News - NVIDIA GeForce GTX 480: pricing and specifications leaked?

[HEXUS]

The latest batch of Fermi leaks shed light on core, shader and memory frequencies, and we've speculative pricing, too.

Read more.

[kalniel]

How does the 5970 end up with 256GB/s memory bandwidth? It should be half that surely - it's just a 256bit bus after all (not x2 as each GPU only accesses it's own set of memory chips)

[Biscuit]

so its about 10% more in cost than its ATI counterpart according to those price estimations - which actually relates to the suggested performance increase. Seems optomistic as it is added to the fact ATI cards could easily be dropped in price.

[dangel]

How does the 5970 end up with 256GB/s memory bandwidth? It should be half that surely - it's just a 256bit bus after all (not x2 as each GPU only accesses it's own set of memory chips)

I suppose you could argue that *2 is valid, given they combine to render.

Interesting though, both top end cards are *2 what I have now choices.. Then again.. what I do like is the 480 is a single chip, which will always therefore scale.. Plus it's cheaper. Oh god.

[GheeTsar]

Fudzilla (not a reliable source mind) have the UK price for the 480 as £429 and the 470 as £299:

http://www.fudzilla.com/content/view/18145/1/

Bargains!

[scaryjim]

Let's just run with those rumoured prices for a minute...

£329 + VAT = £387
£230 + VAT = £270

This being britain, that probably means £399 and £289, and you'll need to add on the early adopter tax for the first couple of weeks, as well. EDIT: that'd make the fudzilla claims about right, then?

At the minute, 5870s can be easily picked up for ~ £320. 5850s seem to be in much shorter supply, and will set you back ~ £250.

So based on that, the GTX470 will only be ~ 10% more expensive than a 5850, for a claimed "up to" 10% faster GPU. Pricing not too bad.

The GTX480, however, will be the best part of 25% more expensive, for a similar "up to" 10% faster. Pricing, not so hot... (shame we won't be able to say the same about the GPU )

[kalniel]

I suppose you could argue that *2 is valid, given they combine to render.

The memory doesn't combine. Each rendered part-scene needs its own isolated buffer. So you should really consider it a 2x1gb card too It's just one of those cases where 2x1gb doesn't equal 2gb. Like 4 times 3ghz in a quadcore doesn't equal 12ghz

[j.o.s.h.1408]

plus the 480gtx max load is 300w as a posed to 188w for the 5870

[scaryjim]

The memory doesn't combine. Each rendered part-scene needs its own isolated buffer. So you should really consider it a 1gb card too

But presumably both GPUs can access their own frame buffer simultaneously? Which should, externally, appear equivalent...

[kalniel]

But presumably both GPUs can access their own frame buffer simultaneously? Which should, externally, appear equivalent...

But they're buffering duplicate information - each GPU needs its own copy of the textures, maps, etc. for the whole scene. You can't fit anything more in there than you could if it was a single card.

[borandi]

Wait a second...

The 480 has 480 FPU units? But it has 512 cuda cores, thus 512 FPU units. It's always been known to have 512.

[scaryjim]

But they're buffering duplicate information - each GPU needs its own copy of the textures, maps, etc. for the whole scene. You can't fit anything more in there than you could if it was a single card.

But presumably each GPU can access the bits it needs for each part-scene it renders simultaneously, so the bandwidth (externally at least) should appear to be 2x? Although I confess I'm not an expert on GPUs Presumably the good scaling on the 5970 means that actually most games, even at high resolutions, don't max out a 1GB frame buffer at the minute?

Wait a second...

The 480 has 480 FPU units? But it has 512 cuda cores, thus 512 FPU units. It's always been known to have 512.

The original planned spec was for 512 cores, but manufacturing defects in TSMCs 40nm process meant they couldn't get them to run stably (at least, not at acceptable voltages / clock speeds), so they've pulled the specs in. ALLEDGEDLY.

Of course, if they can do a redesign that will negate the 40nm manufacturing defects (like ATI did after they released the 4770) the possibility exists that they might be able to put out a 512 core version later this year - maybe around Christmas? If they can, it should also have a lower power draw, because one of the (alledged) problems with this run is that the defects require extra voltage to improve stability. A proper redesign should allow them to reduce the voltage quite dramatically...

[kalniel]

But presumably each GPU can access the bits it needs for each part-scene it renders simultaneously

Nope, or rather: the 'bits it needs' is practically the whole scene anyway as far as things in memory go.

Presumably the good scaling on the 5970 means that actually most games, even at high resolutions, don't max out a 1GB frame buffer at the minute?

Correct.

[Platinum]

For the power draws the 5970 is looking more and more the card I am going to buy come pay day

[dangel]

The memory doesn't combine. Each rendered part-scene needs its own isolated buffer. So you should really consider it a 2x1gb card too It's just one of those cases where 2x1gb doesn't equal 2gb. Like 4 times 3ghz in a quadcore doesn't equal 12ghz

I didn't say the memory combined Do the gpu's on the card have their own discreet memory interface and own discreet memory pool or not? If they do, then quad core analogies are chocolate teapot useful Fair dues if i'm wrong but i thought that dual gpu cards are effectively the same deal as buying cards seperately but with the SLI/crossfire bridge built in.

[j.o.s.h.1408]

bah who cares about memory bandwidth, cuda this, cuda that, im interested in benchmarks on actual games, power consumptions and temps. the rest i do not care of