OK, so the PS4 is stronger than the XBOX ONE...

[QUOTE="AMD655"][QUOTE="Tessellation"]damn Ron stop with the ownage,no wonder why you get so easily under the basement dwellers skin..you prove them wrong with knowledge :cool:tormentos

Careful or Tessellation may claim that he got under your skin..:lol:

[QUOTE="ronvalencia"]

Actually, the 7770 has less usable power than the Xbox One since lesser GCNs with less internal SRAM storage has greater chances for compute overspill. 7770 also has less triangle rate than 7790 or 7850.

My 8870M takes a greater performance hit when I lower the memory speed to DDR3 levels i.e. AMD's GDDR5 bar graphs are applicable for these cases.

-----

7790 doesn't sport X1's memory bandwidth and the prototype 7850 (with 12 CUs) is a closer fit to the X1.

X1's GCN has Pitcairn's crossbar, 256bit memory controllers and related L2 cache. L2 cache bandwidth is important since this is used as a shared cache across 4 CUs.

7790/7790 has less I/O L2 cache bandwidth than the Pitcairn's version.

http://www.behardware.com/articles/848-4/amd-radeon-hd-7970-crossfirex-review-28nm-and-gcn.html

GCN Cache Hierarchy

With 128bit memory controllers, 7790/7770 has two L2 blocks. That's 64 bytes per cycle per L2 block x 2 = 128 bytes per cycle.

With 256bit memory controllers, 7850/7870 has four L2 blocks. That's 64 bytes per cycle per L2 block x 4 = 256 bytes per cycle.

tormentos

The broken song again,stop the damn 660Ti has 100Gb/s of bandwidth less than the 7950 and still out perform it in some test...:lol:

You have no argument,i demostrated how a GPU with 100Gb/s less bandwidth than another can still beat it,so the xbox one miserable 37gb/s over the 7790 should be even less of a problem.

The 7790 is very close to the 7850 even that it has 57Gb/s less in bandwidth,so yeah bandwidth is not the issues is power,so yeah the 7790 >> xbox one..

No matter how you slice it what you try to imply is idiotic,if you put 250Gb/s bandwidth on the xbox one would it beat the 7950.? Just because it has more bandwidth.?

7770/7790 @ 800Mhz, L2 cache I/O bandwidth: 128 bytes x 800Mhz = 95.367 GB/s External memory setup faster than 95.367 GB/s would be useless.

W5000/7850-768 @ 853hz, L2 cache I/O bandwidth: 256 bytes x 853Mhz = 203.37 GB/s. ESRAM was design for Pitcairn level I/O.

 

7770/7790 @ 800Mhz, L2 cache I/O bandwidth: 128 bytes x 800Mhz = 95.367 GB/s External memory setup faster than 95.367 GB/s would be useless.

W5000/7850-768 @ 853hz, L2 cache I/O bandwidth: 256 bytes x 853Mhz = 203.37 GB/s. ESRAM was design for Pitcairn level I/O.

ronvalencia

660Ti =144Gb/s

7950 = 240Gb/s..

Bandwidth mean nothing when you have no power,and the xbox one doesn't have it..

Trying to imply that the xbox one would need 203Gb/s is down right stupid it will not.

[QUOTE="ronvalencia"]

 

7770/7790 @ 800Mhz, L2 cache I/O bandwidth: 128 bytes x 800Mhz = 95.367 GB/s External memory setup faster than 95.367 GB/s would be useless.

W5000/7850-768 @ 853hz, L2 cache I/O bandwidth: 256 bytes x 853Mhz = 203.37 GB/s. ESRAM was design for Pitcairn level I/O.

tormentos

660Ti =144Gb/s

7950 = 240Gb/s..

Bandwidth mean nothing when you have no power,and the xbox one doesn't have it..

Trying to imply that the xbox one would need 203Gb/s is down right stupid it will not.

[QUOTE="ronvalencia"]

 

7770/7790 @ 800Mhz, L2 cache I/O bandwidth: 128 bytes x 800Mhz = 95.367 GB/s External memory setup faster than 95.367 GB/s would be useless.

W5000/7850-768 @ 853hz, L2 cache I/O bandwidth: 256 bytes x 853Mhz = 203.37 GB/s. ESRAM was design for Pitcairn level I/O.

tormentos

660Ti =144Gb/s

7950 = 240Gb/s..

Bandwidth mean nothing when you have no power,and the xbox one doesn't have it..

Trying to imply that the xbox one would need 203Gb/s is down right stupid it will not.

[QUOTE="ronvalencia"]

7770/7790 @ 800Mhz, L2 cache I/O bandwidth: 128 bytes x 800Mhz = 95.367 GB/s External memory setup faster than 95.367 GB/s would be useless.

W5000/7850-768 @ 853hz, L2 cache I/O bandwidth: 256 bytes x 853Mhz = 203.37 GB/s. ESRAM was design for Pitcairn level I/O.

tormentos

660Ti =144Gb/s

7950 = 240Gb/s..

Bandwidth mean nothing when you have no power,and the xbox one doesn't have it..

Trying to imply that the xbox one would need 203Gb/s is down right stupid it will not.

I didn't imply anything. My post includes information on CU's max compute processing memory bandwidth from L1 cache and L2 cache I/O.

For X1's GCN, L1 bandwidth has 64 byte per cycle per CU = 50.8425 GB/s x 12 CUs = 610.11 GB/s.

