[TT] Pascal rumored to use GDDR5X..

[Cookie Monster]

Wasn't it always true that GPUs were rarely bandwidth limited? I remember throughout many generations of nVIDIA/AMD(ATi) GPUs, hitting the memory bandwidth bottleneck was hardly the case because the shader core would come to a crawl first.

I assume thats the case with even today's generation of high end cards? a ~5% performance increase at best from overclocking the memory and usually the larger gains come from overclocking the core..

 

[iiiankiii]

LOL. Are you guys reading each other post at all or are you guys arguing just for the heck of it?

Seems to me, all of you guys are agreeing with each other that HBM is faster, allows for smaller form factor and consumes less energy compared to GDDR5.

I'm confident Big Pascal and Arctic Island will do a much better job of using HBM's advantages over GDDR5.

 

[4K_shmoorK]

LOL. Are you guys reading each other post at all or are you guys arguing just for the heck of it?

That is all the graphics section of this forum is. A bunch of one-upping and 'I told you so' followed up with a dash of 'If you knew as much as me'.

It is hard to have any real conversation when most (or a vocal few) feel the need nitpick every single word of a response in hopes of appearing superior to the person in which they are replying to.

It's no fun, I tell ya. No fun at all.

 

[psolord]

There is little effect, but there is one. That is why i was asking if HBM was really overkill. If it was, I wouldn't expect any effect at all.

Isn't memory overclocking reducing the latency as well?

Maybe this is why it is presenting a measurable increase.

 

[AtenRa]

Isn't memory overclocking reducing the latency as well?

Maybe this is why it is presenting a measurable increase.

yes one reason may be lower latency.

 

[ShintaiDK]

I would guess L2 cache speed could also be affected.

 

[JDG1980]

Maybe I'm missing something, but it seems to me that GDDR5X could conceivably be used to save power and die space compared to standard GDDR5 (even if it doesn't do so to as great an extent as HBM does).

Having a wider memory controller means a bigger die, and more power consumption if all else is equal. In AMD's case, there is also the issue that on newer cards its memory controllers are being run at much higher clocks than they were originally designed for (~1250 -> 1500 MHz), thus hurting power consumption worse. But the memory controllers in Nvidia's Maxwell don't seem to have that issue, which indicates that a high-speed GDDR5 memory controller doesn't have to be inefficient if it is designed that way from the ground up.

Based on the descriptions so far, it looks like GDDR5X doubles bandwidth by various optimization tricks, not by running a higher raw clock. Therefore, 10 Gbps GDDR5X should run at the same (1250 MHz) base clock as 5 Gbps standard GDDR5, and hopefully the memory controller should use no more power. That means that the manufacturers can get away with a narrower memory bus than they're using now.

For instance, the GTX 980 has a 256-bit bus with 7.012 Gbps GDDR5. This provides memory bandwidth of 224 GB/sec. Assume that only the low-end versions of GDDR5X (10 Gbps) are available at first. A hypothetical GTX 980 successor with GDDR5X could then cut the bus width to 192 MHz, and actually see an increase in memory bandwidth:
224 * (192/256) * (10000/7012) = ~239.5

 

[LTC8K6]

There is little effect, but there is one. That is why i was asking if HBM was really overkill. If it was, I wouldn't expect any effect at all.

I really wonder if Fiji pro would be noticeably slower with the same lower cost memory as the 390 series cards.

The BW isn't that much lower at 384GB/s.

Perhaps it could make for a cheaper Fury pro card?

 

[MrTeal]

I really wonder if Fiji pro would be noticeably slower with the same lower cost memory as the 390 series cards.

The BW isn't that much lower at 384GB/s.

Perhaps it could make for a cheaper Fury pro card?

There isn't a GDDR5 controller on Fiji (that's been revealed), so the cost would be very high as they'd need to create a separate die just for that part.

 

[Silverforce11]

Maybe I'm missing something, but it seems to me that GDDR5X could conceivably be used to save power and die space compared to standard GDDR5 (even if it doesn't do so to as great an extent as HBM does).

Having a wider memory controller means a bigger die, and more power consumption if all else is equal. In AMD's case, there is also the issue that on newer cards its memory controllers are being run at much higher clocks than they were originally designed for (~1250 -> 1500 MHz), thus hurting power consumption worse. But the memory controllers in Nvidia's Maxwell don't seem to have that issue, which indicates that a high-speed GDDR5 memory controller doesn't have to be inefficient if it is designed that way from the ground up.

Based on the descriptions so far, it looks like GDDR5X doubles bandwidth by various optimization tricks, not by running a higher raw clock. Therefore, 10 Gbps GDDR5X should run at the same (1250 MHz) base clock as 5 Gbps standard GDDR5, and hopefully the memory controller should use no more power. That means that the manufacturers can get away with a narrower memory bus than they're using now.

