AMD Locks Down HBM Frequency on Fiji Cards

StriderX · 19 Jun 2015 at 12:10

Vram OC is about as useful as Ed Miliband.

bru · 19 Jun 2015 at 12:20

It has probably been done, to ensure it works for the life of the product, remember these things will have two or three years warranty ( saying that they will all come with only a year now and we can all be worried

). HBM is new nobody knows for sure how it will last with the abuse that a bunch a crazy PC nerds will be throwing at it, locking it down seems like a sensible (but disappointing) solution to me.

queamin · 19 Jun 2015 at 12:35

This is the first HBM out if you go back and look at the first GDDR5 specs and you see they are a lot lower than what is out now.

I think if they didn't do this the launch could be messy with a lot of returns if people tried to over clock them like GDDR5 without really understanding the tech first.

rtho782 · 19 Jun 2015 at 12:36

MadMossy said:
It's not just the bandwidth but also the bus width being much much wider than traditional GDDR5 set ups.

4096-bit vs 384/512.

That's a massive chunk of data that can be pushed through up to 8 times faster than GDDR5.

8 times faster?

See, this is the bit that annoys me.

Wider bus, lower clock speed. The two are interchangeable. A 512bit bus at 8GT/s would be the same as Fury X's 4096bit bus at 1GT/s.

We have a 512bit bus at 6GT/s (390x) and we have a 384bit bus at 7GT/s (oc's to 8GT easily) in 980ti/TX.

Fury is faster, but it's 33% faster, not 800% faster.

PCIe 3.0 x4 is the same speed as PCIe 1.0 x16. PCIe 1.0 x16 is wider, 3.0 x4 faster. They are the same though.

IF either has an advantage, it's PCIe 3.0 (narrow, fast) not 1.0 (wide, slow) as for small packets the latency is lower.

TheMorningStar · 19 Jun 2015 at 12:42

Well, if you think about all the power this card uses at stock but with a 50c stock peak load was it? That extra power could well go to the core. have a feeling these will be in range of 1500mhz core clocks. Just wondering what will be the limitation, power draw or thermals.

flopper · 19 Jun 2015 at 12:57

rtho782 said:
8 times faster?

See, this is the bit that annoys me.

Wider bus, lower clock speed. The two are interchangeable. A 512bit bus at 8GT/s would be the same as Fury X's 4096bit bus at 1GT/s.

gddr5 and HBM operates a bit differently. cant use the same logic for two different ways the tech operates

TheMorningStar said:
Well, if you think about all the power this card uses at stock but with a 50c stock peak load was it? That extra power could well go to the core. have a feeling these will be in range of 1500mhz core clocks. Just wondering what will be the limitation, power draw or thermals.

I guess we find out soon 5 days to go.

Silent_Scone · 19 Jun 2015 at 12:58

flopper said:
gddr5 and HBM operates a bit differently. cant use the same logic for two different ways the tech operates

This is likely applicable to many things regarding HBM but not this. You are still limited by the speed of the IC

tommybhoy · 19 Jun 2015 at 13:01

Can the memory actually be oc'ed in the first place or does it need to be static should be the question asked.

For all we know at this moment in time regarding oc'ing core/memory-integration...ie nothing

, perhaps wait a week until the tech bumph comes out-it's not rocket science.

Silent_Scone said:
So the sarcasm wasn't needed.

Honestly don't know if your taking the hit n miss with boom there(never mind the reasoning for the thread-but benefit of the doubt and all that).

Silent_Scone · 19 Jun 2015 at 13:04

tommybhoy said:
Can the memory actually be oc'ed in the first place or does it need to be static should be the question asked.

For all we know at this moment in time regarding oc'ing core/memory-integration...ie nothing, perhaps wait a week until the tech bumph comes out-it's not rocket science.

Honestly don't know if your taking the hit n miss with boom there(never mind the reasoning for the thread-but benefit of the doubt and all that).

This has already been mentioned

Silent_Scone said:
I think it's either one of two things. It will be to do with degradation, it will likely get warm as you suggest. That, or it is likely performance related. There is probably a great deal of error correction in HBM.

Deleted User 234324 · 19 Jun 2015 at 13:06

rtho782 said:
8 times faster?

See, this is the bit that annoys me.

Wider bus, lower clock speed. The two are interchangeable. A 512bit bus at 8GT/s would be the same as Fury X's 4096bit bus at 1GT/s.

We have a 512bit bus at 6GT/s (390x) and we have a 384bit bus at 7GT/s (oc's to 8GT easily) in 980ti/TX.

Fury is faster, but it's 33% faster, not 800% faster.

PCIe 3.0 x4 is the same speed as PCIe 1.0 x16. PCIe 1.0 x16 is wider, 3.0 x4 faster. They are the same though.

IF either has an advantage, it's PCIe 3.0 (narrow, fast) not 1.0 (wide, slow) as for small packets the latency is lower.

While on the surface that may be true, you've missed out on vital part of it all.

There is no memory controller or length interconnects connecting the VRAM to the GPU.