One should look for any bottlenecks and seek the reasons on why it doesn't use extra memory bandwidth.

I have shown you why 7770/7790 @ 800 Mhz would gimp any memory faster than ~96 GB/s. AMD would need to redesign 7770/7790's crossbar/memory controller/L2 cache stack for faster memory based on GDDR6. For the X1, it's easier to "copy-and-paste" from the Pitcairn's crossbar/L2 cache/memory controller designs.

7770/7790 has a massive crossbar/L2 cache I/O bottleneck that gimps higher memory bandwidth e.g. not ready for GDDR6.

Reference http://www.behardware.com/articles/848-4/amd-radeon-hd-7970-crossfirex-review-28nm-and-gcn.html

------------------------

NVIDIA Kelper is different fish when it comes to handling it's compute workloads.

Entry level GK104 (e.g 660 TI) is another class of GPU above AMD Pitcairn i.e. it needs AMD Tahiti LE to battle NV 660 Ti.

K20 vs 7970 vs GTX680 vs M2050 vs GTX580 from http://wili.cc/blog/gpgpu-faceoff.html

K20 = GK110

There's very difference between NVIDIA's OpenCL and CUDA.

App 1. Digital Hydraulics code is all about basic floating point arithmetics, both algebraic and transcendental. No dynamic branching, very little memory traffic.

App 2. Ambient Occlusion code is a very mixed load of floating point and integer arithmetics, dynamic branching, texture sampling and memory access. Despite the memory traffic, this is a very compute intensive kernel.

App 3. Running Sum code, in contrast to the above, is memory intensive. It shuffles data through at a high rate, not doing much calculations on it. It relies heavily on the on-chip L1 cache, though, so it's not a raw memory bandwidth test.

App 4. Geometry Sampling code is texture sampling intensive. It sweeps through geometry data in "waves", and stresses samplers, texture caches, and memory equally. It also has a high register usage and thus low occupancy.

Minimise App 2 type workloads, Kepler's would perform fine.

[QUOTE="ronvalencia"]

Actually, the 7770 has less usable power than the Xbox One since lesser GCNs with less internal SRAM storage has greater chances for compute overspill. 7770 also has less triangle rate than 7790 or 7850.

My 8870M takes a greater performance hit when I lower the memory speed to DDR3 levels i.e. AMD's GDDR5 bar graphs are applicable for these cases.

-----

7790 doesn't sport X1's memory bandwidth and the prototype 7850 (with 12 CUs) is a closer fit to the X1.

X1's GCN has Pitcairn's crossbar, 256bit memory controllers and related L2 cache. L2 cache bandwidth is important since this is used as a shared cache across 4 CUs.

7790/7790 has less I/O L2 cache bandwidth than the Pitcairn's version.

http://www.behardware.com/articles/848-4/amd-radeon-hd-7970-crossfirex-review-28nm-and-gcn.html

GCN Cache Hierarchy

With 128bit memory controllers, 7790/7770 has two L2 blocks. That's 64 bytes per cycle per L2 block x 2 = 128 bytes per cycle.

With 256bit memory controllers, 7850/7870 has four L2 blocks. That's 64 bytes per cycle per L2 block x 4 = 256 bytes per cycle.

tormentos

The broken song again,stop the damn 660Ti has 100Gb/s of bandwidth less than the 7950 and still out perform it in some test...:lol:

You have no argument,i demostrated how a GPU with 100Gb/s less bandwidth than another can still beat it,so the xbox one miserable 37gb/s over the 7790 should be even less of a problem.

The 7790 is very close to the 7850 even that it has 57Gb/s less in bandwidth,so yeah bandwidth is not the issues is power,so yeah the 7790 >> xbox one..

No matter how you slice it what you try to imply is idiotic,if you put 250Gb/s bandwidth on the xbox one would it beat the 7950.? Just because it has more bandwidth.?

You demonstrated a different GPU design.

For 7950, L1 bandwidth has 64 byte per cycle per CU @ 800Mhz = 47 GB/s x 28 CUs = 1.3 TB/s.

Your "250Gb/s bandwidth on the xbox one would it beat the 7950" is just plain stupidity. Stop adding words to my posts.

Again, you can't read a simple bar graph that shows a prototype 7850 with 768 stream processors.

[QUOTE="tormentos"]

[QUOTE="ronvalencia"]

7770/7790 @ 800Mhz, L2 cache I/O bandwidth: 128 bytes x 800Mhz = 95.367 GB/s External memory setup faster than 95.367 GB/s would be useless.

W5000/7850-768 @ 853hz, L2 cache I/O bandwidth: 256 bytes x 853Mhz = 203.37 GB/s. ESRAM was design for Pitcairn level I/O.

04dcarraher

660Ti =144Gb/s

7950 = 240Gb/s..

Bandwidth mean nothing when you have no power,and the xbox one doesn't have it..

Trying to imply that the xbox one would need 203Gb/s is down right stupid it will not.

The useless memory bandwidth increase is due to a bottleneck in the system.

[QUOTE="ronvalencia"]

1. PC is not limited to 1080p.

2. It's minor.

When it's possible, I play games at 5760x1080p on my 7950 (with 900mhz firmware) @ 1000Mhz which is higher than the reference 7950 @ 800Mhz.

tormentos

Stop it you blind biased fanboy that game is unplayable on both the 7950 and 660Ti at that resolution,so 2560x1600 is actually more playable and the difference is not even 2 frames,but nice to miss the point and change the argument..

