROPs/512-bit bus is "terrible." I guess you know more than all of AMD's engineers combined.

Can I use you as a reference, or are engineers and sales in different parts of the hiring process?

I never said 64/512 was terrible, I said only achieving Titan performance which has already been eclipsed and nullified by non reference 780 was terrible.

Please post some benchmarks of your R9 290X card that proves the move to 64 ROPs/512-bit bus provides a trivial gain over 7970GE/R9 280X.

This is a speculation thread, I'm only making comments based on what has been said/leaked/hinted. If that is wrong I am wrong and I wouldn't care if that was the case, I can accept being wrong it's not that important to me in fact in this case I hope I am.

Saving this for later when a sub-440mm2 R9 290/X is trading blows with a 561mm2 780/Titan and beating 770 4GB/780/Titan on price/performance.

Share your benchmarks and shopping cart, lol after just trying to use that against me... You're comical, next you'll be saying we can't compare non reference 780s to the reference 290X then hours later turn around and compare non reference 280X'es to reference 770s.

 

[blackened23]

Heh. Arguing specifications and scaling is literally the biggest waste of time and the most ridiculous argument ever IMO. Performance and noise matter. This garbage? Not so much. 680 > Titan didn't scale linearly with specs as RS mentioned. Nowhere near linear scaling. Carry on.

 

[RussianSensation]

Considering that there are factory clocked R9 280X running at 307.2 GB/s and many models are capable of overclocking to 336GB/s on a 384 bit bus, achieving 320GB/s bandwidth on a 512 bit bus isn't very impressive. In fact, anything less than 352GB/s on a 512 bit bus is rather pathetic for current generation of vram.

The flaw with that assumption is that a videocard's performance does NOT increase linearly with a commensurate increase in memory bandwidth. As has already been shown with HD4890 (vs. HD7790 or HD7850) and GTX580 (vs. GTX570/GTX660Ti), or HD7950 vs. HD7870, or 7970GE vs. GTX680, the extra memory bandwidth can simply be wasted if the GPU isn't powerful enough to take advantage of it. None of these cards benefited tremendously from their high memory bandwidth vs. their own GPU processing power. They had way too much memory bandwidth for their GPU.

You are then assuming that Tahiti XT/XTL takes advantage of the added memory bandwidth. I own Tahiti and I know it's not true. When I overclock my memory to 7200mhz, the increase in most games is nothing worth talking about. Most of the increase comes from GPU overclocking which suggests Tahiti's performance barely improves by moving from 264GB/sec to 345 GB/sec unless the game is very memory bandwidth limited.

The beauty of the 512-bit bus is that if R9 290X scales much better than Tahiti with memory bandwidth, then these cards can be overclocked and the 512-bit bus will provide a larger % gain per mhz than if that card had a 384-bit bus. One guess is that if R9 290X's VRAM can clock to 6Ghz, AMD limited it for power consumption reasons. Another guess is that it's not about 320 GB/sec being "pathetic" for a 512-bit bus. AMD chose the 512-bit bus because the memory controller is smaller than their previous gen 384-bit one. It's a die size saving strategy without a memory bandwidth penalty.

Obviously over time a card with 512-bit bus and GDDR6 on 14-16nm will have a far higher amount of memory bandwidth. From that point of view, Hawaii's total memory bandwidth over 512-bit bus won't seem impressive. But this has nothing to do with downplaying the reasoning of why AMD chose to utilize the 512-bit bus now.

 

[Face2Face]

I know several Automotive Engineers and its painful to listen to the stupidity and ignorance that they have about the very cars they work on and design.they have made some very odd and bad decisions for the companies they work for.

Let me guess.....they work for Toyota..? 😉

 

[BallaTheFeared]

You're of course assuming they're clocked low for reasons other than like with Tahiti where the bus simply couldn't handle higher clocked ram across all samples.

 

[toyota]

Let me guess.....they work for Toyota..? 😉

lol actually they work for Honda. And none of them even drive Hondas so that should tell you something. And really the name Toyota that I picked means nothing at all as I don't work for them and never even owned a Toyota until last year. It was just a random name when I first started getting on the Internet and I wish I could change it on here.

 

[Stuka87]

The flaw with that assumption is that a videocard's performance does NOT increase linearly with a commensurate increase in memory bandwidth. As has already been shown with HD4890 (vs. HD7790 or HD7850) and GTX580 (vs. GTX570/GTX660Ti), or HD7950 vs. HD7870, or 7970GE vs. GTX680, the extra memory bandwidth can simply be wasted if the GPU isn't powerful enough to take advantage of it. Neither of those cards benefited tremendously from their high memory bandwidth vs. their own GPU processing power.

