ATI Radeon HD 5890 Coming in Q1 2010

[faugusztin]

And regarding the original post, that's a table from Wikipedia. Someone made a few guesses about the next round of cards, and now people are assuming it's genuine information. I somehow doubt it.

Actually it got removed from Wikipedia because it is just a damn speculation/dream of someone.

 

[Owenb]

28nm, no chance. 320/384bit bus- possible as the 5000 series is certainly bandwidth starved.

 

[nitram100]

When's the 5870 released?

 

[gurusan]

When's the 5870 released?

? It's been out for awhile...it says right on the chart there.

 

[raralis]

5860 & 5790 look very intersting assuming the info is real.

Wonder if it's worth buying a 5850 now or waiting a few weeks.

To buy or not to buy...

 

[drunkenmaster]

I find the quoted picture as a pure imagination of someone with a great fantasy.

1) if Globalfoundries would have 28nm process done, they wouldn't do their CPUs in 45nm process and wouldn't bother with TMSC 40nm process. If they will have it by end of next year or in 2011, call yourself lucky. 40nm process is what we will have next year. http://www.pcgameshardware.com/aid,698957/AMD/Ati-28-nm-Radeon-chips-from-Globalfoundries/News/

2) 384 bit bus. Yes, right, and you know that one about ... . AMD and 384 bit bus with that core size is not going to happen.
3) The 5890 have less transistors, yet more shader cores ?

Sorry, but that whole picture looks like "Hm, make some nice numbers which will be quoted all around the world".

Those changes are way too big for a AMD half cycle (plus AMD half cycle is going to happen sometimes at end of Q2 2010/start of Q3 2010 - read june-august).

Firstly yes, its incredibly unlikely any of the info is accurate, its certainly possible those are the specs for the 28nm part/refresh but Q1 is incredibly unlikely.

However you need to realise several things about the rest of what you said.

CPU chips are made on a vastly more expensive SOI process, which is far harder to produce, takes far longer to change the process and redesign the core for a process shrink and doesn't tape out quickly. It takes a year, no less, for Intel or AMD to tape out a design and push it into mass production. Current 40nm gpu's, even if only one shader cluster and vastly reduced heat out put could not come close to 2Ghz, let alone 3-4Ghz, the processes aren't compatible, not remotely comparable, you could very easily be making a 28nm gpu and a 45nm CPU on different processes.

Anyway, 28nm bulk will be easier to produce and should come much earlier than 28nm SOI process, likewise the 32nm dates(which most likely will be skipped it seems to go directly to 28nm as 32nm is shaping up to be crap all around without HKMG, which again seems a waste at 32nm), are for the new fab being built, what GloFo decide to do with their other fabs is up to them.

384bus with a core that size? I have no idea what you mean, core size has entirely nothing to do with the amount of bandwidth it needs. Yes bus width increases pin out, traces required and complexity of the pcb, transistor count and a little more power, theres entirely no reason to not do it based on core size. The question is, does the extra cost warrant the performance, to fit onto a 5850 which should eventually be £150, no, to be on an ultra overclocked high end £300 card, very possibly, considering it would still be cheaper than a Fermi and could still make more profit while possibly being faster, why not.

Infact, if they ditch all gddr3/4 support in an upcoming chip gddr5 actually allows some very fancy memory timing, right now every memory chip has to have the exact same length trace from gpu to memory chip, meaning every chip has some ridiculous trace routing that add's layers, cost and complexity. GDDR5 allows the controller to stagger the signals so the traces can be different lengths, ie as short as physically possible. Meaning extra bus width doesn't add anywhere near the cost as it would have a year ago with older memory, meaning a much smaller bump in pcb cost, production time and even amount of metal used for traces.

That list also has the 5890 with 400million more transistors, not less, which would more than account for a minor shader bump and extra memory controller space.

however again, the list is just wrong, there will not be 28nm available right now, no fab in the world is doing real chips on 28nm in risk production let alone mass producing it, its simply not there or available.

Also the bump in shader numbers looks like the classic crap from last gen where someone convinced half the internet based on reading a picture of the 4870 core INCORRECTLY, that there were extra shaders built in for redundancy, that wasn't the case, it won't be the case here either. A new core could certainly bolt more on, and on a smaller process no doubt will. But AMD will not drop from 40nm to 28nm, a FULL node drop in 4 months and add 20% in shaders, that will never in a trillion years happen.

Not least because no manufacturer has dropped a full node before in anything under 2 years and seeing that TSMC is the least reliable manufacturer on the planet right now, just nonsense. Now keep in mind that in general both Nvidia/AMD have added anywhere from 80-100% more transistors on a HALF NODE drop in manufacturing process, a 28nm version of this 5890 would be, just ludicrously small, like, less than the size of a current 5770 probably. You'd be expecting a roughly doubling of transistors every half node drop on GPU's, yet this is what 20% more on a full node.

