hd6670
480 shaders(each can perform 1 each of floating point multiply and add per clock)
24 texture units(includes address units to fetch texture data for the pixels making
up the surface of polygons and sample units for up to 16x anisotropic filtering/128 tap to
correctly render textures on object not at 90 or 45 degrees angles and at a distance)
8 rops(render back ends, mostly for calculating multi sample anti- aliasing/jaggie
reduction and i think high dynamic range lighting, or is that shaders too?)
800mhz gpu clock equals theoretical 768gflops(billions of floating point operations/sec)
128 bit memory interface(at 8 bits per byte is 16 bytes per clock)
gddr3 usually clocked at 800mhz(1600 double data rate) so about 25.6GB/s bandwidth.
gddr5 usually @1ghz(gddr5 is 4 transfers per clock so 4 giga transfers/sec) about 64GB/s.
bandwidth is used mostly in calculating aa( send and receive samples from frame buffer)
2GB gddr5 frame buffer( stores high res texture data, aa sample data, and possibly
several finished frames when buffering to enhance smooth rendering when the number
of frames the graphics card is completing does not match monitor refresh rate)
incidentally, it seems to me that 1GB of frame buffer is most appropriate for resolutions
of about 2 megapixels(like your monitor), 2GB for those higher res 27-30in panels
(about 4 megapixels) and 3GB for 3 panels at 1080p( about 6 megapixels). 2GB of
frame buffer on a 6670 gddr3 is total overkill, as the narrow bandwidth and small
number of rops at lower clock speed will not be fast enough to fill it up and provide
playable framerates. on to 6770 specs and will throw 6750 in too(it's between them)
(actually, a 512MB frame buffer might be more appropriate for cards such as the
hd6670 and 5670, but they may be able to use 1GB in isolated incidents. 1GB is pretty
much a standard minimum for all cards nowadays, whether they can use it or not)
hd6770(rebrand of 5770)
800 shaders
40 texture units
16 rops
850mhz gpu clock(1360gflops)
128 bit memory interface
gddr5 @1.2ghz(4.8 giga transfers/sec) x 128 bit interface is about 76.8GB/s.
1GB gddr5 frame buffer(quite appropriate for the card's rendering power and most
likely target resolutions)
hd6750(rebrand of 5750)
720 shaders
36 texture units
16 rops
700 mhz gpu clock(1008gflops)
128bit memory interface
gddr5 @1.15ghz(4.6 giga transfers/sec) x 128 bit interface is about 73.6GB/s.
512MB-1GB gddr5 frame buffer(the gpu can take advantage of this amount of
frame buffer with its level of performance)
would like to mention about restructuring of shaders in the last coupler generations
of radeon GPUs. for a long time, amd used vliw5(very long instruction word 5) where
4 out of 5 shaders in a group work on a pixel(each pixel up to 32 bit precision in red,
green, blue and the alpha transparency channel) with the last fifth fatter shader
performing the occasional transcendental function like sine, cosine, tangent, etc.
these groups of five were further grouped into 16 vliw 5s, for a block of 80 shaders
times 1(80), 2(160), 4(320), 5(400), 6(480), 9(720) 10(800), 12(960), 14(1120), 18
(1440) and 20(1600) depending on the gpu in question and if units were disable due
to binning. what amd found out was that that last unit was rarely used, so
with the hd6900 cards, amd switched to vliw4, dropping the last fat unit
and applying its extra parts to the other 4 shaders in the group. this made them all
a little fatter and you would loose a bit of performance when a group of 4 was used
to calculate a special function, but they were able to put more 16x4 bunches on the
silicon without much extra size increase at 40nm(hd6950 w/64x22 or 1408 shaders,
6970 w/64x24 or 1536 shaders). these cards also got a boost in clock speed compared
to hd5850(1440 shaders or 80x18) and hd5870(1600 or 80x20). i predicted the change
would result in a 20-25% increase in performance per theoretical gflop, and i was right.
