This is most likely the cause; the 7850 1GB used was clocked at 860Mhz core whilst the 7790 was clocked at 1075Mhz core, that is an enormous difference.
7850 @860Mhz = 1.76Tflops
7850 @1075Mhz = 2.2Tflops
AMD: Southern Islands (7*** series) Speculation/ Rumour Thread
- Thread starter UniversalTruth
- Start date
This is most likely the cause; the 7850 1GB used was clocked at 860Mhz core whilst the 7790 was clocked at 1075Mhz core, that is an enormous difference.
7850 @860Mhz = 1.76Tflops
7850 @1075Mhz = 2.2Tflops
depend what you are using as AA and the quality level ... on max AA, the game use SSAA ( 2x at high and 4x SSA if use the extreme setting if im right ). in Addition if they disable all effect and keep environnement in high, instead of extreme, im not even sure the HD textures quality is used.
"Spanking" ? Is that the official performance gap, or the AMD Intern approved measurement?
TPU ~5% @ 1920x1080
Minimal difference recorded by TR, and Anand basically shows what was already known- that games with a DirectCompute component tip the scales in favour of the GCN arch.
The pricing and free game seem to be the biggest differentiators between the two cards...and both of those evaporate against their own product stack when you have 7850's retailing for $160.
The only game on Anandtech review where the 650Ti is upper is BF3, i was think this is strange, so i have watch other review.. in all review the 7790 is faster in Bf3, not by much, but still . ( including TPU and TR review you have post ). ( TR the resuts are the same, because they use the 650TI AMP OC @ 1033mhz (instead of 928mhz ).
Something funny is the 650TI AMP, can be largely used as an relative indication for the future Nvidia 650TI Boost ( where the OC models will pack 993mh base and 1050+mhz Boost ( 1059mhz on the Zotac one rated at 140W )
Last edited by a moderator:
It's not that comfortable a margin. Compared to Orbis it has 97% texturing performance and shader performance. 63% fill rate and 125% geometry setup/tessellation. The only really big disadvantage is memory bandwidth (55% or 70% if you account for separate system memory bandwidth) and of course memory size. Where they are limiting factors the 7970 won't stand a chance. For everything else it should more or less be able to keep up.
Compared to Durango the picture is far more favourable. 146% texturing and shader performance and 125% fill rate and geometry setup /tessellation performance. But still only 56% (72% with system memory bandwidth) of the memory bandwidth so this and memory size will again be it's big downfalls.
I don't see why Durango or Orbis will be seeing particularly higher real world performance than this GPU in a PC. They are exactly the same architecture so any optimisations in the consoles will translate directly to this GPU and API overhead is much lower in DX11 (which will be the standard) than it was in DX9 which is what the "old" overhead arguments are based on.
The 650Ti Boost should still have quite an advantage due to the 192bit/24 ROP configuration - since the memory clock increased vs. the normal 650Ti that's still a 50% bandwidth advantage over 650Ti AMP (and over 7790). So the 650Ti Boost should really be faster and closer to a 660GTX. In perf/power though it will look rather sad compared to 7790...
While DX11 is an improvement my understanding is that there's still a lot of API crud in the windows world, not to mention PCIe latencies etc.
Sure, but most of the API overhead is going to effect the CPU rather than the GPU. And PCI-E latency shouldn't have any baring on graphics performance. That only comes into play if you need low latency communication between the CPU and GPU, for example if you're performing gameplay effecting GPGPU calculations on the GPU. And if you're doing that on the console then that's actually sapping more GPU performance where the same work on the PC may have to be performed on the bigger CPU since PCI-E latency is too high to allow GPGPU to offer any benefit.
Does somebody know why AMD avoids 192 bit bus?
I mean the card comes really close the hd7850, with more bandwidth it would be even better while still cheaper and more competitive (making it tougher for NV to compete).
I see something a bit weird here as far as product placement is concerned, I would have think that AMD could have "bartized" their previous line, so in the same time replacing anything in between the HD7700 and the HD7850 by that new GPU and soon after putting a product in between Pitcairn and tahiti (so discontinuing pitcairn all together, replace on the lower hand by a cheaper cheaper to produce and on the higher something significantly cheaper than Tahiti), wider bus width would have come handy.
