Tech ARP updated GTX 380, 360, 395 SPECs!

[elpedro]

About 10% faster actually: http://www.madshrimps.be/vbulletin/f22/amd-radeon-hd-5870-hd-5850-performance-chart-66465/

Why is the 8800GT 7% quicker than a 9800GT according to that graph

 

[drunkenmaster]

9600GT?

9600gt 550million transistor part, 9800gtx 750million transistor part, so other than it NOT being salvaged but being a smaller core with the correct amount of shaders... oh wait, so completely and utterly different.

If they'd made 9600gt's out of 9800gtx cores it would have been just as bad, but the point is they didn't. With the 9600gt they cut so much off it, yet it still proved to be very very good performance comparitively(because the 9800gtx was heavily limited somewhere, I can't remember the reasoning).

Fact is the 9600gt, supposely midrange, was a little bit too close to the top end performance from what I recall, but it was a midrange gpu design, well really just a cut down 8800gtx, but thats fine, it was smaller and therefore had less shaders. The 350gtx could be right up on the 360gtx performance for all I care, I don't mind the name, call it a 220gt, or a 495gtx ultra mega uber boobs, the problem is when you make a 256shader part out of a core twice its size, why, because you can't sell it for the same price as a full 512shader part. Why's that bad, the cheaper that 256shader card has to be sold for, the more the higher end cores have to be sold for to even it out.

If each core working perfectly cost $100 out of the oven, if you can only sell half at $50 because they only half work, to BREAK EVEN you HAVE TO increase the cost of the other cores that do work. The fact that so many expensive cores look set to be sold much cheaper, just pushes the 380gtx price higher in the end, which is bad for everyone, AMD fans, Nvidia fans, gaming fans. If they were making 256shader cores out of a 1.8-2billion transistor core that only had 256shaders, that would be fantastic, there might be a price war created from a well priced midrange card. But a core with only 256 shaders, being sold with all its shaders doesn't effect the price of waifers with larger cores, this one would, dramatically.

if these specs turn out to be what they are or very similar, which gpu will you buy ATI or Nvidia?

Already have a 5850, if Nvidia release something with the specs as listed, it will still be better value to go AMD. A second 5850, even at £200 would make a cheaper and faster setup than a 380gtx, which I can't see them launching at under £450, maybe even launch it at £550+ just so no one actually wants one. It's a great way to do business, release a card you can barely make but price it where no one wants it, and no one really notices you can't make them because theres still a few around.

If the 360gtx turns out to be great at circa £300, then the 5850 will drop to £150, undercut it and still 2 in crossfire will be faster than it by a HUGE margine. If a 350gtx is in the works to salvage yields, well it won't be very fast anyway, and it will likely be beaten on price by 5770 in crossfire aswell as smashed in performance.

I have little doubt that Nvidia can make a faster single gpu card than AMD have, AMD's design policy is to not be going for the performance crown, Nvidia's is. The problem is twofold, do you need that much extra juice, can you get the same performance with two cheaper cards, ok threefold, can they even sell it. The other big issue being, the only games that really want a 5970's level of power, for now, would be a game in ridiculous res, which eyefinity gives you. Then again I guess you can say Nvidia has the 3d screens so getting 120fps worth of performance is fairly crucial there also.

I do genetic algorithm research on CUDA, and I'm waiting anxiously for CUDA 3.0 on Fermi. Unlike most, I'm not too bothered (yet) about their massive DP performance increase, rather the simultaneous different kernel executions and increased caching/memory behaviour.

Both features the 5870 has, can execute different programs on different clusters, and increased caching over last gen. This is the problem, AMD really don't want to invest millions marketing in a market thats worth less than $50million turnover to them, however they basically have every last GPGPU feature Nvidia are going to have with Fermi, while pushing arguably the better platform(opencl). The thing is Nvidia have no where else to expand so are spending heavily to get into a market that was only worth $88million in revenue, not profit, last year.

Pretty much everything Fermi can do, the 5870 can already do. I'm not a good enough coder to know this though, for gaming the 4+1 architecture AMD have is VERY hard to get the full power out of, thats why its peak theoretical throughput dwarfs Nvidia's, but most of the time its using about half its capacity. But in non gaming coding, GPGPU work, I'm not sure if its easier or not to use the full 4+1 parts of a shader. If you can, the 5870 WILL spank Nvidia, if a specific gpgpu program can't, then it won't.

