FuryX completely abandoned now, barely matching the 1060 or 580!

[HenrySomeone]

If cost was the main obstacle, then you can bet Nvidia would put it on their prohibitively expensive 2080Ti not to mention Titan RTX, yet they only use it on TItan V an one or two Quadro models

 

[xkm1948]

That may be but aren't all new things going to have break in periods? Also, Crypto mining was decent on the FuryX given the initial cost and the VRAM wasn't a limitation early on.

@ HenrySomeone They still use it. Instinct cards use it and will continue to use it.

I don't know about how early was early. By the time I started going with it there were some sort of difficulty rise that makes a must to have 8GB VRAM to mine.

 

[Keullo-e]

If cost was the main obstacle, then you can bet Nvidia would put it on their prohibitively expensive 2080Ti not to mention Titan RTX, yet they only use it on TItan V an one or two Quadro models

I guess that they're simply not bandwith limited so they do fine with GDDR and a 352/384-bit bus

 

[HenrySomeone]

That's the point - HBM may be "the king of bandwith", but with gddr5x before and now gddr6 you still get enough of it for most use cases with much less hassle and cost...

 

[Keullo-e]

That's the point - HBM may be "the king of bandwith", but with gddr5x before and now gddr6 you still get enough of it for most use cases with much less hassle and cost...

I'd see the best thing in HBM to have cards with a smaller PCB, just like R9 Nano. But then, AIB's totally missed this possibility when they made a "normal size" Fury and Vega cards instead of their possible "Nano" cards.

Hell, even GTX 1080 and 1080 Ti had small versions:

GeForce® GTX 1080 Mini ITX 8G Key Features | Graphics Card - GIGABYTE Global

Discover AORUS premium graphics cards, ft. WINDFORCE cooling, RGB lighting, PCB protection, and VR friendly features for the best gaming and VR experience!

www.gigabyte.com

ZOTAC GeForce® GTX 1080 Ti Mini | ZOTAC

<p>NVIDIA’s new flagship GeForce® GTX 1080 Ti is the most advanced gaming GPU on the planet, driven by the high-performance, power-efficient new NVIDIA Pascal™ architecture. All that performance is now contained in a compact size, enabling top performance even in smaller builds.</p> <p>- The...

www.zotac.com

 

[HenrySomeone]

Well, considering which gpus they were pairing HBM with (i.e. "gas guzzlers"), small cards were only really possible by slapping watercooling on it or they throttled like crazy like the Nano...

 

[biffzinker]

I guess that they're simply not bandwith limited so they do fine with GDDR and a 352/384-bit bus

That would be Nvidia's Delta Color Compression at work. AMD has the feature as well but the implementation wasn't as efficient as Nivida's. I think it was Vega but it might be Navi that AMD had the efficiency matching Nvidia.

 

[Keullo-e]

Yeah, I remember something about that when Maxwell came out, that a 256-bit bus on GTX 980 for example wasn't a bottleneck.

 

[TheoneandonlyMrK]

If cost was the main obstacle, then you can bet Nvidia would put it on their prohibitively expensive 2080Ti not to mention Titan RTX, yet they only use it on TItan V an one or two Quadro models

Nvidia took one memory chip off a 2080ti to make the 11Gb 2080 ,just to save a few quid and it still cost /700£ and you think they could have gone Hbm.
They could sure but Huang isn't passing up a chance to slap consumer asses is he, no.

 

[Keullo-e]

Nvidia took one memory chip off a 2080ti to make the 11Gb 2080

What? I guess you mean that they knocked one memory chip off from Titan to make it a 2080 Ti, just like they did with 1080 Ti.

 

[biffzinker]

but with gddr5x before and now gddr6 you still get enough of it for most use cases with much less hassle and cost...

You left out the additional PCB spaced occupied for GDDR5/GDDR6 unlike HBM on the same package as the GPU die.

 

[Keullo-e]

You left out the additional PCB spaced occupied for GDDR5/GDDR6 unlike HBM on the same package as the GPU die.

Exactly, that's what I meant that they could make a R9 Nano type of a small powerhorse with HBM.

 

[biffzinker]

Exactly, that's what I meant that they could make a R9 Nano type of a small powerhorse with HBM.

Except the AIBs want or need to slap on a 3 slot heatsink so the PCB might as well be humongous.

 

[The Egg]

Nvidia took one memory chip off a 2080ti to make the 11Gb 2080 ,just to save a few quid and it still cost /700£ and you think they could have gone Hbm.
They could sure but Huang isn't passing up a chance to slap consumer asses is he, no.

The memory controller on the chip has to be built to use HBM. The Titan RTX used GDDR6 as well, so there's no way a derivative is going to use HBM.

 

[Keullo-e]

Except the AIBs want or need to slap on a 3 slot heatsink so the PCB might as well be humongous.

Or the heatsink is twice the size of the PCB in length.. About just what Sapphire's R9 Fury was:

 

[kings]

Now I remembered when AMD said 4GB HBM is better than 8GB or even 12GB GDDR5.

And most curiously, many have defended this.

