• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

From ATI to AMD back to ATI? A Journey in Futility @ [H]

https://videocardz.com/69662/raja-koduri-explains-where-radeon-rx-vega-in-reddit-ama

Some answers here

2 stacks of HBM (8 gb each giving 16 gb)

Frontier version not suited for games (drivers are not optimized for games)

16gb for consumer they are looking into that, pretty much it seems will come down to pricing of the card it seems.

Frontier edition, will be the fastest compute output of all the Vega's. That puts gaming vega's at less clocks or are cut down versions (I think its the prior as they have shown those types of cards to public already)

Interesting talking about the water cooled version, "may have more thermal headroom for specific scenarios" that doesn't' sound promising. Sounds like they are pushing the card to the max.

1 6 and 8 pin looks to be ES, retail might have 2 8 pin, so yeah ~300 watt neighborhood looks to be correct, I'm going to guess at around 250 watts stock any amount of overclocking its going to push the power usage up considerably. So that artist rendition was not by mistake.

Big thing though, it will be there at Computex but won't be available for purchase till after. So Q2 purchasing not happening but launch will happen, maybe.
 
Last edited:
What I do not get is AMD is big on BW, so why did they not create the Frontier Edition with 3 or 4 stacks just like Nvidia does with Gx100?
It would give the Frontier the ideal BW while also providing the critical 24GB-32GB that the top card in certain professional market needs (especially HPC).
The cost to design such a GPU would not be excessive, considering it seems they have paid SK Hynix to deliver 1st and foremost 8GB HBM2 to them even though it is costing SK Hynix in terms of not getting 1.6Gbps 4GB HBM2 out the door (now 7 months behind when they said it would be).

The Frontier is being pushed for Data Scientists and other elite fields and long run HPC, they really could had put pressure on Nvidia if they went with the 24GB or 32GB as this is something quite a few of Nvidia clients was expecting with the Tesla.
This model of Vega goes beyond relying upon the drivers with the evolution of the Dynamic RAM (HBCC).
Cheers
 
Last edited:
probably not enough performance to fully saturate that much bandwidth. Or they are not expecting to sell too many because they don't directly compete because of the lack of professional software. Something has to be there not to have more bandwidth, HPC market targets that Vega FE will be in need a butt load of bandwidth.
 
razor1 said:
probably not enough performance to fully saturate that much bandwidth. Or they are not expecting to sell too many because they don't directly compete because of the lack of professional software. Something has to be there not to have more bandwidth, HPC market targets that Vega FE will be in need a butt load of bandwidth.
Click to expand...
They will be able to saturate the bandwidth for HPC, look at how Nvidia increased theirs for this segment with V100.
Beyond that clients really do want more memory in this segment, especially scientists and related modelling/algorithm workloads.
It would had been a notable win over Nvidia with its V100, beyond marketing (and that would had been good as well).
Cheers
 
oh ok yeah I see what you are saying, brain fart on my part.

Its probably cost of a bigger die possible and the increased cost of having more stacks of HBM2.
 
On the compute side of things..Vega FE will be the fastest single GPU solution (>12.5 TFlops FP32) when it's available and our NCU packs several additional optimizations, including Rapid-Packed-Math which delivers >25 TFLops of FP16
Click to expand...
We haven't mentioned any multi GPU designs on a single ASIC like Epyc, but the capability is possible with Infinity Fabric.
Click to expand...
Few Raja quotes I found interesting.

razor1 said:
Frontier edition, will be the fastest compute output of all the Vega's. That puts gaming vega's at less clocks or are cut down versions (I think its the prior as they have shown those types of cards to public already)
Click to expand...
Raja said:
Consumer RX will be much better optimized for all the top gaming titles and flavors of RX Vega will actually be faster than Frontier version!
Click to expand...
If you say so.

CSI_PC said:
What I do not get is AMD is big on BW, so why did they not create the Frontier Edition with 3 or 4 stacks just like Nvidia does with Gx100?
It would give the Frontier the ideal BW while also providing the critical 24GB-32GB that the top card in certain professional market needs (especially HPC).
The cost to design such a GPU would not be excessive, considering it seems they have paid SK Hynix to deliver 1st and foremost 8GB HBM2 to them even though it is costing SK Hynix in terms of not getting 1.6Gbps 4GB HBM2 out the door (now 7 months behind when they said it would be).

