Upcoming Geforce GTX Volta cards Use GDDR5X not HBM2

[P4-630]

"We've written a lot back and forth about GDDR5X versus HBM2 (and GDDR6 these days). Nvidia did not place all their eggs in one basket, AMD pretty much did so for high-end. That being the main and root cause of delays as HBM2 is difficult to fab and expensive. As it stands now, upcoming consumer GeForce cards based on Volta will not use HBM2.

This is a new claim that originates from Fudzilla, who states that this info came from well-informed sources. I tend to agree as nothing is pointing or indicating towards HBM2 from Nvidia at this moment in time. From the looks of it, small and compact HBM2 memory simply is too expensive for consumer products. Nvidia does use HBM2 on its GV100-gpu, but much like what happened with Pascal, here again they will revert to GDDR5X memory for the consumer parts."

"While HBM2 might seem to offer better memory-bus width, latency and lower voltages, it seems harder to fab and implement and likely thus is more costly and complicated to use than expected. Just look at the massive AMD Radeon RX Vega (consumer) delay, this likely is due to the limited availability of HBM2. Meanwhile Nvidia has been plastering GDDR5X in their recent high-end SKUs. GDDR5X still rocks hard this year. Next to HBM2, GDDR6 is the new thing for 2018. The new graphics memory type will offer up to 16Gb (2GB) per IC. So eight ICs would already get you to 16 GB of graphics memory.

Nvidia could make use of SK Hynix and/or Micron GDDR5X memory. "

http://www.guru3d.com/news-story/upcoming-geforce-gtx-cards-use-gddr5x-not-hbm2.html

 

[FordGT90Concept]

Makes me wonder if AMD is going to be forced to ditch HBM2 for consumer Vega cards. It probably wouldn't be called Vega if they did but you know what I mean.

 

[Toothless]

I'm pretty sure AMD screwed up bad with using HBM2 and trying to be the lead dog here. NVIDIA worked their magic with GDDR5 and GDDR5X and it's working pretty good.

 

[the54thvoid]

Makes me wonder if AMD is going to be forced to ditch HBM2 for consumer Vega cards. It probably wouldn't be called Vega if they did but you know what I mean.

People applaud AMD for pushing the boundary on HBM2 but as a metric of profitability, it's like shooting themselves in the foot. With GDDR5x and GDDR6 still making progress, it's absolutely unnecessary to move to HBM2 on consumer cards. Its a crazy fixation that is not serving them well. I disagree with those that say they need to do it for energy reasons, the power draw isnt that much greater. They need to do it because they want to but commercially it's a dead end unless it's on small volume, hugely expensive cards, looking at you GP100.

The 1st person to wade in here with some "Nvidia ripping people off again not using HBM" tirade should walk straight back out because they nothing about successful business models. Use what is cheaper and more abundant because you can make more money and push more units. HBM2 is not AMD's saviour, it's their heroin.

 

[ppn]

Since GDDR6 is the thing of 2018, using GDDR5 means November release and 12nm refresh, no need to postpone.

 

[qubit]

I'm sure that whatever memory type NVIDIA uses their cards will work very well.

The 1st person to wade in here with some "Nvidia ripping people off again not using HBM" tirade should walk straight back out because they nothing about successful business models.

NVIDIA are ripping people off again for not using the best and fastest memory type!!!!!

There, said it.

 

[P4-630]

I'm sure that whatever memory type NVIDIA uses there cards will work very well.

Agreed!
I don't mind not having HBM2 when GDDR5X or if GDDR6 does the job as well, no problem with that!

 

[FordGT90Concept]

HBM2 is lower latency, by a lot.

 

[Dethroy]

While it is probably the way better business move/decision, it still makes me kinda sad as an tech enthusiast.
I was finally planning on upgrading my GPU next year, after Nvidia didn't release anything interesting (from a technical point of view; not performance wise) since Kepler in all this years and AMD's performance simply wasn't there... Volta was said to be a completely new architecture, and the prospect of it sporting HBM2 as well, would of made its release even more exciting/ compelling. Even though most people have given up on Vega by now, I really hope it delivers and pulls a Ryzen so that we'll finally see some competition again.

 

[FordGT90Concept]

Even if it does pull a Ryzen, I think "too little, too late" is applicable.

 

[cdawall]

What is curious is that this contradicts the information that came from hynix. They stated that they are having to ramp production up for a 1st quarter 2018 flagship card release. Now they didn't mention brand at all so maybe that was the hint when amd will finally release vega.

 

[Toothless]

What is curious is that this contradicts the information that came from hynix. They stated that they are having to ramp production up for a 1st quarter 2018 flagship card release. Now they didn't mention brand at all so maybe that was the hint when amd will finally release vega.

Cept Vega is so far late that it won't really matter anymore. NVIDIA will work some magic and still have better performance.

 

[cdawall]

Cept Vega is so far late that it won't really matter anymore. NVIDIA will work some magic and still have better performance.

Nvidia didn't even release their top tier GPU to consumers this past go. I mean how far ahead do you have to be to not even try to release something that is the absolute best?

I guess the original road maps did show vega and volta competing, but hell AMD has been hype training that shit for 8 months now.

 

[HTC]

Cept Vega is so far late that it won't really matter anymore. NVIDIA will work some magic and still have better performance.