It's literally on top of the GPU so latency will be minimal if any at all, that alone can drastically alter the available speed. It's almost like comparing a HDD to SSD, they both can have the same storage capacity but one is just faster at accessing it than the other.

Devrij · 19 Jun 2015 at 13:09

Memory clocking rarely yields good improvements anyway compared to core, so I'm not fussed. It's usually not worth it for the stability problems it can introduce imo.

Jono8 · 19 Jun 2015 at 13:18

Looks like it will have more than enough bandwith anyway so can't see this being an issue.

As it is such a new thing/early production they probably cant overclock much and keep stable anyway.

AgedIdiot · 19 Jun 2015 at 13:26

LoadsaMoney said:
Doesn't really bother me, only got my core overclocked on my 290, memory didn't seem to make much difference, so just put it back to stock and left it.

+ 1

tommybhoy · 19 Jun 2015 at 13:26

Silent_Scone said:
This has already been mentioned

So what?

Everything discussed in the thread is simply assumption and will be clarified on release.

Gerard · 19 Jun 2015 at 13:30

Don't care really, i can clock my vram up hundreds of mhz and what difference do i see in game? Virtually zero. I'd be interested in knowing just how many ram modules are on the board, yes we can see 4 but if they are vertically stacked on top of each other with an interposer in between there could be quite a few ram modules we can't actually see.

r7slayer · 19 Jun 2015 at 13:32

I'll be honest comming from some one who favors AMD more at the moment (Though i go where the best performance for my money is)

From the statement AMD made about the card being a overclockers dream im kinda gutted about the memory being locked down if this is true. And scratching my head to why they would call it a overclockers dream card. But at the same time i'm thinking do we really need anymore bandwidth from the vram?

I saw little performance increase from overclocking the vram on my 290x and i think AMDMatt (LtMatt back then) did a thread showing the effects of overclocking the memory had on the 290 series and it was hardly noticeable. He said that the best gains were from overclocking the core. This could be exactly the same case with the fury x and with the HBM modules being next to the core, just a 25mhz clock bump is quite substantial. People could overclock their memory more than they realize and break their cards. If there is more than enough bandwidth this might be why there are little gains from overclocking the memory and locked it down.

I'm sure AMD did it with good reasoning. But im going to wait and see anyways as im sure you can still overclock the memory if you really wanted to in the bios.

Greebo · 19 Jun 2015 at 13:39

I have seen in lots of articles that overclocking the memory on gddr5 cards makes very little difference to framerates.

And its quite often with my overclocking attempts that its the memory overclock which makes it fail.

Silent_Scone · 19 Jun 2015 at 13:47

tommybhoy said:
So what?

Everything discussed in the thread is simply assumption and will be clarified on release.

Hi,

You're entitled to make your own speculative reason informed or uninformed. Just as much as you're entitled to not care either way, no peer pressure here. If everything was simply waiting and relaying legalisation laid out by manufacturers we might as well all be robots.

meldarthx · 19 Jun 2015 at 13:52

r7slayer said:
I'll be honest comming from some one who favors AMD more at the moment (Though i go where the best performance for my money is)

From the statement AMD made about the card being a overclockers dream im kinda gutted about the memory being locked down if this is true. And scratching my head to why they would call it a overclockers dream card. But at the same time i'm thinking do we really need anymore bandwidth from the vram?

I saw little performance increase from overclocking the vram on my 290x and i think AMDMatt (LtMatt back then) did a thread showing the effects of overclocking the memory had on the 290 series and it was hardly noticeable. He said that the best gains were from overclocking the core. This could be exactly the same case with the fury x and with the HBM modules being next to the core, just a 25mhz clock bump is quite substantial. People could overclock their memory more than they realize and break their cards. If there is more than enough bandwidth this might be why there are little gains from overclocking the memory and locked it down.

I'm sure AMD did it with good reasoning. But im going to wait and see anyways as im sure you can still overclock the memory if you really wanted to in the bios.

I have a feeling that's exactly how HBM OC will work - instead of looking at bumping it by 100s mhz; we're looking at 10s of mhz or single numbers.

People won't know this; thinking something not like GDDR and boom their goes their card...

Better to lock it down this round; get people used to it; know what you can do with it before unleashing them.

I have a feeling the cores will OC a lot.....and I mean a lot 500mhz+

rtho782 · 19 Jun 2015 at 13:53

MadMossy said:
While on the surface that may be true, you've missed out on vital part of it all.

There is no memory controller or length interconnects connecting the VRAM to the GPU.

It's literally on top of the GPU so latency will be minimal if any at all, that alone can drastically alter the available speed. It's almost like comparing a HDD to SSD, they both can have the same storage capacity but one is just faster at accessing it than the other.

The latency from the signal travelling along the circuitboard, at the speed of light, is nothing.

We don't actually know what the memory timings are like, I suspect measured in nanoseconds they will be similar to other memory.

The things you have named are big advantages - lower power, smaller footprint - but don't have anything to do with how fast it is.

It's 33% faster than 390x. Not 800% faster.

An SSD is no faster on a 1inch SATA cable than it is on an 18inch one. mSATA SSDs tend to be slower.