The Frontier is being pushed for Data Scientists and other elite fields and long run HPC, they really could had put pressure on Nvidia if they went with the 24GB or 32GB as this is something quite a few of Nvidia clients was expecting with the Tesla.
This model of Vega goes beyond relying upon the drivers with the evolution of the Dynamic RAM (HBCC).
Cheers
Click to expand...
It's back to the quality of bandwidth question. For certain workloads HBM is 4x more effective than GDDR5 and even more compared to GDDR5X. Extra bandwidth might not have been as necessary given the HBCC design. The whole point of HBCC is to lean on system memory or nonvolatile(SSG) for massive amounts of storage. That's far more practical for most HPC uses when programmers don't have to worry about programming around the limited VRAM capacity. Naples/Threadripper also have lots of PCIE lanes to accommodate those designs. Even some of the NVLink benchmarks show that large problems scale with link bandwidth as opposed to all the VRAM bandwidth. Even for graphics, it should have far more effective bandwidth than nearly all current Nvidia products due to the HBM2. Not to mention a move towards FP16 would lessen the burden with comparable loads.
 
Anarchist4000 said:
Few Raja quotes I found interesting.

If you say so.
Click to expand...

Did you actually read ALL of what he stated? They specifically stated in a comment just after the one you quoted Compute wise VEGA FE will be the most performant compute card on the market.

RK: On the compute side of things..Vega FE will be the fastest single GPU solution (>12.5 TFlops FP32) when it’s available and our NCU packs several additional optimizations, including Rapid-Packed-Math which delivers >25 TFLops of FP16
Click to expand...

Yeah If I say so lol, no if Raja says so! That leaves no wiggle room for interpretation, he said she said BS man.

It's back to the quality of bandwidth question. For certain workloads HBM is 4x more effective than GDDR5 and even more compared to GDDR5X. Extra bandwidth might not have been as necessary given the HBCC design. The whole point of HBCC is to lean on system memory or nonvolatile(SSG) for massive amounts of storage. That's far more practical for most HPC uses when programmers don't have to worry about programming around the limited VRAM capacity. Naples/Threadripper also have lots of PCIE lanes to accommodate those designs. Even some of the NVLink benchmarks show that large problems scale with link bandwidth as opposed to all the VRAM bandwidth. Even for graphics, it should have far more effective bandwidth than nearly all current Nvidia products due to the HBM2. Not to mention a move towards FP16 would lessen the burden with comparable loads.
Click to expand...

Really think they can double "quality of bandwidth"? probably not man. And CSI isn't talking about GDDR versions of nV cards that is the same mistake I made, he is talking about the HBM versions.

We are seeing a big difference in the way Raja is talking about products now than when he did Polaris. There is no room for hype, Why would he state in the financial presentation, its not about beating anyone? He is covering his ass and AMD's.
 
Last edited:
razor1 said:
Yeah If I say so lol, no if Raja says so! That leaves no wiggle room for interpretation, he said she said BS man.
Click to expand...
I think your reading comprehension leaves something to be desired.

razor1 said:
Did you actually read ALL of what he stated? They specifically stated in a comment just after the one you quoted Compute wise VEGA FE will be the most performant compute card on the market.
Click to expand...
Yeah, using the exact quote you provided, he said the FE will be the fastest COMPUTE card available when it first releases. So the only available card from the line will be the fastest available card. I guess that means it's the slowest as well. Seems pretty simple and doesn't even apply to graphics or consumer products as they aren't compute.

razor1 said:
Really think they can double "quality of bandwidth"? probably not man.
Click to expand...
Yeah, just need divergence or scalar workloads like SM6 is making available.
 
Anarchist4000 said:
I think your reading comprehension leaves something to be desired.


Yeah, using the exact quote you provided, he said the FE will be the fastest COMPUTE card available when it first releases. So the only available card from the line will be the fastest available card. I guess that means it's the slowest as well. Seems pretty simple and doesn't even apply to graphics or consumer products as they aren't compute.
Click to expand...

He never stated FIRST releases. That is something you added in

Yes I did say compute, right

this was my quote

Frontier edition, will be the fastest compute output of all the Vega's. That puts gaming vega's at less clocks or are cut down versions (I think its the prior as they have shown those types of cards to public already)
Click to expand...

Did you see gaming performance in that, no actually I stated gaming performance for Vega FE is not suited for gaming prior to that too, in this quote in the same post.

Frontier version not suited for games (drivers are not optimized for games)
Click to expand...

That is implied the gaming Vega's will be better for gaming.

Yeah, just need divergence or scalar workloads like SM6 is making available.
Click to expand...

Where is the software to use that? Not there yet, when do you think they will come out? Not anytime soon.

You can try to point out to things that aren't there in my posts all day long, just won't happen, the things I stated about Vega close to year ago, look to be coming true...... All that is left is performance. Specs wise its tallying up.
 