For instance, the GTX 980 has a 256-bit bus with 7.012 Gbps GDDR5. This provides memory bandwidth of 224 GB/sec. Assume that only the low-end versions of GDDR5X (10 Gbps) are available at first. A hypothetical GTX 980 successor with GDDR5X could then cut the bus width to 192 MHz, and actually see an increase in memory bandwidth:
224 * (192/256) * (10000/7012) = ~239.5

We don't know enough about GDDR5X to go into that detail on speculation with accuracy. Lots of known, the major one being: the doubling of sampling frequency to double the bandwidth, does that have a drawback? How does that relate to the current GDDR5 MC?

Current gen mid-range may do with ~250 GB/s, obviously next-gen on a new node mid-range will need double that unless they vastly improve their compression tech.

 

[railven]

LOL. Are you guys reading each other post at all or are you guys arguing just for the heck of it?

Seems to me, all of you guys are agreeing with each other that HBM is faster, allows for smaller form factor and consumes less energy compared to GDDR5.

I'm confident Big Pascal and Arctic Island will do a much better job of using HBM's advantages over GDDR5.

Second post in this thread pretty much summed up the conversation.

How it turned into HBM (ie AMD) vs GDDR5X (ie NV) is beyond me. If post #2 turns out true and AMD ends up using GDDR5X, then 90% of the bickering in this thread is pointless.

"GDDR5x is a band aid fix to HBM's growing pain."
"So what you're saying is that HBM is a failure."

I hate those effin' posts. If you start your post with "so what you're saying" just stop.

 

[thesmokingman]

Maybe I'm missing something, but it seems to me that GDDR5X could conceivably be used to save power and die space compared to standard GDDR5 (even if it doesn't do so to as great an extent as HBM does).

Nobody knows anything. This is another clickbait piece by tweaktown yet again. It's hilarious when they are quoting themselves imagining what would be what. Read the link, it's crazy.

In our post about NVIDIA testing its new Pascal GPU internally, I did say "I would love to see a GDDR5-based offering for the cheaper mainstream side, and a HBM2-powered GeForce 1000 series card for the enthusiasts" and it looks like I could be right on the mark if the current rumors are true. It looks like we could expect a GDDR5X-powered, Pascal-based GeForce card while the enthusiast-class cards will enjoy the much faster, but more expensive HBM2 technology.

lol what is this? A self fulfilling prophecy?

 

[tential]

Second post in this thread pretty much summed up the conversation.

How it turned into HBM (ie AMD) vs GDDR5X (ie NV) is beyond me. If post #2 turns out true and AMD ends up using GDDR5X, then 90% of the bickering in this thread is pointless.

"GDDR5x is a band aid fix to HBM's growing pain."
"So what you're saying is that HBM is a failure."

I hate those effin' posts. If you start your post with "so what you're saying" just stop.

Because the second you talk HBM, someone HAS to bring up how Nvidia can use GDDR5/x to compete just fine.
HBM can't be talked about by itself until both AMD and Nvidia have it. Then, people will be able to think about HBM vs GDDR5 for their vendor.

It just can't be seen as a whole when so many people are obsessed with how it effects their favorite GPU vendor rather than the technology by itself.

HBM has benefits, but it's not like it's a requirement for a card to be good. Why this idea can't be understood is beyond me.

All I care about at the end of the day is the end product, and it was not like I disliked the 980Ti. So I'm excited for Big Pascal! My guess is it comes out first and sets a baseline for what the high end performance will be. That will let me know just how excited I need to be about spending my money. It will decide whether I rush out day 1, or whether I wait til a sale as I have a backlog of games that won't benefit much from higher res as they're just that old, and I'm expecting SOMETHING worthwhile from this new node.

Pascal performance should be impressive, end statement, or I'll be very upset.

 

[LTC8K6]

There isn't a GDDR5 controller on Fiji (that's been revealed), so the cost would be very high as they'd need to create a separate die just for that part.

I don't think it's too farfetched an idea that AMD might have a GDDR5 Fiji chip design ready.

Either as a hedge for HBM problems/delays, or as a future cheaper card.

But I really just meant theoretically.

 

[thesmokingman]

I don't think it's too farfetched an idea that AMD might have a GDDR5 Fiji chip design ready.

Either as a hedge for HBM problems/delays, or as a future cheaper card.

But I really just meant theoretically.

It's doubtful, when you consider the level of integration of HBM into the silicon. New technologies had to be invented and perfected for it to happen. I would be surprised if AMD put aside what they developed, and Nvidia passing on the future and any chance to catch up to AMD on the HBM front, just for a stop gap that brings a lot of power concerns. GDDR5X is going in the opposite direction to HBM.