The 660Ti has 144Gb/s bandwith is even lower in Bandwidth than the 7850 yet it perform as good or better in some test vs the 7950 which has 240Gb/s bandwidth,so yeah my was proven the 7790 >>> xbox one..

Nice to miss the point and changed the argument by using a different GPU design. Your not even factoring the GK104's cache designs.

My 7950 is NOT running at 800Mhz; it's running at 1Ghz (from 900Mhz "out-of-the box" 7950).

http://www.hardocp.com/article/2013/02/01/xfx_radeon_hd_7950_black_edition_video_card_review/8#.UhHM3dB--po

the stock XFX Radeon HD 7950 with the factory overclock stood its ground, and provided nearly identical performance to the Radeon HD 7970. The overclock we achieved pushed performance beyond the Radeon HD 7970.

XFX Radeon HD 7950 Black Edition is "out-of-the-box" 900Mhz version.

I have already shown you what a flagship PC GPU design (i.e. 7950 @ 950Mhz) with GDDR5-2400 + 1080p Tomb Raider.

With 660 Ti, you're missing the fact that the entry level GK104 is a class above Pitcairn Pro and it's clocked higher/higher GFLOPS than 7850.

[QUOTE="22Toothpicks"][QUOTE="AMD655"]

Do Tormentos and Ron not realise they have gone in circles 17 times throughout this thread?

ronvalencia

[QUOTE="ronvalencia"][QUOTE="22Toothpicks"] This literally happens in every PS4 v. X1 spec thread they post in. On a side note I am convinced that Ron is a robot of some form.btk2k2

Neither did you.

Your formulas didn't work for 8 ROPS and your multi-layer counter statement was a joke.

Formula all you want, you are pig ignorant to basic GPU understanding Ron, let alone anything more advanced. 

You demonstrated a different GPU design.

For 7950, L1 bandwidth has 64 byte per cycle per CU @ 800Mhz = 47 GB/s x 28 CUs = 1.3 TB/s.

Your "250Gb/s bandwidth on the xbox one would it beat the 7950" is just plain stupidity. Stop adding words to my posts.

Again, you can't read a simple bar graph that shows a prototype 7850 with 768 stream processors.

ronvalencia

[QUOTE="btk2k2"][QUOTE="ronvalencia"] Tormentos didn't do computer science 101.ronvalencia

Neither did you.

Your formulas didn't work for 8 ROPS and your multi-layer counter statement was a joke.

Memory bandwidth means squat after a certain point and the difference between the X1 with DDR3+esram and PS4's GDDR5 will be negligible. The difference seen will be the gpu's processing power04dcarraher

At least we agree for once now tell that to Ron he refuse to admit it,even that i show him how the 660Ti performed side by side with the 7950 hell even surpass it while having almost 100GB/s less bandwidth.

 

I didn't imply anything. My post includes information on CU's max compute processing memory bandwidth from L1 cache and L2 cache I/O.

 

For X1's GCN, L1 bandwidth has 64 byte per cycle per CU = 50.8425 GB/s x 12 CUs = 610.11 GB/s.

 

 

One should look for any bottlenecks and seek the reasons on why it doesn't use extra memory bandwidth.

 

I have shown you why 7770/7790 @ 800 Mhz would gimp any memory faster than ~96 GB/s. AMD would need to redesign 7770/7790's crossbar/memory controller/L2 cache stack for faster memory based on GDDR6. For the X1, it's easier to "copy-and-paste" from the Pitcairn's crossbar/L2 cache/memory controller designs.

 

7770/7790 has a massive crossbar/L2 cache I/O bottleneck that gimps higher memory bandwidth e.g. not ready for GDDR6.

 

Reference http://www.behardware.com/articles/848-4/amd-radeon-hd-7970-crossfirex-review-28nm-and-gcn.html

------------------------

 

NVIDIA Kelper is different fish when it comes to handling it's compute workloads.

Entry level GK104 (e.g 660 TI) is another class of GPU above AMD Pitcairn i.e. it needs AMD Tahiti LE to battle NV 660 Ti.

K20 vs 7970 vs GTX680 vs M2050 vs GTX580 from http://wili.cc/blog/gpgpu-faceoff.html

 

K20 = GK110

 

There's very difference between NVIDIA's OpenCL and CUDA.

 

App 1. Digital Hydraulics code is all about basic floating point arithmetics, both algebraic and transcendental. No dynamic branching, very little memory traffic.

App 2. Ambient Occlusion code is a very mixed load of floating point and integer arithmetics, dynamic branching, texture sampling and memory access. Despite the memory traffic, this is a very compute intensive kernel.

App 3. Running Sum code, in contrast to the above, is memory intensive. It shuffles data through at a high rate, not doing much calculations on it. It relies heavily on the on-chip L1 cache, though, so it's not a raw memory bandwidth test.

App 4. Geometry Sampling code is texture sampling intensive. It sweeps through geometry data in "waves", and stresses samplers, texture caches, and memory equally. It also has a high register usage and thus low occupancy.

 

 

Minimise App 2 type workloads, Kepler's would perform fine.

ronvalencia

W5000/7850-768 @ 853hz, L2 cache I/O bandwidth: 256 bytes x 853Mhz = 203.37 GB/s. ESRAM was design for Pitcairn level I/O.

