[Various]Radeon Fury X and Radeon Fury coming

[Azix]

http://www.amd.com/Documents/High-Bandwidth-Memory-HBM.pdf

I just noticed in the footnotes that the bandwidth is referenced to an HBM-1 device, but the size graphic is for an HBM-2 stack.

Is HBM-1 actually bigger than HBM-2?

Also the wattage comparison is to a 290X, so HBM-1 will be a little worse next to NV's current crop of cards as far as BW per watt.

Doesn't nvidia use more watts for their memory because of the higher clock?

 

[sm625]

My dream scenario that is not likely impossible.

AMD Fury X is 60% and AMD Fury is 30% faster than Titan X

AMD Fury X price point is $750

AMD Fury price point is $600

If this sound like trolling, it really isn't.

That is what I expect as well. AMD needs to beat by those margins just to combat the marketing hype. There is a very good chance that the nvidia cards will still outsell, even if AMD hits those numbers.

 

[exar333]

Doesn't nvidia use more watts for their memory because of the higher clock?

I believe the most recent info has alluded to about 20-25w of power savings using HBM. That's not bad at all...that climbs as you include more memory.

 

[SimianR]

My dream scenario that is not likely impossible.

AMD Fury X is 60% and AMD Fury is 30% faster than Titan X

AMD Fury X price point is $750

AMD Fury price point is $600

If this sound like trolling, it really isn't. There should be no reason why the GTX 980ti should be more than $500 just because the Titan X is an overprice piece of garbage.

As much as I'm sure lots of people would like to see this, I have a feeling the Watercooled variant of Fury might be ahead of the Titan X 10-15% and the air cooled version will be trading blows. I'm curious as to where the cut down version of the Fury will land (if it exists). Between the 980 and the 980Ti?

 

[IEC]

As much as I'm sure lots of people would like to see this, I have a feeling the Watercooled variant of Fury might be ahead of the Titan X 10-15% and the air cooled version will be trading blows. I'm curious as to where the cut down version of the Fury will land (if it exists). Between the 980 and the 980Ti?

Would be great for gamers if "Fury X" is closer to thehotsung8701A's numbers AND comes in at a closer to $650 price point.

If it crushes Titan X it might actually be worth considering, even as a late 28nm part.

 

[exar333]

As much as I'm sure lots of people would like to see this, I have a feeling the Watercooled variant of Fury might be ahead of the Titan X 10-15% and the air cooled version will be trading blows. I'm curious as to where the cut down version of the Fury will land (if it exists). Between the 980 and the 980Ti?

Honestly, while I am not agreeing or disagreeing with what you say, the benchmark is the 980Ti not the TX anymore. A week ago, yes. Today, not so much.

 

[Gikaseixas]

Honestly, while I am not agreeing or disagreeing with what you say, the benchmark is the 980Ti not the TX anymore. A week ago, yes. Today, not so much.

If it's faster than Titan X then it makes it a Titan X rival (too)

 

[AnandThenMan]

My dream scenario that is not likely impossible.

AMD Fury X is 60% and AMD Fury is 30% faster than Titan X

Don't think this is physically possible the transistor budget for this kind of performance would be impossible to manufacture on 28nm.

 

[tential]

Don't think this is physically possible the transistor budget for this kind of performance would be impossible to manufacture on 28nm.

It's plain and simple not happening. No way the fury x has 60% + perf on Titan x and is cheaper. We all know that lol

 

[raghu78]

My dream scenario that is not likely impossible.

AMD Fury X is 60% and AMD Fury is 30% faster than Titan X

AMD Fury X price point is $750

AMD Fury price point is $600

If this sound like trolling, it really isn't. There should be no reason why the GTX 980ti should be more than $500 just because the Titan X is an overprice piece of garbage.

unrealistic and will not happen at 28nm. Best case is Fury X is 20-25% faster than Titan X due to significant microarchitectural improvements to improve perf/sp, perf/sq mm and perf/watt. Coupled with HBM thats probably the best we can expect. Realistically even 15% faster than Titan-X means only at 1400+ Mhz Titan-X can catch up with a stock Fury X. So I think even if AMD achieves that at stock with another 15-20% perf through OC headroom they are going to be in a good position. :thumbsup:

 

[soccerballtux]

I believe the most recent info has alluded to about 20-25w of power savings using HBM. That's not bad at all...that climbs as you include more memory.

those dinky little GDDR5 chips do NOT use more than 1W/piece

maybe the savings is on the die itself

 

[Tuna-Fish]

those dinky little GDDR5 chips do NOT use more than 1W/piece

maybe the savings is on the die itself

The savings is on the interface. Almost all of the power use of high-speed ram is rapidly loading and discharging the long wires between the controller and ram. That's why tech like bus invert signal is used to reduce the amount of wires that need to switch.

 

[IEC]

those dinky little GDDR5 chips do NOT use more than 1W/piece

maybe the savings is on the die itself

AnandTech covered this in detail already:
http://anandtech.com/show/9266/amd-hbm-deep-dive/4

Could be anywhere from a 30%-70% savings in power over GDDR5 depending on clockspeeds, etc. Power budget that is then rolled into the core.