The funny part is, well not funny, almost sad, Nvidia will spend millions, maybe hundreds of millions expanding a tiny market over the next few years, while Intel and AMD simply intergrate massive fpu monsters onto the cpu die. In the not to distant future we'll be at the point where there will be a interger line of I 17 cpu's with 24 int cores and 8 fpu cores, and a fpu moster version with 4 in cores and 28 fpu cores. Aswell as both having the abilility to rather seemlessly(software wise) add extra discrete gpgpu's to share the workload and offer a gpgpu platform top to bottom to the big companies. That means reliability, compatibility complete flexibility and better pricing, aswell as both companies producing all their own stuff cutting out the middle man in manufacturing to decrease prices. So Nvidia are set to spend money turning this tiny market into a bigger market, for both Intel and AMD to step in and take it easily. AMD are keeping up with features, improving gen to gen and when the markets big enough that extra marketing and spending means profits(which they don't now) they'll go for it.

 

[Biffa]

I think that the 5870 doesn't have ECC memory, which is a biggie for the scientific bods apparently

 

[Greebo]

Not sure if 1600MHz shaders is right - tho that was originally the target... but there seemed to be a change recently.

EDIT: If thats right and they really have bumped the 360 to 480SP its going to be a good margin faster than the 5870...

At 384SP the 360GTX would have been competitive against the 5870 either very slightly slower or very slightly faster... but with 480SP it would be quite a bit faster.

Am I reading that link wrong then? Assuming the figures are correct the pixel fillrate, texture fillrate and memory bandwidth of the GTX360 are all less than a GTX295 so how does the GTX360 magically outperform it?

With the texture fillrate being 26% slower than GTX295 albeit the other figures almost matching it, I would say that chart shows that the GTX360 will be slower than GTX295 and hence as I have said with a 5870 being about 10% slower than a GTX295 that the GTX360 will match the 5870 on performance.

I know shaders are important but doesn't a graphics cards performance relates to the actual pixel and texture fillrates they acheive rather than number of cores/shaders etc? (driver optimisation excluded)

Even the GTX380 can't match the texture fillrate of the GTX295 although it blows it away with pixel fillrate and memory bandwidth.

To me, assuming those figures are correct which is a big IF, the GTX360 will trade blows with the 5870 just like the gtx260 did with the 4870/4890 and the GTX380 will be the fastest single core graphics card you can buy but will still not beat a 5970 so as I have stated before, price will be the key.

 

[thevet]

9600gt 550million transistor part, 9800gtx 750million transistor part, so other than it NOT being salvaged but being a smaller core with the correct amount of shaders... oh wait, so completely and utterly different.

If they'd made 9600gt's out of 9800gtx cores it would have been just as bad, but the point is they didn't. With the 9600gt they cut so much off it, yet it still proved to be very very good performance comparitively(because the 9800gtx was heavily limited somewhere, I can't remember the reasoning).

Fact is the 9600gt, supposely midrange, was a little bit too close to the top end performance from what I recall, but it was a midrange gpu design, well really just a cut down 8800gtx, but thats fine, it was smaller and therefore had less shaders. The 350gtx could be right up on the 360gtx performance for all I care, I don't mind the name, call it a 220gt, or a 495gtx ultra mega uber boobs, the problem is when you make a 256shader part out of a core twice its size, why, because you can't sell it for the same price as a full 512shader part. Why's that bad, the cheaper that 256shader card has to be sold for, the more the higher end cores have to be sold for to even it out.

If each core working perfectly cost $100 out of the oven, if you can only sell half at $50 because they only half work, to BREAK EVEN you HAVE TO increase the cost of the other cores that do work. The fact that so many expensive cores look set to be sold much cheaper, just pushes the 380gtx price higher in the end, which is bad for everyone, AMD fans, Nvidia fans, gaming fans. If they were making 256shader cores out of a 1.8-2billion transistor core that only had 256shaders, that would be fantastic, there might be a price war created from a well priced midrange card. But a core with only 256 shaders, being sold with all its shaders doesn't effect the price of waifers with larger cores, this one would, dramatically.



Already have a 5850, if Nvidia release something with the specs as listed, it will still be better value to go AMD. A second 5850, even at £200 would make a cheaper and faster setup than a 380gtx, which I can't see them launching at under £450, maybe even launch it at £550+ just so no one actually wants one. It's a great way to do business, release a card you can barely make but price it where no one wants it, and no one really notices you can't make them because theres still a few around.