Yeah you did..

In fact a W5000 768 @853 mhz = and over clock GPU.

The speed of the W5000 is 825mhz to 853mhz would mean is slightly over clocked,so how many over clocked GPU have you see on consoles.?

ESRAM wasn't design for Pitcairn ESRAM is a damn band aid use by MS to help those starved components because MS chose DDR3,if the xbox one used GDDR5 even 102Gbs the xbox now would not even had ESRAM.

So for you a gimped 7790 will perform better than a 7950 if you give it more bandiwdth you are a joke,i proved my point 100GB/s less yet the performance is the same.

 

Nice to miss the point and changed the argument by using a different GPU design. Your not even factoring the GK104's cache designs.

 

 

My 7950 is NOT running at 800Mhz; it's running at 1Ghz (from 900Mhz "out-of-the box" 7950).

http://www.hardocp.com/article/2013/02/01/xfx_radeon_hd_7950_black_edition_video_card_review/8#.UhHM3dB--po

the stock XFX Radeon HD 7950 with the factory overclock stood its ground, and provided nearly identical performance to the Radeon HD 7970. The overclock we achieved pushed performance beyond the Radeon HD 7970.

 

XFX Radeon HD 7950 Black Edition is "out-of-the-box" 900Mhz version.

 

I have already shown you what a flagship PC GPU design (i.e. 7950 @ 950Mhz) with GDDR5-2400 + 1080p Tomb Raider.

 

With 660 Ti, you're missing the fact that the entry level GK104 is a class above Pitcairn Pro and it's clocked higher/higher GFLOPS than 7850.

 

ronvalencia

You are a joke of a poster where the fu** did i say that your 7950 was stock.? I compare the stock 7950 vs the stock 660Ti for a fair comparison,and the excuse you use against the 660Ti is a joke because the 7950 is not a Pitcairn pro is Tahiti,so yeah even with 100GB/s bandwidth the 660Ti beat the 7950 i could care less about your particular 7950 since i used the first model that came and not the one that suited you best.

My point was to preve Bandwidth mean sh** not that the 7950 is sh**.

:lol:

How hypocrite you are you accuse me of using different hardware for a comparison,when up until today without a single link to back you up you claim the xbox one has a W5000 no wait it has an over clocked W5000 since we all know the GPU clock on the W5000 is not 1ghz is 825mhz,so yeah you want to imply that the xbox one has an over clocked Pitcairn..:lol:

You loss the arguments be a fu**ing man and grow up admit that you loss.

.

.

marklarmer

If only Microsoft made their GPU at least 16 CUs instead of 12.

How much would've been the cost to Microsoft if they added 4 more CUs to their design?

also, Ronvalencia,

That last video you posted about physix contributed to convincing myself of taking the plunge and ordering nVidia Titan from amazon (couldn't wait for the 8xxx series as it's taking AMD forever to release it).

Athough my current HD 6990 runs pretty much everything maxed out with superb framerates, I wanted a signle GPU solution as my HD 6990 is having issues with the Oculus Dev Kit (inability to clone monitors in full screen mode due to the dual GPU nature of the card) which is forcing audio to go through only the active screen (i.e. I won't have audio if I play a game through the Oculus).

If only Microsoft made their GPU at least 16 CUs instead of 12.

How much would've been the cost to Microsoft if they added 4 more CUs to their design?

also, Ronvalencia,

That last video you posted about physix contributed to convincing myself of taking the plunge and ordering nVidia Titan from amazon (couldn't wait for the 8xxx series as it's taking AMD forever to release it).

Athough my current HD 6990 runs pretty much everything maxed out with superb framerates. But I just wanted a signle GPU solution as my HD 6990 is having issues with the Oculus Dev Kit (innability to clone monitors in full screen mode due to the dual GPU nature of the card) which is forcing audio to go through only the active screen (i.e. I won't have audio if I play a game through the Oculus).

Mystery_Writer

I think the GPU on the X1 has less CUs mostly because of die space. The eSRAM takes 1/4 of the chip size, so they had to cut the number of CUs down to 12.

If only Microsoft made their GPU at least 16 CUs instead of 12.

How much would've been the cost to Microsoft if they added 4 more CUs to their design?

also, Ronvalencia,

That last video you posted about physix contributed to convincing myself of taking the plunge and ordering nVidia Titan from amazon (couldn't wait for the 8xxx series as it's taking AMD forever to release it).

Athough my current HD 6990 runs pretty much everything maxed out with superb framerates. But I just wanted a signle GPU solution as my HD 6990 is having issues with the Oculus Dev Kit (innability to clone monitors in full screen mode due to the dual GPU nature of the card) which is forcing audio to go through only the active screen (i.e. I won't have audio if I play a game through the Oculus).

Mystery_Writer

If you care about CUs units so much you would just buy a PC. 

If only Microsoft made their GPU at least 16 CUs instead of 12.

How much would've been the cost to Microsoft if they added 4 more CUs to their design?

also, Ronvalencia,

That last video you posted about physix contributed to convincing myself of taking the plunge and ordering nVidia Titan from amazon (couldn't wait for the 8xxx series as it's taking AMD forever to release it).