Anarchist4000 said:
Not if it wasn't relevant to the current quarter. Analyst day is different from SEC filings as well. In the case of a MCM it probably wouldn't be a licensing deal, but a direct product integration. The same way Naples is 4 CPUs in a package, Intel could create an APU with a CPU+GPU over PCIE all within a socketed package. That would be a small form factor and not involve licensing of IP, but the creation of an embedded system.
Click to expand...

I think it would be a very important piece of information, and Idoubt investors would be happy if they found out Lisa or Raja failed to inform them.
 
Anarchist4000 said:
Few Raja quotes I found interesting.



If you say so.


It's back to the quality of bandwidth question. For certain workloads HBM is 4x more effective than GDDR5 and even more compared to GDDR5X. Extra bandwidth might not have been as necessary given the HBCC design. The whole point of HBCC is to lean on system memory or nonvolatile(SSG) for massive amounts of storage. That's far more practical for most HPC uses when programmers don't have to worry about programming around the limited VRAM capacity. Naples/Threadripper also have lots of PCIE lanes to accommodate those designs. Even some of the NVLink benchmarks show that large problems scale with link bandwidth as opposed to all the VRAM bandwidth. Even for graphics, it should have far more effective bandwidth than nearly all current Nvidia products due to the HBM2. Not to mention a move towards FP16 would lessen the burden with comparable loads.
Click to expand...

HBCC is not really that important in its entire role but specifically required as a cohesive Unified Memory and virtual pool just like the P100/V100 in this specific segment for HPC and data scientists/modelling/algorithm workloads; VRAM capacity and BW are.
The bold is not directed at you but just generally before anyone thinks I am dismissing HBCC for all segments and solutions.

Which NVLink benchmark are you thinking of regarding scaling?
The issue is probably not primarily NVLink or BW (for Nvidia) but the framework/application/coding with scale-up of GPUs.
You seen how well Caffe/GoogleNet/etc scales with it?

image011.png



image003.png



But anyway BW and VRAM would be important in the type of HPC workloads I mentioned before and would use the system's environment-tools and Unified Memory under their control rather than relying upon the HBCC to swap as used with games or rendering (but that is the SSG or say gaming Vega and so separate product, not the Frontier).

I am thinking primarily of the Frontier for this 3 or 4-stack due to its segment and operation focus.
SSG could be 16GB quite happily for its specific role (which is not HPC/DL/etc).
Cheers
 
Last edited:
razor1 said:
oh ok yeah I see what you are saying, brain fart on my part.

Its probably cost of a bigger die possible and the increased cost of having more stacks of HBM2.
Click to expand...
I believe AMD solution for more ram will be a dual design later, two GPU's and a total of four stacks of HBM2. The only way they will be able to compete against V100 performance and I would think if such was available would beat it as well.

As for more memory, high bandwidth controller and SSD should solve that issue with super size data sets.
 
Stoly said:
I think it would be a very important piece of information, and Idoubt investors would be happy if they found out Lisa or Raja failed to inform them.
Click to expand...


unless they are under NDA, but Intel making a statement like that seems to be shifting the focus away from what they are doing, its not a direct answer, so we still don't know what those talks are about.
 
noko said:
I believe AMD solution for more ram will be a dual design later, two GPU's and a total of four stacks of HBM2. The only way they will be able to compete against V100 performance and I would think if such was available would beat it as well.
Click to expand...


Ram is important but its the bandwidth, 480 Gb/s is really starving that chip for certain work loads.
 
razor1 said:
Ram is important but its the bandwidth, 480 Gb/s is really starving that chip for certain work loads.
Click to expand...
Agreed but with these types of workloads you get full benefit from two gpu's and combined bandwidth of them. Memory wise AMD looks to slaughter Nvidia with their high bandwidth controller and onboard SSD or any other storage device.
 
noko said:
I believe AMD solution for more ram will be a dual design later, two GPU's and a total of four stacks of HBM2. The only way they will be able to compete against V100 performance and I would think if such was available would beat it as well.
Click to expand...

Unfortunately it seems that will be quite awhile away and by then who knows what HBM memory capacity V100 will have.
If done today, they would have a clear advantage and also benefit from more than having 480GB/s that will be a limitation for certain tasks-segments, remember Frontier is HPC related and also for data scientists.
And importantly it would had been a win over Nvidia, in a way that is not just marketing but real world viable and important.
Most interested in Frontier from the data scientist/algorith/DL will also then be considering the MI25 for larger scale solutions such as a node comparable to DGX-1 at minimum.
Makes more sense to give the capability to Frontier and then MI25 if there are supply challenges for now.
IMO anyway.
Cheers
 
noko said:
Agreed but with these types of workloads you get full benefit from two gpu's and combined bandwidth of them. Memory wise AMD looks to slaughter Nvidia with their high bandwidth controller and onboard SSD or any other storage device.
Click to expand...