If the 360gtx turns out to be great at circa £300, then the 5850 will drop to £150, undercut it and still 2 in crossfire will be faster than it by a HUGE margine. If a 350gtx is in the works to salvage yields, well it won't be very fast anyway, and it will likely be beaten on price by 5770 in crossfire aswell as smashed in performance.

I have little doubt that Nvidia can make a faster single gpu card than AMD have, AMD's design policy is to not be going for the performance crown, Nvidia's is. The problem is twofold, do you need that much extra juice, can you get the same performance with two cheaper cards, ok threefold, can they even sell it. The other big issue being, the only games that really want a 5970's level of power, for now, would be a game in ridiculous res, which eyefinity gives you. Then again I guess you can say Nvidia has the 3d screens so getting 120fps worth of performance is fairly crucial there also.



Both features the 5870 has, can execute different programs on different clusters, and increased caching over last gen. This is the problem, AMD really don't want to invest millions marketing in a market thats worth less than $50million turnover to them, however they basically have every last GPGPU feature Nvidia are going to have with Fermi, while pushing arguably the better platform(opencl). The thing is Nvidia have no where else to expand so are spending heavily to get into a market that was only worth $88million in revenue, not profit, last year.

Pretty much everything Fermi can do, the 5870 can already do. I'm not a good enough coder to know this though, for gaming the 4+1 architecture AMD have is VERY hard to get the full power out of, thats why its peak theoretical throughput dwarfs Nvidia's, but most of the time its using about half its capacity. But in non gaming coding, GPGPU work, I'm not sure if its easier or not to use the full 4+1 parts of a shader. If you can, the 5870 WILL spank Nvidia, if a specific gpgpu program can't, then it won't.

The funny part is, well not funny, almost sad, Nvidia will spend millions, maybe hundreds of millions expanding a tiny market over the next few years, while Intel and AMD simply intergrate massive fpu monsters onto the cpu die. In the not to distant future we'll be at the point where there will be a interger line of I 17 cpu's with 24 int cores and 8 fpu cores, and a fpu moster version with 4 in cores and 28 fpu cores. Aswell as both having the abilility to rather seemlessly(software wise) add extra discrete gpgpu's to share the workload and offer a gpgpu platform top to bottom to the big companies. That means reliability, compatibility complete flexibility and better pricing, aswell as both companies producing all their own stuff cutting out the middle man in manufacturing to decrease prices. So Nvidia are set to spend money turning this tiny market into a bigger market, for both Intel and AMD to step in and take it easily. AMD are keeping up with features, improving gen to gen and when the markets big enough that extra marketing and spending means profits(which they don't now) they'll go for it.

i see your points...interesting....like greebo said then the price is the key

 

[Vince]

I think that the 5870 doesn't have ECC memory, which is a biggie for the scientific bods apparently

"Evergreen family of GPU chips comes with a fully-optimized memory controller for the GDDR5, and built-in ECC function into every GDDR5 chip means you can "drive the cells" to the yet unseen points in the world of DDR SDRAM."

Source : Bright side of news

I am not sure if this is equivilant tech to the nvidia card but was sure I remembered error correction on the memory from release.

 

[thevet]

released yesterday at Tech ARP
http://www.techpowerup.com/110471/N...X_300_Series_to_Clinch_Performance_Crown.html

 

[gurusan]

dude those have been floating around for quite awhile now. For some reason they keep popping up on sites every day as "new and just released" or something.

 

[mk17]

...
Pretty much everything Fermi can do, the 5870 can already do. I'm not a good enough coder to know this though, for gaming the 4+1 architecture AMD have is VERY hard to get the full power out of, thats why its peak theoretical throughput dwarfs Nvidia's, but most of the time its using about half its capacity. But in non gaming coding, GPGPU work, I'm not sure if its easier or not to use the full 4+1 parts of a shader. If you can, the 5870 WILL spank Nvidia, if a specific gpgpu program can't, then it won't.
...

2 very good points - have a google on ATI uptake for research, there's a very good reason why NV have the edge significantly.

However I agree with you 100%, I wish ATI could get their coding model into a state of maturity. We're self-financed at the moment, so ATI (and just having them in the market) would make a huge difference.

OpenCL is a dream yet-to-be atm. It can be done, but hugely unwieldy at present.

All that said, it is expected that GPGPU will go the way of maths-coprocessors of old, and soon become standard fare.

Sooner the better, imo Still changes very little on the gfx card wars though!

