Topic: Speculation about a big AMD GPU (or HBM ram) on the new process (490x etc) - page 2. (Read 4062 times)

 NVidia targeted higher-end cards at the start, which sell in smaller quantites - and I AM still seeing some shortage issues despite that and despite NVidia having had the 1080 and the 1070 on the market a fair bit longer than the RX 480 has been available.
 Try finding that Gigabyte "ITX" 1070 card in stock anywhere for example - it's been VERY VERY spotty so far (and I have a specific use for 2 or 3 of them over the next week or 2).

 AMD isn't so much "behind" NVidia, as they started out targeting the "mainstream" market - and have probably sold quite a few more RX 480 units than NVidia has sold on all GTX 1xxx + Titan X Pascal units COMBINED, though the total on DOLLAR sales is going to be a lot closer.

 The real tale will be next year, when AMD finally releases cards INTENDED to go head-to-head against the current NVidia offerings.

Interesting. It's clear that it's easier to stack two gpus side by side than to develop a new bigger one. In raw power and in gaming amd is really behind nvidia now. A part Ethash I don't see them performing very well.

I found this a while ago that speculated that they would launch a multi-chip card this year in order to compete at the high end.

http://wccftech.com/amd-rx-490-4k-gaming-card/

For what I see there is no shortage on the nvidia side. My argument for a bigger amd before end of year is that the rx480 isn't that powerful compared to nvidia offerings. For end of year sell it would be quite disappointing that on amd side the more powerful gpu be from the previous generations. let's wait and see.

There is some variance on yields between the manufacturers - and I suspect some between different fabs at the SAME manfacturer - but they've all been reporting major gains on yield this year, it's not that much of an issue any more.

 The REAL limit is flat out total capasity limitations.


 From what I've been seeing around the web, it doesn't look likely that there will be any more major "new cards" announced for the rest of the year - and even if any WERE announced there would be major "availability" issues for a while.

 Keep in mind we're still seeing major issues with availability on all of the CURRENT new-gen cards.

Okay thanks, I didn't knew that the yield was the same for both. I haven't read on the subject recently.

So wait&see until the bottleneck get loser and more chips are produced.

Do you think that there is still a possibility for this "big" amd gpu for the end of this year?

The only major difference between 14nm and 16nm is the manufacturer - in practice, they're pretty much identical for efficiency and in yield.

 The main issue is that there are only a very few founderies so far that can make the stuff at all - and there's a TON of demand for it from folks that make VERY HIGH VOLUME stuff like smartphones, not just GPUs and CPUs and ASIC for mining (though GPU/CPU sales from folks like Intel/AMD/NVidia/IBM/etc aren't exactly LOW volume).

 Also keep in mind that a lot of the output of Global Founderies in particular is probably "locked in" to AMD and IBM due to the history and contracts that resulted in Global Founderies existance and current fab ownerships. To a real degree, there's only what, 4 companies that actually own fabs any more - TSMC, Global Founderies, Intel, and maybe Samsung? Might be a couple others I'm forgetting, and perhaps some SMALL foundery owners I'm not aware of, but when it comes to CURRENT process it's a very short list.



 So yes, there IS a bottleneck on current process, vs a rather high "pent up demand" for chips from current process....

Thank you very much for your post. it's very informative. And then come the optimization of the mining software. It can too make a huge difference... specially if they are kept private.


That would be expected... there is already a problem with the supply of the current cards. Is there a problem with the process? 14nm is too ambitious and not yielding enough yet? Maybe Nvidia was right to go with a 16nm...

http://www.fudzilla.com/news/graphics/41467-leaked-slides-warn-of-vega-delays

We may have to wait a bit longer.

 For Ethereum, very much so - even counting core units AND clock rate has little to no relation to hashrate, as Ethereum is very reliant on the MEMORY SYSTEM on the card.


 There are many other usages where Number of Cores x Clock Rate *IS* a good indicator of performance, but even there differences in card design can have a major effect - just not usually the same way as on Ethereum.

 Folding@Home for example is dominated by NVidia cards despite AMD having more "cores" and competative clock speeds on "comparable" cards, as the design OF the "cores" in the NVidea cards works quite a bit better for the type of work F@H does.
 distributed.net RC-5 and OGR work favors AMD cards very heavily, as for those projects the AMD "cores" are pretty much identical efficiency per core/clock vs NVidia but "comparable" AMD cards normally have a LOT more "cores" at comparable clock. This also applied to SHA256, Scrypt, and X11 mining back in the days that those were viable to mine with GPUs and applies to at least some degree to many other current "GPU mineable" algorythms.




 As far as the current Polaris vs. Pascal battle - keep in mind that AMD *STARTED* with a card intended for "mainstream" usage, NVidia "STARTED" with a card intended for VERY HIGH END usage - the expansions of both lines are only just starting to overlap, and even with that the RX 480 is still apparently aimed at a slightly LOWER performance point than the GTX 1060, though the announced GTX 1060 is FINALLY aiming into the same price point range as the RX 470/480 are at.
 It's still early days, way too early to predict how this battle is going to turn out since neither manufacturer has put out anywhere near a "full range" of cards in the current generation YET.

It's a complete nonsense comparing different architecture just counting the number of the core units.
It's also impossible to make speculations about thr coming chip from amd, dur the presence of a new kind of RAM, chip and driver implementation.