Athough my current HD 6990 runs pretty much everything maxed out with superb framerates. But I just wanted a signle GPU solution as my HD 6990 is having issues with the Oculus Dev Kit (innability to clone monitors in full screen mode due to the dual GPU nature of the card) which is forcing audio to go through only the active screen (i.e. I won't have audio if I play a game through the Oculus).

Mystery_Writer

Its not really the cost for more stream processors its the size of the die along with the constrains in the thermal output of the APU. Part of the problem is that MS wants to avoid another RRoD scandal, and the the decision to go with DDR3 instead of GDDR5 required MS to include a large buffer to offset the bandwidth for the system bus. You should have gotten a GTX 780 much better price to performance ratio.

04dcarraher

Actually I wanted to get the GTX 780, but didn't know if it'll be a noticable upgrade from my 2.5 year old HD 6990.

Somehow Titan, apparently based on older Geforce 600 technology, is outperforming the new 700 series single GPU cards (according to anandtech).

I'm puzzled as to why nVidia didn't release a 700 series with the same CUDA cores count of Titan.

Do you think Titan was a good choice? (aside from the price drawback)

Edit: Also, that's a really interesting take regarding why MS didn't increase the CU count. I thought it was purely due to trying to keep the design within certain cost bracket and Kinect upset that bracket so they skimped on CUs to compensate.

Edit 2: Is Titan considered a Geforce 600 or 700 series GPU?

[QUOTE="04dcarraher"]

Its not really the cost for more stream processors its the size of the die along with the constrains in the thermal output of the APU. Part of the problem is that MS wants to avoid another RRoD scandal, and the the decision to go with DDR3 instead of GDDR5 required MS to include a large buffer to offset the bandwidth for the system bus. You should have gotten a GTX 780 much better price to performance ratio.

Mystery_Writer

Actually I wanted to get the GTX 780, but didn't know if it'll be a noticable upgrade from my 2.5 year old HD 6990.

Somehow Titan, apparently based on older Geforce 600 technology, is outperforming the new 700 series single GPU cards (according to anandtech).

I'm puzzled as to why nVidia didn't release a 700 series with the same CUDA cores count of Titan.

Do you think Titan was a good choice? (aside from the price drawback)

Edit: Also, that's a really interesting take regarding why MS didn't increase the CU count. I thought it was purely due to trying to keep the design within certain cost bracket and Kinect upset that bracket so they skimped on CUs to compensate.

Here is performance summary for GTX 780

Here is performance summary for GTX 780

04dcarraher

You're 100% right. Here is what Anand had to say about the GTX 780

The end result is that with the GTX 780 delivering an average of 90% of Titans gaming performance for 65% of the price, this is by all rights the Titan Mini, the cheaper video card Titan customers have been asking for.

From that perspective the GTX 780 is nothing short of an amazing deal for the level of performance offered, especially since it maintains the high build quality and impressive acoustics that helped to define Titan.

anandtech

Although I don't know how much difference this will make, but the Titan I got is supposedly an overclocked edition link

btw, do you know by any chance the overall performance difference between HD 6990 and Titan?

[QUOTE="04dcarraher"]

Here is performance summary for GTX 780

Mystery_Writer

You're 100% right. Here is what Anand had to say about the GTX 780

The end result is that with the GTX 780 delivering an average of 90% of Titans gaming performance for 65% of the price, this is by all rights the Titan Mini, the cheaper video card Titan customers have been asking for.

From that perspective the GTX 780 is nothing short of an amazing deal for the level of performance offered, especially since it maintains the high build quality and impressive acoustics that helped to define Titan.

anandtech

Although I don't know how much difference this will make, but the Titan I got is supposedly an overclocked edition link

btw, do you know by any chance the overall performance difference between HD 6990 and Titan?

[QUOTE="04dcarraher"]

Its not really the cost for more stream processors its the size of the die along with the constrains in the thermal output of the APU. Part of the problem is that MS wants to avoid another RRoD scandal, and the the decision to go with DDR3 instead of GDDR5 required MS to include a large buffer to offset the bandwidth for the system bus. You should have gotten a GTX 780 much better price to performance ratio.

Mystery_Writer

Actually I wanted to get the GTX 780, but didn't know if it'll be a noticable upgrade from my 2.5 year old HD 6990.

Somehow Titan, apparently based on older Geforce 600 technology, is outperforming the new 700 series single GPU cards (according to anandtech).

I'm puzzled as to why nVidia didn't release a 700 series with the same CUDA cores count of Titan.

Do you think Titan was a good choice? (aside from the price drawback)

Edit: Also, that's a really interesting take regarding why MS didn't increase the CU count. I thought it was purely due to trying to keep the design within certain cost bracket and Kinect upset that bracket so they skimped on CUs to compensate.

Edit 2: Is Titan considered a Geforce 600 or 700 series GPU?

[QUOTE="Mystery_Writer"]

If only Microsoft made their GPU at least 16 CUs instead of 12.

How much would've been the cost to Microsoft if they added 4 more CUs to their design?

also, Ronvalencia,

That last video you posted about physix contributed to convincing myself of taking the plunge and ordering nVidia Titan from amazon (couldn't wait for the 8xxx series as it's taking AMD forever to release it).

Athough my current HD 6990 runs pretty much everything maxed out with superb framerates. But I just wanted a signle GPU solution as my HD 6990 is having issues with the Oculus Dev Kit (innability to clone monitors in full screen mode due to the dual GPU nature of the card) which is forcing audio to go through only the active screen (i.e. I won't have audio if I play a game through the Oculus).