You'd expect on a similar architecture a 32nm "5870" style chip to have anything around 4billion transistors, and a 28nm version to have 6-8billion transistors, not 2.5billion.

 

[faugusztin]

Yes bus width increases pin out

This. With smaller core, you will simply not have enough place to route those pins without having interference etc.

That list also has the 5890 with 400million more transistors, not less, which would more than account for a minor shader bump and extra memory controller space.

Yep, i misread the number - 2154 vs 2150, not 2154 vs 2510 .

 

[Owenb]

Firstly yes, its incredibly unlikely any of the info is accurate, its certainly possible those are the specs for the 28nm part/refresh but Q1 is incredibly unlikely.

However you need to realise several things about the rest of what you said.

CPU chips are made on a vastly more expensive SOI process, which is far harder to produce, takes far longer to change the process and redesign the core for a process shrink and doesn't tape out quickly. It takes a year, no less, for Intel or AMD to tape out a design and push it into mass production. Current 40nm gpu's, even if only one shader cluster and vastly reduced heat out put could not come close to 2Ghz, let alone 3-4Ghz, the processes aren't compatible, not remotely comparable, you could very easily be making a 28nm gpu and a 45nm CPU on different processes.

AMD use SOI, Intel use bulk-infact nearly everyone uses bulk. We will see GPU logic on SOI when fusion arrives at 32nm, but AMD still isn't sure if it will use bulk or SOI after that.

 

[muon]

Well that table is obviously in wikipedia's style.

http://en.wikipedia.org/wiki/Compar...g_units#Radeon_Evergreen_.28HD_5xxx.29_series

If it was there before, its been reverted because the source is weak.

edit:

Going through the history, the original person provided no source. Hence other users have removed the speculation.

For things which have uncertainty, wikipedia is not where you want to be looking. However, users are pretty good at removing claims with weak or no source.

 

[psychas]

so many and 28 nm already

 

[muon]

so many and 28 nm already

It was made up. Someone edited the wikipedia article (now reverted). The actual article posted by OP mentions no specific details.

 

[gurusan]

WIKI WARRIORS UNITE!

 

[MoRT489]

how can it be unconfirmed if it's due to be released so soon and 28nm is just dreaming isnt it?

 

[faugusztin]

how can it be unconfirmed if it's due to be released so soon and 28nm is just dreaming isnt it?

It is not going to be released soon and right, 28nm in Q1 2010 is just dreaming.

 

[mmj_uk]

So it's just a rebranded 5870 then?

 

[CBJ]

It's not anything. It's just a piece of speculation that some people are reading too much into.

 

[drunkenmaster]

This. With smaller core, you will simply not have enough place to route those pins without having interference etc.


Yep, i misread the number - 2154 vs 2150, not 2154 vs 2510 .

You can make a bigger core, the extra bus would be a very small amount of transistors, the biggest increase was extra traces, more layered pcb and increased costs, you're talking about a 28nm core with 20% more shaders, you're looking at a core much less than half the size now, the extra bus would increase die size by 3-4%, that would be nothing compared to now on a 28nm process, but as I said, they simply would not do this right now a 28nm part would be twice as many transistors minimum otherwise it would be an entire waste.

The 5870 is actually bordering on too small with too small a bus width, the next gen will almost certain get a bump in bus width and the reduced tracing issues with gddr5 radically decreases the penalities of a larger bus. The extra pin out for memory connections really wouldn't introduce any extra interference in the slighest, you're not talking about huge high speed shaders doing complex work, its basic added core logic.

AMD use SOI, Intel use bulk-infact nearly everyone uses bulk. We will see GPU logic on SOI when fusion arrives at 32nm, but AMD still isn't sure if it will use bulk or SOI after that.

Intel are using HKMG, its in no way the same as current GPU bulk production and I actually can't remember but I believe Intel are adding SOI in the future, or well they are adding multiple things over the generations, they could not make an I7 above 800Mhz on current 40nm gpu bulk production, its not designed for it.

Likewise the first fusion gpu's will be much smaller than current top end, won't be running at the same speeds and will still work on SOI, it will cost more to make the same core on SOI than bulk though, remember, SOI is better all ways, it won't make a 5870 capable of 3Ghz + but it will decrease one form of leakage significantly and one of the biggest problems on TSMC's 40nm process(there are multiple types of leakage, the 40nm tsmc suffers from gate leakage IIRC which is the main thing SOI works to combat.

The problem with making gpu's in a fab that does bulk, rather than intergrated on die with a cpu that can only be made on SOI are greater than making an all in one die with the GPU done at extra cost on SOI though.

 

[theforce]

Can't see ATI adding more shaders to the 5890. It will probably be a 5870 with higher clocks and if we are lucky a 384-bit memory interface. I guess ATI will have an estimation of how much of a performance increase a 384-memory interface will provide.