hd6950 @800mhz with 2253gflops performs generally equal to or better than hd5870
@850mhz with 2720 gflops. they both have 32 rops and a 1-2GB frame buffer. the only
other difference is the # of texture units( always 4 for a block of 64-80 shaders), but
since even bottom level cards can do 16xaf w/o issues, i see this as irrelevant.
with the low/mid-high end hd7000 cards(lower cards are rebrands of older GPUs), they
enhanced performance more by making each shader block of 64 go from 16 groups of
4 shaders each to 4 groups of 16 shaders, making the architecture more similar to what
nvidia has been doing since the geforce 8 series. this is part of what allows a card such
as the hd7850 to perform as well at times as an hd6970 or gtx570(the two older cards
used to slug it out in price/performance from us315-350) despite an otherwise similar
architecture to 6970, lower core and vram clocks and massively reduced theoretical
gflop performance. this from a card that started life at us250 and can now be had for
a little under us200. pretty amazing, and newer drivers have been making it even better.
find out more clicking links to pages for recent radeon families on page here
http://en.wikipedia.org/wiki/Comparison_of_AMD_graphics_processing_units
might you possibly reconsider keeping your current psu and getting 1 of these hd7750s?
vtx3d 81.70 pounds with free delivery in the UK(i know nothing of this company)
http://www.amazon.co.uk/VTX3D-Graphics-128-Bit-PCI-Express-Architecture/dp/B0078XMRF8/ref=sr_1_2?ie=UTF8&qid=1346100314&sr=8-2
club3d 82.58 pounds with free delivery in the uk( decent company rep in th past)
http://www.amazon.co.uk/Club3D-Radeon-GDDR5-128-Bit-Graphics/dp/B007A3SQ7Y/ref=sr_1_7?ie=UTF8&qid=1346100314&sr=8-7
doesn't say whether these prices are before of after VAT( that's an issue, right?).
480 shaders(each can perform 1 each of floating point multiply and add per clock)
24 texture units(includes address units to fetch texture data for the pixels making
up the surface of polygons and sample units for up to 16x anisotropic filtering/128 tap to
correctly render textures on object not at 90 or 45 degrees angles and at a distance)
8 rops(render back ends, mostly for calculating multi sample anti- aliasing/jaggie
reduction and i think high dynamic range lighting, or is that shaders too?)
800mhz gpu clock equals theoretical 768gflops(billions of floating point operations/sec)
128 bit memory interface(at 8 bits per byte is 16 bytes per clock)
gddr3 usually clocked at 800mhz(1600 double data rate) so about 25.6GB/s bandwidth.
gddr5 usually @1ghz(gddr5 is 4 transfers per clock so 4 giga transfers/sec) about 64GB/s.
bandwidth is used mostly in calculating aa( send and receive samples from frame buffer)
2GB gddr5 frame buffer( stores high res texture data, aa sample data, and possibly
several finished frames when buffering to enhance smooth rendering when the number
of frames the graphics card is completing does not match monitor refresh rate)
incidentally, it seems to me that 1GB of frame buffer is most appropriate for resolutions
of about 2 megapixels(like your monitor), 2GB for those higher res 27-30in panels
(about 4 megapixels) and 3GB for 3 panels at 1080p( about 6 megapixels). 2GB of
frame buffer on a 6670 gddr3 is total overkill, as the narrow bandwidth and small
number of rops at lower clock speed will not be fast enough to fill it up and provide
playable framerates. on to 6770 specs and will throw 6750 in too(it's between them)
(actually, a 512MB frame buffer might be more appropriate for cards such as the
hd6670 and 5670, but they may be able to use 1GB in isolated incidents. 1GB is pretty
much a standard minimum for all cards nowadays, whether they can use it or not)
hd6770(rebrand of 5770)
800 shaders
40 texture units
16 rops
850mhz gpu clock(1360gflops)
128 bit memory interface
gddr5 @1.2ghz(4.8 giga transfers/sec) x 128 bit interface is about 76.8GB/s.