ManatuBeard

I think the GPU on the X1 has less CUs mostly because of die space. The eSRAM takes 1/4 of the chip size, so they had to cut the number of CUs down to 12.

[QUOTE="ronvalencia"]

I didn't imply anything. My post includes information on CU's max compute processing memory bandwidth from L1 cache and L2 cache I/O.

For X1's GCN, L1 bandwidth has 64 byte per cycle per CU = 50.8425 GB/s x 12 CUs = 610.11 GB/s.

One should look for any bottlenecks and seek the reasons on why it doesn't use extra memory bandwidth.

I have shown you why 7770/7790 @ 800 Mhz would gimp any memory faster than ~96 GB/s. AMD would need to redesign 7770/7790's crossbar/memory controller/L2 cache stack for faster memory based on GDDR6. For the X1, it's easier to "copy-and-paste" from the Pitcairn's crossbar/L2 cache/memory controller designs.

7770/7790 has a massive crossbar/L2 cache I/O bottleneck that gimps higher memory bandwidth e.g. not ready for GDDR6.

Reference http://www.behardware.com/articles/848-4/amd-radeon-hd-7970-crossfirex-review-28nm-and-gcn.html

------------------------

NVIDIA">http://www.behardware.com/articles/848-4/amd-radeon-hd-7970-crossfirex-review-28nm-and-gcn.html">http://www.behardware.com/articles/848-4/amd-radeon-hd-7970-crossfirex-review-28nm-and-gcn.html

------------------------

NVIDIA Kelper is different fish when it comes to handling it's compute workloads.

Entry level GK104 (e.g 660 TI) is another class of GPU above AMD Pitcairn i.e. it needs AMD Tahiti LE to battle NV 660 Ti.

K20 vs 7970 vs GTX680 vs M2050 vs GTX580 from http://wili.cc/blog/gpgpu-faceoff.html

K20">http://wili.cc/blog/gpgpu-faceoff.html">http://wili.cc/blog/gpgpu-faceoff.html

K20 = GK110

There's very difference between NVIDIA's OpenCL and CUDA.

App 1. Digital Hydraulics code is all about basic floating point arithmetics, both algebraic and transcendental. No dynamic branching, very little memory traffic.

App 2. Ambient Occlusion code is a very mixed load of floating point and integer arithmetics, dynamic branching, texture sampling and memory access. Despite the memory traffic, this is a very compute intensive kernel.

App 3. Running Sum code, in contrast to the above, is memory intensive. It shuffles data through at a high rate, not doing much calculations on it. It relies heavily on the on-chip L1 cache, though, so it's not a raw memory bandwidth test.

App 4. Geometry Sampling code is texture sampling intensive. It sweeps through geometry data in "waves", and stresses samplers, texture caches, and memory equally. It also has a high register usage and thus low occupancy.

Minimise App 2 type workloads, Kepler's would perform fine.

tormentos

W5000/7850-768 @ 853hz, L2 cache I/O bandwidth: 256 bytes x 853Mhz = 203.37 GB/s. ESRAM was design for Pitcairn level I/O.

Yeah you did..

In fact a W5000 768 @853 mhz = and over clock GPU.

The speed of the W5000 is 825mhz to 853mhz would mean is slightly over clocked,so how many over clocked GPU have you see on consoles.?

ESRAM wasn't design for Pitcairn ESRAM is a damn band aid use by MS to help those starved components because MS chose DDR3,if the xbox one used GDDR5 even 102Gbs the xbox now would not even had ESRAM.

So for you a gimped 7790 will perform better than a 7950 if you give it more bandiwdth you are a joke,i proved my point 100GB/s less yet the performance is the same.

I didn't imply anything.

7950 @ 800Mhz, L2 cache I/O bandwidth: L2 cache I/O bandwidth: 364 bytes per cycle x 800 Mhz = 300 GB/s.

7950 @ 950Mhz (fastest out-of-the-box edition), L2 cache I/O bandwidth: L2 cache I/O bandwidth: 364 bytes per cycle x 950 Mhz = ~337.7GB/s.

X1's L2 cache I/O bandwidth is not even close to 7950 @800Mhz. Your wrong again.

AMD.com's W5000 @ 1.3 TFLOPS would need about 853Mhz clock speed. There are W5000s with slightly clock speeds. AMD doesn't ultimately dictate the final GPU clock speed i.e. that's AIB (Add In Board)'s or OEM's decisions.

10Mhz overclock is minor i.e. I overclocked my 7950-900mhz to 950Mhz without any voltage increase. MS overclocked X1's GPU by 53 Mhz. Your issue on this subject is just noise i.e. it's a "who cares" episode.

-------

X1 has GCN's 256bit memory controllers and related L2 cache I/O bandwidth i.e. 203.37 GB/s which can cover 133 GB/s (via ESRAM) alpha blend was the effective bandwidth for the prototype X1.

A 7770 or 7790's L2 cache I/O bandwidth would not cover 133 GB/s (via ESRAM) alpha blend. This is why the prototype-7850 with 768 stream processors @ 860Mhz is a closer fit to the X1's GCN.