Also keep in mind that AIO coolers will further reduce power consumption over most air coolers by a measurable 10W-30W depending on how bad the air cooler is. By keeping temps lower less power is used. Or, as is often the case, higher clockspeeds can be achieved at the same power limit.

 

[exar333]

AnandTech covered this in detail already:
http://anandtech.com/show/9266/amd-hbm-deep-dive/4

Could be anywhere from a 30%-70% savings in power over GDDR5 depending on clockspeeds, etc. Power budget that is then rolled into the core.

Also keep in mind that AIO coolers will further reduce power consumption over most air coolers by a measurable 10W-30W depending on how bad the air cooler is. By keeping temps lower less power is used. Or, as is often the case, higher clockspeeds can be achieved at the same power limit.

Exactly. Thanks for jogging my memory, that's where I saw that...

Of course AMD is then investing some of those gains back in to coming up with more memory bandwidth, so it’s not as simple as saying that memory power consumption has been cut by 70%. Rather given our earlier bandwidth estimate of 512GB/sec of memory bandwidth for a 4 stack configuration, we would be looking at about 15W of power consumption for a 512GB/sec HBM solution, versus 30W+ for a 320GB/sec GDDR5 solution. The end result then points to DRAM power consumption being closer to halved, with AMD saving 15-20W of power.

I would imagine that is higher and would be closer to 25w or maybe even 30w if you are talking about 8GB HBM (if that is even released)

Some people just like to argue everything, with no justification....

 

[LTC8K6]

Well, if the pictures were accurate, Fiji apparently still needs more power than 980ti.
Two 8-pin connectors versus the Ti's one 6-pin and one 8-pin.

 

[exar333]

Well, if the pictures were accurate, Fiji apparently still needs more power than 980ti.
Two 8-pin connectors versus the Ti's one 6-pin and one 8-pin.

Good point. Assuming this is true, that probably is telling on why AMD wanted gen1 HBM. They needed the power savings to allocate to compute...

NV had power consumption a little more in control, and could wait it out until gen2.

 

[beginner99]

Don't think this is physically possible the transistor budget for this kind of performance would be impossible to manufacture on 28nm.

I don't think it will happen but compared to Hawaii AMD can save die space on memory controller because HBM, can make a bigger die, can make a denser design (if made at GF 28 nm) and can cut out DP support just like GM200. All of this gives die space and lower TDP. I'm not saying it's happening and I very much doubt it, but it doesn't seem that impossible.

 

[Despoiler]

Well, if the pictures were accurate, Fiji apparently still needs more power than 980ti.
Two 8-pin connectors versus the Ti's one 6-pin and one 8-pin.

Maybe, maybe not. I read that AIBs won't be able to customize the Fury X card (or whatever the top water cooled card is). If that's the case a 2x8 pin would make sense to give overclockers room to work with.

 

[cbrunny]

Well, if the pictures were accurate, Fiji apparently still needs more power than 980ti.
Two 8-pin connectors versus the Ti's one 6-pin and one 8-pin.

Fiji (from the pics leaked) also has a water pump and probably a 120mm fan. NO idea how much power that takes though. I assume it isn't marginal though.

 

[Despoiler]

Fiji (from the pics leaked) also has a water pump and probably a 120mm fan. NO idea how much power that takes though. I assume it isn't marginal though.

It's marginal.

 

[AnandThenMan]

I don't think it will happen but compared to Hawaii AMD can save die space on memory controller because HBM, can make a bigger die, can make a denser design (if made at GF 28 nm) and can cut out DP support just like GM200. All of this gives die space and lower TDP. I'm not saying it's happening and I very much doubt it, but it doesn't seem that impossible.

I think 25% is not out of the realm of possibility. Anything more to me seems too good to be true.

 

[LTC8K6]

Maybe, maybe not. I read that AIBs won't be able to customize the Fury X card (or whatever the top water cooled card is). If that's the case a 2x8 pin would make sense to give overclockers room to work with.

If Fury uses a decent amount less power than the 980Ti, then a 6/8pin setup should give Fury plenty of room.

980ti is listed at 250W but we have 75W slot + 75W 6-pin + 150W 8-pin or 300W total available to the Ti card.

So, Ti theoretically already has some power overhead.

 

[xthetenth]

I can see pushing an overclock to the limit with the extra room given by water taking it close to 300w. May as well give the big OC crowd room.

 

[Face2Face]

I wouldn't use the ATX12V Spec as concrete power consumption guide/measurement. Just look at the R9 295X2, it only had two 8-pin PEG connectors and still pulled close to 500w alone under load. Also, I have two 8-pin PEG connectors on my Lightning, and I have also pulled around 500w from my card with 1.6v applied to core.

 

[LTC8K6]

I can see pushing an overclock to the limit with the extra room given by water taking it close to 300w. May as well give the big OC crowd room.

Well yeah, it's either 300W available, or you jump to 375W available.

You either add 225W or 300W from the PS to the 75W from the slot.

Will overclockers be adding voltage to the HBM?

Do we know much about voltage tolerances of HBM?