 

[Andric]

released yesterday at Tech ARP
http://www.techpowerup.com/110471/N...X_300_Series_to_Clinch_Performance_Crown.html

old news - and even fudzilla has agreed these are fake & fudo <3's nvidia.
http://www.fudzilla.com/content/view/16843/1/

 

[thevet]

old news - and even fudzilla has agreed these are fake & fudo <3's nvidia.
http://www.fudzilla.com/content/view/16843/1/

oh ok i did not know

 

[Valleyman]

GTX 395: Max Board TDP = 380W
HAHAHAHA, really? Really? REEAAALLLY??

Might wanna look at the PCI Express regulations there... Max power consuption allowable on a 12V current draw is 300W. Fair enough you can specifiy a TDP of greater than max power consumption if you're making sure you can shift over and above the thermal energy produced, but it HAS to be close, which this isn't. It's no-where near a valid safety factor. Fail fail fail.

People faking specs and pretending to be an electronics engineer should at least try and grasp basic knowledge of electrical engineering first.

 

[Biffa]

"Evergreen family of GPU chips comes with a fully-optimized memory controller for the GDDR5, and built-in ECC function into every GDDR5 chip means you can "drive the cells" to the yet unseen points in the world of DDR SDRAM."

Source : Bright side of news

If that report is accurate then its wierd that more ppl don't realise this.

Lots of conflicting reports on the net about this, although HardOCP does say:

http://www.hardocp.com/article/2009/09/22/amds_ati_radeon_hd_5870_video_card_review/5

Memory Controller

The Radeon HD 5800 series memory controller has also received some optimizations. ATI has now turned on GDDR5’s Error Detection Code (EDC) similar to ECC in DRAM. This allows GDDR5 modules to reach higher frequencies while reducing errors. This could help make video cards as a whole more stable and robust, especially as you clock memory frequencies higher.

One interesting result of enabling EDC however is that overclocking consequences have now changed. In the past, when you overclocked memory on a video card it would ramp up in frequency until the video card would start throwing out artifacts, and then crash. Now when you overclock memory frequency your performance will increase until it reaches its climax and EDC kicks in fixing errors in the memory due to the frequency being too high. This means EDC will fix these problems and your performance will start to actually degrade as you increase frequency and EDC has to work harder at fixing the errors. There won’t be as much flashing polygons and artifacts anymore and no outright crashing from overclocking memory, you will only see a performance decline. Therefore, the goal is to now overclock memory until performance starts declining, then back it off until you’ve reached the point right before the decline, and that is your maximum stable memory overclock.

 

[Mr Paul]

It's accurate.

 

[evoyear]

wait and watching, because the budget, so only looking for the price drop of older generation cards

 

[Greebo]

If that report is accurate then its wierd that more ppl don't realise this.

I always knew this from when they first came out and widely reported in reviews.

In fact I have tried it on my card. Once I go past 1300Mhz on memory my vantage scores keep decreasing. I can put the memory as high as I like without artifacts or crashing but framerates drop.

It makes it a much more long winded process to find your maximum memory overclock as you have to run benchmarks after every setting to check.

 

[Lightnix]

These specs fly in the face of what has already been confirmed by nvidia. Fermi will have a 384bit memory bus, not 512bit. I wouldn't take anything from this list at all.

Exactly this. Feel free to quote me on this when Fermi is released if I'm wrong, but frankly the entries for the Fermi parts on this table contain very little but wishful thinking in terms of clock speeds and memory bandwidth. Also what the hell at the 3-billion transistor Fermi part with only 256 shaders enabled. Least economically viable card ever.

 

[Gills]

Indeed. Thanks to these threads my ignore list is looking pretty healthy

I should have known better than to expect reasoned discussion about up and coming tech.

Seriously, please tell me how to ignore, I got tonnes of people I want to add.

 

[Kribby2006]

9600gt 550million transistor part, 9800gtx 750million transistor part, so other than it NOT being salvaged but being a smaller core with the correct amount of shaders... oh wait, so completely and utterly different.

If they'd made 9600gt's out of 9800gtx cores it would have been just as bad, but the point is they didn't. With the 9600gt they cut so much off it, yet it still proved to be very very good performance comparitively(because the 9800gtx was heavily limited somewhere, I can't remember the reasoning).