You are then assuming that Tahiti XT/XTL takes advantage of the added memory bandwidth. I own Tahiti and I know it's not true. When I overclock my memory to 7200mhz, the increase in most games is nothing worth talking about. Most of the increase comes from GPU overclocking which suggests Tahiti's performance barely improves by moving from 264GB/sec to 345 GB/sec.

The beauty of the 512-bit bus is that if R9 290X scales much better than Tahiti with memory bandwidth, then these cards can be overclocked and the 512-bit bus will provide a larger % gain per mhz than if that card had a 384-bit bus.

I think the reason Tahiti does not scale well with memory overclocking is because they are not starved for bandwidth. GTX680's show quite a bit of gain with memory overclocking because they are memory bottlenecked. Tahiti's are not.

I do not see memory bandwidth being an issue with 290X. A 512bit bus will allow them to use lower cost memory without a hit to performance. We will find out just how it performs in 5 days.

 

[GaiaHunter]

I never said 64/512 was terrible, I said only achieving Titan performance which has already been eclipsed and nullified by non reference 780 was terrible.

No, what you said was that 20-30% performance increase over 7970GE/R9 280X is pitiful considering the amount of hardware increase.

I've shown you another case where the hardware increase was even bigger and the performance was around the 30% mark, at least for 1080 and 1600.

So is Titan, GTX780 performance gain also pitiful considering the amount of hardware increase or do you have a different basis to consider that R9 290X needs to be x% (x>30)?

 

[BallaTheFeared]

No because when you compare max clocked 780s/Titans against max clocked 680s the performance delta is closer to 50%.

 

[Slomo4shO]

The flaw with that assumption is that a videocard's performance does NOT increase linearly with a commensurate increase in memory bandwidth.

I never made any arguments regarding correlations between performance and memory bandwidth, that is your own assumption. I was simply stating that memory speeds have increased over time and that 320GB/s bandwidth over a 512 bit bus is not impressive by any means.

 

[Face2Face]

lol actually they work for Honda. And none of them even drive Hondas so that should tell you something. And really the name Toyota that I picked means nothing at all as I don't work for them and never even owned a Toyota until last year. It was just a random name when I first started getting on the Internet and I wish I could change it on here.

Not to get too off topic, but you can change your name. Send a msg to the mods, I changed mine from Greenhell6, which is a name I came up with when I was a teen. I love my Toyota, great truck!

 

[omeds]

I never made any arguments regarding correlations between performance and memory bandwidth, that is your own assumption. I was simply stating that memory speeds have increased over time and that 320GB/s bandwidth over a 512 bit bus is not impressive by any means.

It's not impressive in omg leet uber bandwidth terms, but it may be impressive from a power use and cost perspective, while still offering a little more bandwidth. That is to say I dont think the choice to go 512 bit was because the were chasing uber high amounts of bandwidth. It is a design choice based on other benefits.

 

[GaiaHunter]

You're of course assuming they're clocked low for reasons other than like with Tahiti where the bus simply couldn't handle higher clocked ram across all samples.

It does seem both NVIDIA and AMD are using 6Gbs memory with their 384 bit bus, although AT has articles on both R9 280X and GTX780 getting 7Gbs overclocks for memory.

http://www.anandtech.com/show/7406/the-sapphire-r9-280x-toxic-review/5
http://www.anandtech.com/show/7356/capsule-review-evga-geforce-gtx-780-superclocked-acx/4

 

[GaiaHunter]

I never made any arguments regarding correlations between performance and memory bandwidth, that is your own assumption. I was simply stating that memory speeds have increased over time and that 320GB/s bandwidth over a 512 bit bus is not impressive by any means.

Can you point a 512 bit bus reaching higher?

 

[railven]

No because when you compare max clocked 780s/Titans against max clocked 680s the performance delta is closer to 50%.

Now this strikes me as interesting, because 7970 was considered bad when it was ~30% faster than HD 6970, but max clocks both that gap was now >50%.

So, when we consider a product, what do we use for the basis? Stock, max clocks, - I'm just curious because following these forums this is constantly changing.

I remember people defending heavily OC'ed GTX 460's against HD 6870, but you couldn't use an OC'ed 7970 against a GTX 580.

 

[GaiaHunter]

No because when you compare max clocked 780s/Titans against max clocked 680s the performance delta is closer to 50%.

Then when we see the difference between max OC R9 290X vs max OC R9 280X we can talk.

Unless you are sitting in benchmarks you aren't posting?

 

[GaiaHunter]

Now this strikes me as interesting, because 7970 was considered bad when it was ~30% faster than HD 6970, but max clocks both that gap was now >50%.

So, when we consider a product, what do we use for the basis? Stock, max clocks, - I'm just curious because following these forums this is constantly changing.

I remember people defending heavily OC'ed GTX 460's against HD 6870, but you couldn't use an OC'ed 7970 against a GTX 580.

I can see a pattern in there.