 

[biffzinker]

Now I remembered when AMD said 4GB HBM is better than 8GB or even 12GB GDDR5.

Well then, I wasn't aware but I'm also not surprised.

 

[Vya Domus]

Anyway, this is pretty neat troll thread. The OP, a well known avid fanboy which I had the pleasure to argue with on several occasions is nowhere to be seen.

I must say the quality of most threads on here as of late really went downhill.

 

[Vario]

Anyway, this is pretty neat troll thread. The OP, a well known avid fanboy which I had the pleasure to argue with on several occasions is nowhere to be seen.

I must say the quality of most threads on here as of late really went downhill.

Yep dump a controversial topic, sit back, and watch fanboys argue over it.

 

[biffzinker]

Isn't the 980ti still a monster?

At this point I would be calling the 1080 Ti the monster. The 980 Ti had that moment but it's long past now.

 

[Solaris17]

At this point I would be calling the 1080 Ti the monster. The 980 Ti had that moment but it's long past now.

Thanks I knew it held its own for quite some time but didn't know it had diminished so much.

 

[Fluffmeister]

The 980 Ti was always the better card, it's reference 1000Mhz base clock always makes it appear slower in charts too, when it could easily gain 25-30% over that.

 

[Keullo-e]

Even 980 can do still fine, I had one about an year ago and ran it with stable 1500MHz.

 

[Vayra86]

That's the point - HBM may be "the king of bandwith", but with gddr5x before and now gddr6 you still get enough of it for most use cases with much less hassle and cost...

That was the main argument surrounding HBM at the time. A year prior to Fury's launch every nerd yelled their next card 'had to have HBM' and Nvidia was 'cheaping out with shitty GDDR and bandwidth limited cards'... and then 980ti came out and smoked the HBM offering using GDDR5 - and even kept up at 4K most of the time.

The other hope and dream for Fury X was its 'tremendous overclocking potential' as referenced on an AMD slide. And then Fury X came out.

The reality is, Fury X was bare necessity because AMD didn't get the brilliant idea of delta compression yet, nor did they ever get their boost technology in order. Nvidia rode Maxwell doing that alone and made do with old tech - and started running away from AMD's performance cap. Meanwhile, AMD got stuck with HBM-equipped technology trying to push ancient GCN forward that really wasn't able to cope anymore regardless - HBM was AMD"s good old 'moar cores/hardware' approach to fix an architecture lacking efficiency - and was even not balanced quite right for Fiji either.

And note: they still haven't fixed it really. They're now using new technology (again) to keep it viable. This time its called 7nm.

 

[Valantar]

That was the main argument surrounding HBM at the time. A year prior to Fury's launch every nerd yelled their next card 'had to have HBM' and Nvidia was 'cheaping out with shitty GDDR and bandwidth limited cards'... and then 980ti came out and smoked the HBM offering using GDDR5 - and even kept up at 4K most of the time.

The other hope and dream for Fury X was its 'tremendous overclocking potential' as referenced on an AMD slide. And then Fury X came out.

The reality is, Fury X was bare necessity because AMD didn't get the brilliant idea of delta compression yet, nor did they ever get their boost technology in order. Nvidia rode Maxwell doing that alone and made do with old tech - and started running away from AMD's performance cap. Meanwhile, AMD got stuck with HBM-equipped technology trying to push ancient GCN forward that really wasn't able to cope anymore regardless - HBM was AMD"s good old 'moar cores/hardware' approach to fix an architecture lacking efficiency - and was even not balanced quite right for Fiji either.

And note: they still haven't fixed it really. They're now using new technology (again) to keep it viable. This time its called 7nm.

You're missing a key point here: Nvidia's advantage with Maxwell (and what has kept them ahead since) isn't just DCC, but a generally better balanced architecture that allows for (far) better utilization of available resources. That's also why they're so damn efficient. AMD is finally gaining on Nvidia in this regard - with Navi more or less matching Turing in perf/W - but the process node advantage AMD has definitely shows that they're still quite a bit behind. Still, HBM doesn't matter much in this regard - while GCN has indeed pretty much always been bandwidth starved, the general lack of utilization of resources is a far bigger issue as performance increases. That's why a 40-CU Navi card (admittedly at higher clocks, but still) can beat a 64-CU Vega card even though it has slightly lower memory bandwidth. AMD's VRAM compression is also improving, but that's just one part of the puzzle. Still, unless you're specifically talking about the Radeon VII and ignoring the RX 5700 series, your 'moar coars' analogy misses the mark. The architectural efficiency improvements of RDNA are definitely not insignificant.

It is definitely worth saying that GDDR6 (and in part 5X before that) has made HBM's main advantage a lot smaller. And with cost being roughly double, the other advantages (density/board space and power) are tough to swallow. Personally I would much rather have a compact PCB where more of the power budget is given to the GPU (a normal 8GB GRRD5 or GDDR6 setup needs ~40W of power, HBM halves that, and 20W more GPU power can make a noticeable difference), but outside of premium segments (>$700 or so, I'd say), the extra $70-80 of going HBM (not including the interposer cost, IIRC) would indeed be a tough pill to swallow for most people.