Fact is the 9600gt, supposely midrange, was a little bit too close to the top end performance from what I recall, but it was a midrange gpu design, well really just a cut down 8800gtx, but thats fine, it was smaller and therefore had less shaders. The 350gtx could be right up on the 360gtx performance for all I care, I don't mind the name, call it a 220gt, or a 495gtx ultra mega uber boobs, the problem is when you make a 256shader part out of a core twice its size, why, because you can't sell it for the same price as a full 512shader part. Why's that bad, the cheaper that 256shader card has to be sold for, the more the higher end cores have to be sold for to even it out.

If each core working perfectly cost $100 out of the oven, if you can only sell half at $50 because they only half work, to BREAK EVEN you HAVE TO increase the cost of the other cores that do work. The fact that so many expensive cores look set to be sold much cheaper, just pushes the 380gtx price higher in the end, which is bad for everyone, AMD fans, Nvidia fans, gaming fans. If they were making 256shader cores out of a 1.8-2billion transistor core that only had 256shaders, that would be fantastic, there might be a price war created from a well priced midrange card. But a core with only 256 shaders, being sold with all its shaders doesn't effect the price of waifers with larger cores, this one would, dramatically.



Already have a 5850, if Nvidia release something with the specs as listed, it will still be better value to go AMD. A second 5850, even at £200 would make a cheaper and faster setup than a 380gtx, which I can't see them launching at under £450, maybe even launch it at £550+ just so no one actually wants one. It's a great way to do business, release a card you can barely make but price it where no one wants it, and no one really notices you can't make them because theres still a few around.

If the 360gtx turns out to be great at circa £300, then the 5850 will drop to £150, undercut it and still 2 in crossfire will be faster than it by a HUGE margine. If a 350gtx is in the works to salvage yields, well it won't be very fast anyway, and it will likely be beaten on price by 5770 in crossfire aswell as smashed in performance.

I have little doubt that Nvidia can make a faster single gpu card than AMD have, AMD's design policy is to not be going for the performance crown, Nvidia's is. The problem is twofold, do you need that much extra juice, can you get the same performance with two cheaper cards, ok threefold, can they even sell it. The other big issue being, the only games that really want a 5970's level of power, for now, would be a game in ridiculous res, which eyefinity gives you. Then again I guess you can say Nvidia has the 3d screens so getting 120fps worth of performance is fairly crucial there also.



Both features the 5870 has, can execute different programs on different clusters, and increased caching over last gen. This is the problem, AMD really don't want to invest millions marketing in a market thats worth less than $50million turnover to them, however they basically have every last GPGPU feature Nvidia are going to have with Fermi, while pushing arguably the better platform(opencl). The thing is Nvidia have no where else to expand so are spending heavily to get into a market that was only worth $88million in revenue, not profit, last year.

Pretty much everything Fermi can do, the 5870 can already do. I'm not a good enough coder to know this though, for gaming the 4+1 architecture AMD have is VERY hard to get the full power out of, thats why its peak theoretical throughput dwarfs Nvidia's, but most of the time its using about half its capacity. But in non gaming coding, GPGPU work, I'm not sure if its easier or not to use the full 4+1 parts of a shader. If you can, the 5870 WILL spank Nvidia, if a specific gpgpu program can't, then it won't.

The funny part is, well not funny, almost sad, Nvidia will spend millions, maybe hundreds of millions expanding a tiny market over the next few years, while Intel and AMD simply intergrate massive fpu monsters onto the cpu die. In the not to distant future we'll be at the point where there will be a interger line of I 17 cpu's with 24 int cores and 8 fpu cores, and a fpu moster version with 4 in cores and 28 fpu cores. Aswell as both having the abilility to rather seemlessly(software wise) add extra discrete gpgpu's to share the workload and offer a gpgpu platform top to bottom to the big companies. That means reliability, compatibility complete flexibility and better pricing, aswell as both companies producing all their own stuff cutting out the middle man in manufacturing to decrease prices. So Nvidia are set to spend money turning this tiny market into a bigger market, for both Intel and AMD to step in and take it easily. AMD are keeping up with features, improving gen to gen and when the markets big enough that extra marketing and spending means profits(which they don't now) they'll go for it.

Wow man I gotta hand it to you, you certainly go to town with your threads, did you write the bible btw ?

Dont get me wrong its not a criticism but a compliment, fair play .

 

[straxusii]

according to that the 395 breaks ATX limits by being over 300W TDP, strange

I think if it is a PCIe2 card only then it's OK? 150W from the pcie slot compared to 75w