Not sure but i think that, while both camps can take advantage of using HBM2 over GDDR RAM, AMD need HBM2 for the cards to work as intended while nVidia can "get by" with using GDDR: it's like the penalty for using GDDR over HBM2 is quite allot higher for AMD then for nVidia.

 

[cdawall]

Not sure but i think that, while both camps can take advantage of using HBM2 over GDDR RAM, AMD need HBM2 for the cards to work as intended while nVidia can "get by" with using GDDR: it's like the penalty for using GDDR over HBM2 is quite allot higher for AMD then for nVidia.

The term you are looking for is AMD is bandaiding poor performance with HBM.

 

[Fluffmeister]

At this point, Vega needs to blow the GTX 1080 Ti out of the water... I suspect it won't.

 

[HTC]

The term you are looking for is AMD is bandaiding poor performance with HBM.

Not quite: it seems that Vega was designed with only HBM2 in mind while Volta was not. Because of the VRAM type differences, switching to GDDR actually hurts Vega cards allot more then Volta.

 

[Toothless]

Not quite: it seems that Vega was designed with only HBM2 in mind while Volta was not. Because of the VRAM type differences, switching to GDDR actually hurts Vega cards allot more then Volta.

Yeah and using HBM2 gave NVIDIA a lot of time to be free with the 1080TI and have nothing come against it.

 

[Fluffmeister]

Not quite: it seems that Vega was designed with only HBM2 in mind while Volta was not. Because of the VRAM type differences, switching to GDDR actually hurts Vega cards allot more then Volta.

That also begs the question, what are they doing about the pesky GP104? A 314mm² chip in the form of the GTX 1070/1080 that still has no competition, cutting down a HBM2 chip sounds expensive to me, especially when nVidia have been laughing all the way to the bank for over a year now.

 

[FordGT90Concept]

I think AMD has been completely preoccupied with Ryzen that Vega has been on the backburner. I wouldn't be surprised if what cdawall sniffed out what is happening here: Vega is going to be in low production for corporate compute customers until the beginning of 2018. Then we'll start seeing mass production of consumer Vega cards. Why it makes sense:
1) the pressure to get Ryzen fully released will be gone (they'll have consumer CPU, consumer APU, HEDT, and server processors out by then)
2) it gives AMD time to iron out issues with the hardware and software. Major changes to the rasterizer requires a lot of polish. This was something that AMD had to do sooner or later to compete with NVIDIA in consumer graphics. These things really aren't necessary for compute which is why AMD is targeting compute right now in low volumes.
3) it gives AMD and Hynix time to refine fabrication techniques for improved yields (cheaper unit price).


If #2 is true, Vega Frontier Edition cards are going to be really poor for gaming and not very reflective of Vega's gaming performance when they do finally debut the consumer cards.

 

[Vya Domus]

AMD pushed for HBM so early in order to mitigate their losses in terms of production costs and time for a technology that will certainly be the way to go at one point. They didn't do it to band-aid poor performance or to be "top dog" in innovation. Nvidia on the other hand has the cash to switch over to HBM for consumer products much faster albeit at probably a much higher cost. That being said HBM is still not necessary , latency is certainly not as important as bandwidth for GPU's ,and in this regard you can still use plain old GDDR and slap a huge bus interface and your done. Note how Nvidia is holding back on bus sizes for their last couple of generations , they are doing this so that they can give a big boost to memory bandwidth should they ever need it with little cost.

AMD is in a position where they need to push for newer technologies more than Nvidia or even Intel. Nvidia is focusing on other things at the moment.

 

[FordGT90Concept]

I think AMD initially pushed for HBM because they were hoping Sony/Microsoft would be interested in it for consoles. It was also relatively cheap and easy for them to make it.

Considering how much effort AMD put into the HBCC, I'm thinking they're thinking HBM2 might make its way onto future APUs (MCM cache for the graphics chip). Hell, it might serve as an L4 cache for the CPU too.

 

[Vya Domus]

Considering how much effort AMD put into the HBCC, I'm thinking they're thinking HBM2 might make its way onto future APUs (MCM cache for the graphics chip). Hell, it might serve as an L4 cache for the CPU too.

Exactly , they had long term plans for HBM that go beyond just gaming. We'll see if they come to fruition.

 

[RejZoR]

HBC exists even on already available cards. Remember those with onboard SSD? Well, those already use HBC system. They just had to re-purpose it for gaming cards. Besides, while everyone take AMD as bunch of dummies, they are planning things very long term, sort of "world domination" style. Intel is almost planning for the moment and AMD is clumsily but efficiently been shaping the ecosystem landscape to their favor. Just look what they've done with APU's, what Ryzen has done with "moar cores", what HBM is doing, how pretty much entire console ecosystem is shaped around AMD etc etc.

 

[Leothelesser]

I would love to know the numbers of GPUs sold by AMD in the last 6 months. It’s hard to find RX580/570 anywhere and stock sells out immediately. Data Mining loves Polaris and AMD can’t seem to make them fast enough to keep them on the shelves. AMD not having top end cards at all, let alone able to challenge NVidia top end, probably isn’t hurting the bottom line.

I also note the price hike on 580/570 with Rx 580 over $500 in Australia. Rx 480s were $440 initially and about $380 when RX 580 launched