For now, it's just clear that the polaris chip it's underperforming with the majority of algos, in opposition to pascal, considering the same power consumption. The only exception it's, thankfully to the owners, dagger hashimoto, but JUST with a modded bios, that could in theory make the vga a lot more unstable with other algos.

AMD will need to raise the efficiency of his units, and it will, but in the meantime release a chip that will be able to gain at least 80% more than the polaris XT; if history repeat itself, we could have a new 7970-like vga, well balanced, with potential but stretched to gain framerates, with a tdp up to 250W. With the new HBM 2.0 I wouldn't expect more than 220, but also not more than 50% thsn the polaris XT.

The main issue for amd will be the 1080 Ti, that will force vega to compete with a, wonderful, 1080. And that's an optimistic scenario, because vega could also compete with an overclocked 1070

 The real comparison should be to the R9 290 though, not the X, as the non-X mines at pretty much IDENTICAL hashrate to the non-X.

 The duo cards have higher hashrate due to having DOUBLE THE MEMORY SYSTEMS, not due to higher core count - the Fury hashes SLOWER than the R9 290 despite having ALMOST DOUBLE THE CORE COUNT AND MUCH HIGHER CLOCK SPEED.

 For many cryptocoin algorythms, the core count is critical - but Ethereum's algorythm seems to rely a lot more on ability to access memory efficiently, higher core counts past a certain point on the same archetecture DOES NOT HELP or the R9 290X would be a noticeably faster miner card than the R9 290 AND THE FURY WOULD BLOW BOTH OUT OF THE WATER, which is not even close to the truth.

 Yes, the Fury is more EFFICIENT - but it does NOT have a higher hashrate.


 Also, there is no reason for hashrate to "decrease for any gpu 74 degrees celsius or above" - trying to limit a R9 290 to 74C is almost impossible if you want to maximise hashrate as you have to UNDERCLOCK the cards too much to get there.
 Some cards are DESIGNED to have to run hot due to limits of their process tech at the time they were designed.
 Keeping the cards cooler for better longevity, SURE, that's a good idea - but the idea that there is some specific mythical "maximum temperature for best efficiency for ALL GPUs" is totally bogus BS.




 STOCK clock on the R9 290 varied - many early ones had a STOCK clock at 947 MH, not 1GH (and tended to overheat even THEN in mining usage unless you kicked the fans up to 100% or modded the BIOS) - and while I suspect some later models with high-end cooling might hit 1200 MH reliably, none of mine have ever managed more than 1120 with any stability (but they're Reference design with the early blower-model coolers) and they tend to like 1100 max for long-term stability.

I had  very good profits with fury x on sia before pool mining. A fury x reach 1800-1900 mh/s, which more than a R9 390X (1100 mh/s), GTX 980 ti ( 1600-1750) and just a little bit under the GTX 1080.

HBM2 entered production Q2 2016, so it might become available for AMD and NVidia shortly. and zcash mining is not quite clear yet.

This is a very goog point. Algo versatility is very important to me. Who knows what this one will require. I have no idea. Anyone of you know something?

Yeah about HBM the true advantage is the huge bus. Isn't HBM2 already available? 

I have had experience with the 290x, 295x2, 390x, fury x and pro duo. There are a few things to consider... First, the hashrate will decrease for any gpu 74 degrees celsius or above. Second, for a non-commercial setup the easiest way to overclock the cards is with msi afterburner with power set to +50 up to 1137-42 leaving the voltage alone with for a 4 way crossfire x a power supply that does 1300+ continuous watts. Third, use the most current mining software for example "claymore's dual eth and decreed" for more efficient mining and increase in profit. Fourth, it helps to have an eth proxy server if you have 2 or more nodes. For example we have a setup with 1 295x2 ( 2 290x ) and 2 fury x all overclocked and it does 133 non-calculated mh for ether whereas another one of our setups with either 2 295x2's or 4 290x's were getting around 116-118 MH. Our rig with 2 pro duos (2x2 fury x) 147 mh a considerable difference. A tip is to leave your cases open with secondary household fans venting the heat from the already 100% liquid cooled system ( can be done with closed loop cpu, gpu ).

And what about deifferent anlgo? when rx490 will be released there will be z-cash which can be more profitable tha eth..

Do you have any experience with the card? my card won't go anywhere above 1125Mhz. while the stock clock is 1050Mhz. A 290's stock is 1GHz and it might clock to 1.2GHz on good samples. but the Fury X does beat the 290 and even the 390X in most of the games, but not in mining.

I would have to say I have a difference of opinion of the hashrate of the 290x beating the fury x. First, the fury x is more power efficient, stays cooler, and has 4000 cores at the same speed as the 290x with 2800 cores. The fury x also has higher overclocking ability. The fury x definitely outperforms the 290x but its not due to the HBM. A pro duo has 2 fury x gpus with 8000 total cores where as its predecessor the 295x2 has 5600 cores with 2 290x GPU's so the core difference is a noticeable increase in speed in terms of GPU mining especially with the newest mining software.

HBM is a tradeoff.

 The WIDE bus in theory can give better performance - but so far the clock rates on the actual chips are pretty low, making it little if any better than current GDDR5 implimentations.


 5 years down the road might be another story - HBM tech is quite new after all.