1GB gddr5 frame buffer(quite appropriate for the card's rendering power and most
likely target resolutions)
hd6750(rebrand of 5750)
720 shaders
36 texture units
16 rops
700 mhz gpu clock(1008gflops)
128bit memory interface
gddr5 @1.15ghz(4.6 giga transfers/sec) x 128 bit interface is about 73.6GB/s.
512MB-1GB gddr5 frame buffer(the gpu can take advantage of this amount of
frame buffer with its level of performance)
would like to mention about restructuring of shaders in the last coupler generations
of radeon GPUs. for a long time, amd used vliw5(very long instruction word 5) where
4 out of 5 shaders in a group work on a pixel(each pixel up to 32 bit precision in red,
green, blue and the alpha transparency channel) with the last fifth fatter shader
performing the occasional transcendental function like sine, cosine, tangent, etc.
these groups of five were further grouped into 16 vliw 5s, for a block of 80 shaders
times 1(80), 2(160), 4(320), 5(400), 6(480), 9(720) 10(800), 12(960), 14(1120), 18
(1440) and 20(1600) depending on the gpu in question and if units were disable due
to binning. what amd found out was that that last unit was rarely used, so
with the hd6900 cards, amd switched to vliw4, dropping the last fat unit
and applying its extra parts to the other 4 shaders in the group. this made them all
a little fatter and you would loose a bit of performance when a group of 4 was used
to calculate a special function, but they were able to put more 16x4 bunches on the
silicon without much extra size increase at 40nm(hd6950 w/64x22 or 1408 shaders,
6970 w/64x24 or 1536 shaders). these cards also got a boost in clock speed compared
to hd5850(1440 shaders or 80x18) and hd5870(1600 or 80x20). i predicted the change
would result in a 20-25% increase in performance per theoretical gflop, and i was right.
hd6950 @800mhz with 2253gflops performs generally equal to or better than hd5870
@850mhz with 2720 gflops. they both have 32 rops and a 1-2GB frame buffer. the only
other difference is the # of texture units( always 4 for a block of 64-80 shaders), but
since even bottom level cards can do 16xaf w/o issues, i see this as irrelevant.
with the low/mid-high end hd7000 cards(lower cards are rebrands of older GPUs), they
enhanced performance more by making each shader block of 64 go from 16 groups of
4 shaders each to 4 groups of 16 shaders, making the architecture more similar to what
nvidia has been doing since the geforce 8 series. this is part of what allows a card such
as the hd7850 to perform as well at times as an hd6970 or gtx570(the two older cards
used to slug it out in price/performance from us315-350) despite an otherwise similar
architecture to 6970, lower core and vram clocks and massively reduced theoretical
gflop performance. this from a card that started life at us250 and can now be had for
a little under us200. pretty amazing, and newer drivers have been making it even better.
find out more clicking links to pages for recent radeon families on page here
http://en.wikipedia.org/wiki/Comparison_of_AMD_graphics_processing_units
might you possibly reconsider keeping your current psu and getting 1 of these hd7750s?
vtx3d 81.70 pounds with free delivery in the UK(i know nothing of this company)
http://www.amazon.co.uk/VTX3D-Graphics-128-Bit-PCI-Express-Architecture/dp/B0078XMRF8/ref=sr_1_2?ie=UTF8&qid=1346100314&sr=8-2
club3d 82.58 pounds with free delivery in the uk( decent company rep in th past)
http://www.amazon.co.uk/Club3D-Radeon-GDDR5-128-Bit-Graphics/dp/B007A3SQ7Y/ref=sr_1_7?ie=UTF8&qid=1346100314&sr=8-7
doesn't say whether these prices are before of after VAT( that's an issue, right?).
Facebook
Twitter
Instagram