Nah with two chips each chip needs to be feed, the performance of the Vega FE, each Vega chip should be able to handle 800 gb/s at least. Easiest solution would have been 2 more stacks, but who know maybe it was too late to make that change to the memory controller....
 
razor1 said:
Nah with two chips each chip needs to be feed, the performance of the Vega FE, each Vega chip should be able to handle 800 gb/s at least. Easiest solution would have been 2 more stacks, but who know maybe it was too late to make that change to the memory controller....
Click to expand...
Yeah, just very interesting how this is developing - the ball really hangs on RTG more so than Nvidia since Nvidia is already established in the HPC arena.
 
razor1 said:
https://videocardz.com/69662/raja-koduri-explains-where-radeon-rx-vega-in-reddit-ama

Frontier edition, will be the fastest compute output of all the Vega's. That puts gaming vega's at less clocks or are cut down versions (I think its the prior as they have shown those types of cards to public already)
Click to expand...

Elmnator: Will the consumer RX version be as fast at the Frontier version?
RK: Consumer RX will be much better optimized for all the top gaming titles and flavors of RX Vega will actually be faster than Frontier version!
 
fingerbob69 said:
Elmnator: Will the consumer RX version be as fast at the Frontier version?
RK: Consumer RX will be much better optimized for all the top gaming titles and flavors of RX Vega will actually be faster than Frontier version!
Click to expand...

This is even after Raja said the Water cooled edition (which they did not demo, that was air cooled FE) will have higher clocks...

So there's other VEGA SKUs that has higher performance in games than the Vega FE demos.
 
fingerbob69 said:
Elmnator: Will the consumer RX version be as fast at the Frontier version?
RK: Consumer RX will be much better optimized for all the top gaming titles and flavors of RX Vega will actually be faster than Frontier version!
Click to expand...

Bahanime said:
This is even after Raja said the Water cooled edition (which they did not demo, that was air cooled FE) will have higher clocks...

So there's other VEGA SKUs that has higher performance in games than the Vega FE demos.
Click to expand...



Yes in games consumer RX will be faster than the Vega FE. Not hard to understand is it? Different drivers for the two cards Vega FE will use pro drivers when consumer Vega will use drivers optimized for games lol

How hard it is to understand what Raja stated, its pretty black and white.

RK: The Frontier Edition was designed for a variety of use-cases like Machine Learning, real-time visualization, and game design. Can you play games on Frontier Edition? Yes, absolutely. It supports the RX driver and will deliver smooth 4K gaming. But because it is optimized for professional use cases (and priced accordingly), if gaming is your primary reason for buying a GPU, I’d suggest waiting just a little while longer for the lower-priced, gaming-optimized Radeon RX Vega graphics card
Click to expand...

Two separate drivers one for pro and one for games oriented, This is like Fire GL and Quardro cards simple.

God he stated it twice.

RK: RX will be fully optimized gaming drivers, as well as a few other goodies that I can’t tell you about just yet….But you will like FE too if you can’t wait:)
Click to expand...

Right after this quote about the RX drivers he talked about the compute performance for the Vega FE

RK: On the compute side of things..Vega FE will be the fastest single GPU solution (>12.5 TFlops FP32) when it’s available and our NCU packs several additional optimizations, including Rapid-Packed-Math which delivers >25 TFLops of FP16
Click to expand...
 
Last edited:
Dudes, these are video cards and companies, nothing here is life threatening or dangerous. Damn, lets get over ourselves and actually take the time to purchase what you cannot in most other areas of the world.
 
fingerbob69 said:
Fud doubles down
http://www.fudzilla.com/news/graphics/43705-intel-is-making-radeon-powered-cpu
Click to expand...

Really, I could see the possibility of an AMD license to cover Intels ass in the GPU wars. To me there was an almost even chance that Intel would either A) Licence Patents from NVidia again, B) Licence Patents from AMD, C) Figure they have differentiated their design and key patents have expired, so they can ignore A) and B).

But Radeon IGPs inside Intel future CPU packages?

AMD would be pretty much giving up the only advantage their soon to be released, and finally competitive APUs with Zen/Vega internals.

That part of the rumor never made any sense, it would be like racing a marathon behind one of your competitors and catching him right near the finish line, only to give him a nice shove across it, to make sure he finishes first.

I also don't think Intel is concerned enough about the IGP to want to adopt AMDs design.

