AMD: Volcanic Islands R1100/1200 (8***/9*** series) Speculation/ Rumour Thread

I will hope for something new from AMD, but well look like it will not happend before 2015

 

IIRC most rumors have R300 at H1 2015 though the question should be is "what will it be like", the current architecture just cannot compete against Maxwell and they are probably having very low margins (if at all) with their price drops. Just scaling up Tonga probably will be ineffectual (what was the exact point of Tonga again? seems like a waste of R&D).

 

- I think it's unrealistic that they don't have decent gross margins (let alone no margins.) This is such a mature process. That fact that prices were still this high was more or an anomaly. - Tonga is a cost reduction that improves gross margins compared to Tahiti due to lower DRAM requirements. Its architectural improvements (especially the compression) will help higher segments as well. Not entirely useless, but compared to gm204 definitely insufficient, of course. But with Fiji (or whatever it's called) still a bit away, you'd think that they made other improvements as well.

BTW: don't forget that Richard "a buffer adds a full frame of latency" The Scientist Huddy said that AMD did not reduce its prices in reaction to Maxwell. Hard to believe, but if true, it's the board makers who are eating the margins.

 

The " current " architecture ? no seriously, Nvidia gpu cores. after Fermi are looking more and more like GCN .. Nvidia dont realize their magic out of the hat... The next big thing in GPU's is not the GM204,, it will beGM210-200 vs AMD .. Yes Maxwell improve in many segment and specially power consumption, efficiency of the SMM , but this is a real limited chip in term of computing. it offer a lot of efficiency ( for games ) but it still under used yet in this actual configuration. it is just a damn low 5Tflops SP configuration, with a 1:32DP rate configuration.. .. What i want to see, is how this is working in a Tesla M20 configuration ( and we could be suprised in good ) ..

Nvidia need something ASAP for the professional market, because there, they cant compet with AMD Hawaii based gpu, they are overshadowed completely.. so i can imagine( no , i know, it is coming, i see a lot of agitation and heard some rumor sound in the corridors ) we could see something untill the end of the year ( hoping it will not be like the first Kepler based Tesla who was a shame GK204 + a dual GK104 )

 

I've seen this claim elsewhere, but I don't understand it: how is the Maxwell architecture closer to GCN than Kepler?

 

The recent deal with Apple should increase AMD's professional market share to 30% from 20% by end of 2014.

http://finance.yahoo.com/news/amd-professional-graphics-attain-milestone-193408277.html

 

Impressive numbers on the surface....then you look at the fact that you can upgrade from dual D500's to dual D700's for $600 and it tends to show things in a different light. Bearing in mind the pricing of their FirePro analogues, it would seem that AMD are all but buying this market share.

 

I will not use MacPro as a marketshare signal in the professional segement anyway

 

I don't think that it's closer to GCN than it is to Kepler, but its partitioning of that couples register files, schedulers, and SIMDs instead of having multiple schedulers managing multiple SIMDs drawing from a common register file is closer to the CU setup. Splitting local memory out into its own distinct pool is another change.
However, going back a number of years, the introduction of LDS aligned AMD's GPU architectures more closely to the CUDA/OpenCL model back then.

The simplified operand gathering model is still not quite the same as what GCN does, although I also see a higher-level similarity in the increased mention of wait counts, which GCN also has for various reasons.
I think it's less an AMD/Nvidia thing than the designs operating in similar terrain when it comes to local design optima.

For me, the compiler-exposed register reuse values evokes memories of AMD's VLIW GPUs, however.

 

If that really is true "AMD having its board makers eat the price reductions" then AMD in the future will have less board makers.

 

It's true that he said AMD didn't cut their prices. Whether or not he was telling the truth is a different story.

 

Afaik, AMD didn't raise prices when the frenzy was on for cards that could mine. So, one can argue that maybe there is a larger picture.

 

How so?
Mining card price hikes were instituted by (r)etail and wholesale sellers were they not?

Even if you're arguing that AIB's increased prices (which I doubt since prices didn't escalate in markets where mining didn't become fashionable), I doubt that any company would look at current belt-tightening and come to the conclusion that they're happy with it because of profit made many months ago.

 

Hawaii seems to be competing right with GM204...

If they release something 20-30% faster than GM204 at a 250w TDP, that would put it in the same neighborhood in regards to performance/watt.

 

I don't think anyone expects its new high-end to go against gm204.

And if gm200 perf is limited by 250W, with the same perf/W as gm204, it will be 35% faster.

 

My memory is that it was posted that AMD didn't raise prices on chips that were suddenly so popular with miners.* If they didn't raise prices then, and they aren't lowering prices now, I think that maybe there are other considerations beyond their relative worth in the retail market which is driven by immediate demand.

*I remember prices rising on the boards. lol, I was tickled at my HD 7950, no spring chicken, suddenly worth not that much less than what I paid for it.

 

Like you have said this was a wholesale retailers price increase, and so they have increase their margin with it.... Im not sure AIB or AMD have do any profit from it for be honest outside sell cards. I have 2x 5870 and 2x 7970 here, if i had wanted to sold them at this period,
this will have make me a good margin price in Ebay... I will have sold them way higher that the price i have paid them initially. But at this period i was more in overclocking, and thoses 2 5870 are pure gem for it.. ( golden samples i cant say more they support everything.. )

 

GM204 has almost twice the performance per watt of Hawaii as it is. So unless they manage to do something similar to what NV did with Maxwell, I dont see that happening.

If GM200's TDP is 250W, and we scale performance exactly to power, it should have almost 50% more performance than GM204 with its 165W TDP.

However, I think GM200's performance will be limited more by the die size than power. I do not think it can be anything more than a 24 SMM part.

 

I have no clue why NVIDIA came up with that marketing-TDP for GTX 980, but it's real TDP is 180W

 

I'm using the GTX980 TDP of 185W. (Turns out it's 180W.)

Agreed. And it may be less because of DP FP. But even with 24, it's 50% more than gm204 and that goes past the 250W mark. Not that I would mind going past that. Going on the cooler that AMD has been preparing, they're going to do that also...