 

[LTC8K6]

It's doubtful, when you consider the level of integration of HBM into the silicon. New technologies had to be invented and perfected for it to happen. I would be surprised if AMD put aside what they developed, and Nvidia passing on the future and any chance to catch up to AMD on the HBM front, just for a stop gap that brings a lot of power concerns. GDDR5X is going in the opposite direction to HBM.

GDDR5, not 5X.

5X isn't even in production yet, and isn't expected on the market until late 2016.

 

[MrTeal]

I don't think it's too farfetched an idea that AMD might have a GDDR5 Fiji chip design ready.

Either as a hedge for HBM problems/delays, or as a future cheaper card.

But I really just meant theoretically.

Well, as a theoretical you might be able to get at least an idea if you can get someone with a Fury to downclock their RAM to 380MHz to simulate the bandwidth, and see what kind of effect there is on framerates. There's scores of other uncontrolled variables, but it'd be interesting to see.

 

[Goatsecks]

I don't think it's too farfetched an idea that AMD might have a GDDR5 Fiji chip design ready.

Either as a hedge for HBM problems/delays, or as a future cheaper card.

But I really just meant theoretically.

I think a GDDR5 fiji would be a great idea for AMD, would be very interesting to compare to HBM fiji.

On a similar note, is it not fairly likely that AMD are going to be using GDDR5, or a variation of, on at least some of their next gen cards?

 

[railven]

I think a GDDR5 fiji would be a great idea for AMD, would be very interesting to compare to HBM fiji.

On a similar note, is it not fairly likely that AMD are going to be using GDDR5, or a variation of, on at least some of their next gen cards?

If any of the HBM rumors are true, I'd expect to see AMD roll out something with GDDR5 on their new node. Possibly treating it as a pipe cleaner? While I wouldn't put it pass AMD, but if HBM2 is seriously delayed, I doubt AMD wants to repeat where NV is basically left uncontested for over a year.

 

[Techhog]

If any of the HBM rumors are true, I'd expect to see AMD roll out something with GDDR5 on their new node. Possibly treating it as a pipe cleaner? While I wouldn't put it pass AMD, but if HBM2 is seriously delayed, I doubt AMD wants to repeat where NV is basically left uncontested for over a year.

They don't have the money for heavy chip redesigns, though. They have to make a decision about which direction to go in long before they know if it'll be ready.

 

[sm625]

GTX280 with GDDR-3 was faster than HD4870 with GDDR-5.

Yeah but the GTX280 die was literally twice as big as a 4870, and then some. Yet it was only a few percentage points faster. AMD had nvidia completely beaten back in those days. The competition wasnt even close in terms of fps per area of silicon. Yet I think nvidia might have sold more GTX2xx cards anyway? I'm not sure, but I do know that AMD did not claim as much market share as they should have given the product disparity.

 

[railven]

They don't have the money for heavy chip redesigns, though. They have to make a decision about which direction to go in long before they know if it'll be ready.

I meant more so, we might see Hawaii/Grenada[Tahiti/Tonga?] with a node shrink make another showing.

If AMD is heavily tied to HBM for Arctic Islands, unless they want to repeat the long gap between Tahiti and Hawaii and then Hawaii and Fiji, they're going to put something out there. [Hell, of opinion had they refreshed Hawaii into Grenada in Q3/Q4 2014, they wouldn't have lost so much marketshare]. With the DX12 perfs everyone is happy with, why not? If Hawaii/Grenada can compete with GTX 980 Ti, let it go against GTX 1070 or whatever NV decides to do.

This is all assuming any of the rumors have any credibility. Otherwise, we'll see Arctic Islands rocking HBM1 or HBM2 and nothing to worry about.

 

[Techhog]

I meant more so, we might see Hawaii/Grenada[Tahiti/Tonga?] with a node shrink make another showing.

If AMD is heavily tied to HBM for Arctic Islands, unless they want to repeat the long gap between Tahiti and Hawaii and then Hawaii and Fiji, they're going to put something out there. [Hell, of opinion had they refreshed Hawaii into Grenada in Q3/Q4 2014, they wouldn't have lost so much marketshare]. With the DX12 perfs everyone is happy with, why not? If Hawaii/Grenada can compete with GTX 980 Ti, let it go against GTX 1070 or whatever NV decides to do.

This is all assuming any of the rumors have any credibility. Otherwise, we'll see Arctic Islands rocking HBM1 or HBM2 and nothing to worry about.

I guess. Also, Tahiti and Tonga are different chips. Tahiti is retired for good.

 

[thesmokingman]

GDDR5, not 5X.

5X isn't even in production yet, and isn't expected on the market until late 2016.

Same difference, either way I don't see it happening as it is going backwards in the TDP reduction course especially in their current precarious situation.

 

[LTC8K6]

Higher density GDDR5 chips should use less power since you need less of them?

I'm not sure if 20nm is smaller than current GDDR5 lithography, but if it is, that should reduce power as well?