Intels CPUs with AMD IGP is something that has always seemed absurd.
 
razor1 said:
Yes in games consumer RX will be faster than the Vega FE. Not hard to understand is it? Different drivers for the two cards Vega FE will use pro drivers when consumer Vega will use drivers optimized for games lol

How hard it is to understand what Raja stated, its pretty black and white.



Two separate drivers one for pro and one for games oriented, This is like Fire GL and Quardro cards simple.

God he stated it twice.



Right after this quote about the RX drivers he talked about the compute performance for the Vega FE
Click to expand...



I was responding to your assertion that

"Frontier edition, will be the fastest compute output of all the Vega's. That puts gaming vega's at less clocks or are cut down versions (I think its the prior as they have shown those types of cards to public already)"

So wind your patronising neck back in and go shove the attitude where the sun don't shine....


....please.
 
fingerbob69 said:
I was responding to your assertion that

"Frontier edition, will be the fastest compute output of all the Vega's. That puts gaming vega's at less clocks or are cut down versions (I think its the prior as they have shown those types of cards to public already)"

So wind your patronising neck back in and go shove the attitude where the sun don't shine....


....please.
Click to expand...

if compute performance is the fastest for single GPU cards, the Vega FE which is >25 fp 16 flops and >12.5 fp 32 flops, all other cards have be less that what ever that >25 fp 16 flops and >12.5 fp 32 flops.

We are not talking about application performance directly, he was talking about flops and mentioned flops.

RK: On the compute side of things..Vega FE will be the fastest single GPU solution (>12.5 TFlops FP32) when it’s available and our NCU packs several additional optimizations, including Rapid-Packed-Math which delivers >25 TFLops of FP16
Click to expand...

So how do you get other cards less than the >25 fp 16 flops and >12.5 fp 32 flops figures? LESS clocks or CUT cores. Right? Only two explanations.
 
You were talking clocks. "less clocks".

The Frontier Edition clocks at 1600.

You stated RX Vega will be "at less clocks" compared to the Frontier.

Raja stated RX Vega will be faster than Frontier.
 
I am getting an informed answer together now. Funny the fool did big stories on rumors then days don't trade on those. Lol
 
fingerbob69 said:
You were talking clocks. "less clocks".

The Frontier Edition clocks at 1600.

You stated RX Vega will be "at less clocks" compared to the Frontier.

Raja stated RX Vega will be faster than Frontier.
Click to expand...


Raja never stated clocks will be faster on the consumer Vega vs Frontier dude.

He stated gaming performance will be faster due to drivers not clocks

RK: Consumer RX will be much better optimized for all the top gaming titles and flavors of RX Vega will actually be faster than Frontier version!
Click to expand...

What are they optimizing? The card? NO its the driver.
 
So in comparing Pro cards to Consumer versions the only optimisation Raja's talking about are driver optimisations? That's your interpritation, no fact.

But let's suppose you're right. The only change from Pro to Consumer is the driver ...then RX Vega clocked at 1600 isn't going to be any slouch!
 
fingerbob69 said:
So in comparing Pro cards to Consumer versions the only optimisation Raja's talking about are driver optimisations? That's your interpritation, no fact.

But let's suppose you're right. The only change from Pro to Consumer is the driver ...then RX Vega clocked at 1600 isn't going to be any slouch!
Click to expand...


They are the same chip, that's how AMD and nV been making them for years lol, its just driver differences.
 
fingerbob69 said:
I was responding to your assertion that

"Frontier edition, will be the fastest compute output of all the Vega's. That puts gaming vega's at less clocks or are cut down versions (I think its the prior as they have shown those types of cards to public already)"

So wind your patronising neck back in and go shove the attitude where the sun don't shine....


....please.
Click to expand...

Pro Drivers. RX Vega can't even run some pro apps, and certainly will fail in machine learning apps without driver support. So it's a rather iffy statement.
 
Either way AMD's stock took another hit. Got nothing to prop up the stock until end of June. Yeah they will reveal Vega at Computex but release won't be til end of June next to Epyc.
 
OKThen said:
Licensing rumor just denied by Lisa Su: http://www.marketwatch.com/story/am...ense-deal-2017-05-22?siteid=yhoof2&yptr=yahoo
So now we have Intel and AMD denying this deal, but "you have it on good authority it's a done deal"?
Click to expand...

All that statement says is that AMD understands what advantages they have, and aren't stupid enough to give them away.

In my opinion, this does not preclude a third-party asking for a specific product combining both Intel and AMD tech.

Edit: for that 3rd party's exclusive use.
 
You must log in or register to reply here.
Back
Top