Giving 7790 more bandwidth is nearly pointless since it's 128byte per cycle L2 cache I/O would gimp it. Each memory controller, has 64byte per cycle L2 cache bandwidth. The joke is you since you haven't identified the bottleneck in the 7790.

Also, 7950 has substantially larger internal SRAM storage than 7790.

If only Microsoft made their GPU at least 16 CUs instead of 12.

How much would've been the cost to Microsoft if they added 4 more CUs to their design?

also, Ronvalencia,

That last video you posted about physix contributed to convincing myself of taking the plunge and ordering nVidia Titan from amazon (couldn't wait for the 8xxx series as it's taking AMD forever to release it).

Athough my current HD 6990 runs pretty much everything maxed out with superb framerates, I wanted a signle GPU solution as my HD 6990 is having issues with the Oculus Dev Kit (inability to clone monitors in full screen mode due to the dual GPU nature of the card) which is forcing audio to go through only the active screen (i.e. I won't have audio if I play a game through the Oculus).

Mystery_Writer

[QUOTE="04dcarraher"] Memory bandwidth means squat after a certain point and the difference between the X1 with DDR3+esram and PS4's GDDR5 will be negligible. The difference seen will be the gpu's processing powertormentos

At least we agree for once now tell that to Ron he refuse to admit it,even that i show him how the 660Ti performed side by side with the 7950 hell even surpass it while having almost 100GB/s less bandwidth.

Not for Eyefinity level resolutions when 240 GB/s (theoretical) memory bandwidth is actually count for something. You haven't applied the principle on why server CPUs has larger caches.

[QUOTE="ronvalencia"]You demonstrated a different GPU design.

For 7950, L1 bandwidth has 64 byte per cycle per CU @ 800Mhz = 47 GB/s x 28 CUs = 1.3 TB/s.

Your "250Gb/s bandwidth on the xbox one would it beat the 7950" is just plain stupidity. Stop adding words to my posts.

Again, you can't read a simple bar graph that shows a prototype 7850 with 768 stream processors.

btk2k2

[QUOTE="btk2k2"][QUOTE="ronvalencia"]You demonstrated a different GPU design.

For 7950, L1 bandwidth has 64 byte per cycle per CU @ 800Mhz = 47 GB/s x 28 CUs = 1.3 TB/s.

Your "250Gb/s bandwidth on the xbox one would it beat the 7950" is just plain stupidity. Stop adding words to my posts.

Again, you can't read a simple bar graph that shows a prototype 7850 with 768 stream processors.

ronvalencia

 

Not for Eyefinity level resolutions when 240 GB/s (theoretical) memory bandwidth is actually count for something. You haven't applied the principle on why server CPUs has larger caches.

 

ronvalencia

Not for a resolution that basically killed both card because the test you showed killed both cards the performance was under 30 FPS on PC that is horrible since PC has higher variable frame rate than console.

 Don't come with the whole theorical crap the 660Ti 144Gb/s is also theorical all peak on everything is theorical,so save your lame excuses.

[QUOTE="Mystery_Writer"]

If only Microsoft made their GPU at least 16 CUs instead of 12.

How much would've been the cost to Microsoft if they added 4 more CUs to their design?

also, Ronvalencia,

That last video you posted about physix contributed to convincing myself of taking the plunge and ordering nVidia Titan from amazon (couldn't wait for the 8xxx series as it's taking AMD forever to release it).

Athough my current HD 6990 runs pretty much everything maxed out with superb framerates. But I just wanted a signle GPU solution as my HD 6990 is having issues with the Oculus Dev Kit (innability to clone monitors in full screen mode due to the dual GPU nature of the card) which is forcing audio to go through only the active screen (i.e. I won't have audio if I play a game through the Oculus).

xhawk27

If you care about CUs units so much you would just buy a PC. 

[QUOTE="ManatuBeard"]

[QUOTE="Mystery_Writer"]

If only Microsoft made their GPU at least 16 CUs instead of 12.

How much would've been the cost to Microsoft had they added 4 more CUs to their design?

ronvalencia

I think the GPU on the X1 has less CUs mostly because of die space. The eSRAM takes 1/4 of the chip size, so they had to cut the number of CUs down to 12.

Ron, if you were to design the X1 APU, which solution would you pick as the more elegant architecture;

a) 12 CUs + ESRAM + DDR3

b) 18 CUs + GDDR5

[QUOTE="ronvalencia"]

Not for Eyefinity level resolutions when 240 GB/s (theoretical) memory bandwidth is actually count for something. You haven't applied the principle on why server CPUs has larger caches.

tormentos

Not for a resolution that basically killed both card because the test you showed killed both cards the performance was under 30 FPS on PC that is horrible since PC has higher variable frame rate than console.

Don't come with the whole theorical crap the 660Ti 144Gb/s is also theorical all peak on everything is theorical,so save your lame excuses.

660 Ti doesn't win on AMD bias games. You not factoring ALU behaviour differences.

----

660Ti has a larger 64KB L1 cache per SMX unit that covers 192 stream processors. There are 7 SMX units = 448 KB L1 cache @ 915Mhz.

7850 has 16 CUs x 16 KB L1 cache = 256 KB L1 cache @ 860Mhz.

7950@ 800 Mhz: 28 CU x 16 KB L1 = 448 KB @ 800 Mhz.

When games are not bound by GK014's register file limits, GK104 can match Tahiti.