Though not that I imply that there is anything with the product it seems to perform greatly, but I would think that AMD has a good shot a repositioning their offer (and releasing 87xx and 88xx as "Bart type of product").
I mean the card comes really close the hd7850, with more bandwidth it would be even better while still cheaper and more competitive (making it tougher for NV to compete).
I see something a bit weird here as far as product placement is concerned, I would have think that AMD could have "bartized" their previous line, so in the same time replacing anything in between the HD7700 and the HD7850 by that new GPU and soon after putting a product in between Pitcairn and tahiti (so discontinuing pitcairn all together, replace on the lower hand by a cheaper cheaper to produce and on the higher something significantly cheaper than Tahiti), wider bus width would have come handy.
Though not that I imply that there is anything with the product it seems to perform greatly, but I would think that AMD has a good shot a repositioning their offer (and releasing 87xx and 88xx as "Bart type of product").
Last edited by a moderator:
Maybe because there's no natural 192 bits memory controller, same with 320bits memory controller ? nvidia disable just a part of a larger memory controller, it is not a true 192 bits controller . They just disable one . If you are using a 256bits memory controller on a sku, well you can disable one part and it become a 192bits.. Nvidia just use sku who was not dedicated for be on this range, and diminish the performance of higher sku for make them enter in this one.. But without compet with the cards using the same sku.
Now by using 2gb instead of 1gb.. they could get some advance on some games, as bandwith become more and more critical today on this mid range gpu ..
AMD could have use Pitcain and just put some variation for release the 7790 sku .. but it seems they had other plans.. Nvidia is going " on the fight" by using their upper sku , and disable some parts .. ok. its " a good war" .. ( in french " ' c' est de bonne guerre " )
Last edited by a moderator:
There are native 384-bit controllers. I'm sure 192-bit ones would be no more challenging to make. But AMD probably figured it wasn't worth the extra cost, and considering how Bonaire performs, I think they were right.
Blazkowicz
Legend
Nvidia has done native 192bit memory controllers, the GF106, GF116 and GK106 physically have 192bit bus - though physically, what you really have is a collection of 64bit controllers with nvidia GPU having one - on GF119, two, three, four or six of them.
They also developed that weird use of mixed memory densities that allows you to make 1GB and 2GB cards, rather than 1.5GB or 3GB. AMD didn't do that. (GF106 was put on almost no 192bit cards, there was an OEM one with one SM disabled and 1.5GB ddr3 and that's it)
They also developed that weird use of mixed memory densities that allows you to make 1GB and 2GB cards, rather than 1.5GB or 3GB. AMD didn't do that. (GF106 was put on almost no 192bit cards, there was an OEM one with one SM disabled and 1.5GB ddr3 and that's it)
I am quite sure both gk106 and gf106 have native 192bit memory controllers. Well the channels are just 64bit so there's really no problem there. The resulting asymmetric memory configuration (if you want to go for 1GB/2GB sizes) though _may_ pose some problems.
Also I don't know if AMD could partition into non-power-of-two sizes easily (that goes for both ROPs and MCs). All their designs disable ROPs so the ROP number stays a power of two and I can't remember what the last card was which disabled memory controllers at all.
That said yes a theoretical 192bit chip would at least equal HD7850 in performance (with 1.5Ghz gddr5). But the chip would probably get a bit close in size to Pitcairn (and of course still be slower than the full version of it). Or if you'd wanted it to keep the same size, you'd have to probably sacrifice two CUs along with the second setup unit. And it's not obvious it would be faster that way. But yes this chip is (compared to the rest of the series) high on shader/tmu and especially prim setup rate, and low on ROPs and memory bandwidth.
Also I don't know if AMD could partition into non-power-of-two sizes easily (that goes for both ROPs and MCs). All their designs disable ROPs so the ROP number stays a power of two and I can't remember what the last card was which disabled memory controllers at all.