 

660 Ti doesn't win on AMD bias games. You not factoring ALU behaviour differences.

----

660Ti has a larger 64KB L1 cache per SMX unit that covers 192 stream processors. There are 7 SMX units = 448 KB L1 cache @ 915Mhz.

7850 has 16 CUs x 16 KB L1 cache = 256 KB L1 cache @ 860Mhz.

 

7950@ 800 Mhz: 28 CU x 16 KB L1 = 448 KB @ 800 Mhz.

 

When games are not bound by GK014's register file limits, GK104 can match Tahiti.ronvalencia

But on other games it does,with 100GB/s less there is not a single excuse you can use to justify it,bandwidth mean little when your GPU is weak i proved that.

 ESRAM consumed ~32 precent from the ~5 billion transistor budget i.e. 1.6 billion for T6 SRAM. Mystery_Writer

Ron, if you were to design the X1 APU, which solution would you pick as the more elegant architecture;

a) 12 CUs + ESRAM + DDR3

b) 18 CUs + GDDR5

Oh he saw your post and ignores it because everything he knows about PC tell him that.

18CU +GDDR5 >>> 12CU +ESRAM+ DDR3.

Shadow of the Beast is PS4 exclusive toodrakekratos

[QUOTE="drakekratos"]Shadow of the Beast is PS4 exclusive too04dcarraher

[QUOTE="Mystery_Writer"]

 ESRAM consumed ~32 precent from the ~5 billion transistor budget i.e. 1.6 billion for T6 SRAM. tormentos

Ron, if you were to design the X1 APU, which solution would you pick as the more elegant architecture;

a) 12 CUs + ESRAM + DDR3

b) 18 CUs + GDDR5

 Oh he saw your post and ignores it because everything he knows about PC tell him that.

18CU +GDDR5 >>> 12CU +ESRAM+ DDR3.

So I wonder what MS is going to to talk about at Hot Chips?

http://www.hotchips.org/

An off the shelf AMD 7790 gimped GPU? They may get laughed out of there huh.

[QUOTE="ronvalencia"]

660 Ti doesn't win on AMD bias games. You not factoring ALU behaviour differences.

----

660Ti has a larger 64KB L1 cache per SMX unit that covers 192 stream processors. There are 7 SMX units = 448 KB L1 cache @ 915Mhz.

7850 has 16 CUs x 16 KB L1 cache = 256 KB L1 cache @ 860Mhz.

7950@ 800 Mhz: 28 CU x 16 KB L1 = 448 KB @ 800 Mhz.

When games are not bound by GK014's register file limits, GK104 can match Tahiti.tormentos

But on other games it does,with 100GB/s less there is not a single excuse you can use to justify it,bandwidth mean little when your GPU is weak i proved that.

Your not comparing apple to apples. Both X1 and PS4 doesn't have NVIDIA Kepler GK104.

For 5760x1080p resolutions, 660 Ti doesn't scale as good as 7950. BattleField 3's 26.5 fps results can be made to reach 30 fps by slightly reducing the detail settings (a PC user feature) e.g. shadows @ ultra to high, MSAA 4X to 2X. Also, Techpowerup did not use the latest 7950 Boost Edition (aka 8950) variant.

-----

NV SMX's L1 cache/shared memory can do 256 byte per cycle.

660 Ti's 7 SMX x 256 byte per cycle x 915Mhz = 1.6 TB/s. L1/shared memory size is 448 KB.

AMD CU's L1 cache can do 64 byte per cycle.

7950's 28 CU x 64 byte per cycle x 800 Mhz = 1.4 TB/s, L1/shared memory size is 448 KB.

Again, you are not factoring GK104's SRAM storage capability and it's clocked faster than 7950 @800Mhz.

There's a reason why AMD updated 7950 into 7950 Boost Edition. My 7950 doesn't need to be updated into 7950 BE since it's already clocked at 900Mhz out-of-the-box.

Larger SRAM storage = less trip to the external memory.

AMD has somewhat duplicated 660 Ti's setup with 7870 XT (925Mhz with 975Mhz boost, 192 GB/s). Again, AMD updated 7950 to 7950 Boost Edition so it wouldn't be in conflict with 7870 XT.

The prototype 7850 with 768 stream processors would be closer to X1's GCN.

[QUOTE="ronvalencia"][QUOTE="ManatuBeard"]

I think the GPU on the X1 has less CUs mostly because of die space. The eSRAM takes 1/4 of the chip size, so they had to cut the number of CUs down to 12.

Mystery_Writer

Ron, if you were to design the X1 APU, which solution would you pick as the more elegant architecture;

a) 12 CUs + ESRAM + DDR3

b) 18 CUs + GDDR5

Selection B. It's closer to my PC specs.

Since Q2 2013, Intel is already using PS4's high density GDDR5 with their Xeon Phi products i.e. 512bit 16 GB GDDR5, hence 8 GB GDDR5 at 256bit.

http://ark.intel.com/products/75799/Intel-Xeon-Phi-Coprocessor-7120P-16GB-1_238-GHz-61-core

Oh he saw your post and ignores it because everything he knows about PC tell him that.

18CU +GDDR5 >>> 12CU +ESRAM+ DDR3.

tormentos

Fictional post.

I haven't bought any GCN products with non-GDDR5.

Formula all you want, you are pig ignorant to basic GPU understanding Ron, let alone anything more advanced. 

AMD655