That said yes a theoretical 192bit chip would at least equal HD7850 in performance (with 1.5Ghz gddr5). But the chip would probably get a bit close in size to Pitcairn (and of course still be slower than the full version of it). Or if you'd wanted it to keep the same size, you'd have to probably sacrifice two CUs along with the second setup unit. And it's not obvious it would be faster that way. But yes this chip is (compared to the rest of the series) high on shader/tmu and especially prim setup rate, and low on ROPs and memory bandwidth.
UniversalTruth
Veteran
Hmm, 7790 has more transistors but in the end shows either lower or on par performance with the 6870.
_vs_
http://www.techpowerup.com/reviews/Powercolor/HD_7790_Turbo_Duo/26.html
I don't know exactly but can't they achieve whatever memory bandwidth they want through using memory modules with different frequency?
http://www.techpowerup.com/reviews/Powercolor/HD_7790_Turbo_Duo/26.html
I don't know exactly but can't they achieve whatever memory bandwidth they want through using memory modules with different frequency?
Hmm, 7790 has more transistors but in the end shows either lower or on par performance with the 6870.
Well it's a lot more power efficient and smaller too remember.
It's a typical mid-node AMD card, compare it to the 7770 which is the card it was probably designed to replace before AMD's roadmap and strategy got thrown out the window. It was never going to be a balanced part with only a 128-bit bus, but it gets the most performance likely out of such a small chip I'd wager.
Better still compare it to Juniper and you'll see how much they've gained for what are very similar parts. Some 50%-100% faster at around the same power and die size isn't bad given how spectacular Juniper's perf/watt was already.
http://www.anandtech.com/bench/Product/540?vs=538 (assuming equal performance to the 6870).
They don't want it out performing the 7850 at $150. There is no mystery. They slotted in a part to fill a gap, they don't want it cannibalizing sales of their $200 part.
Note the 7790 also handles the 6870 in some newer titles like sleeping dogs and bf3. In a couple years I wouldn't be surprised if the 7790 looks much stronger.
Note the 7790 also handles the 6870 in some newer titles like sleeping dogs and bf3. In a couple years I wouldn't be surprised if the 7790 looks much stronger.
That's right. GCN is way better at compute, and this particular chip also happens to have a stronger front-end with 2 primitives/clock. So Bonaire may lose to Barts in a lot of old and current games, but I suspect it will be faster in future ones.
Some of the additional transistors may be there for power-efficiency more than for raw performance, but of course power-efficiency ultimately translates into performance. Plus, Barts has a 256-bit interface, which doesn't cost a lot of transistors, but does cost a lot of silicon area, not to mention the number of required DRAM chips and the added PCB complexity. So just looking at the transistor count doesn't tell the whole story.
Have you ever seen Nvidia do this type of controller ? his question was why AMD dont use this type of controller ? This will be completely stupid to design a memory controller of this type anyway with a GDD bankson the end .
I think have respond to him... thoses controllers have never been developped, ( by AMD/ATI or Nvidia ) because the results are bad.. Nvidia dont developp them, they use a 256bit memory controller, and erase 64bits of it for reduce the perfomances of the inital 256 bit memory controller who was used on the initial sku .. it is not a pure 192bits controller or a 320bts.. it is a bottlenecked one for limit the performance and sell this card on the middle range ..
Last edited by a moderator:
I think NVIDIA's GK106 (found in the GTX 660 Ti) has a native 192-bit interface, and it performs quite well. It makes it a bit awkward to manage memory capacities if you want powers of 2, but otherwise it's fine.
You will need to recover your source ( or maybe if it is right i should do it myelf, so i let the door open if i was wrong ), but it is not a native 192bits interface .. The 660TI use the same sku of the 670-680 ( GK104 ) and the same chip and memory controller with some part disabed.
The GTX660TI use the GK104 as base .. but not the GTX 660 ( who use the GK106 )
Last edited by a moderator:
You're right, but the GTX 660 (non-Ti) does feature GK106, and GK106 does have a native 192-bit controller: http://www.hardware.fr/articles/876-2/gk106-5-smx-192-bit.html
Similar threads
