*UPDATED* Radeon VII's HBM 2 Costs Almost Half the GPU's Price

[CarlBar]

6 minutes ago, Stefan Payne said:

Exactly

that's what I meant

 

Sorry you went right over my head first time.

[YoloSwag]

7 hours ago, Stefan Payne said:

It would be pointless as its basically a Standard RX Vega...

Wouldn't say it's pointless though. Since they compared it to an RTX 2080 which had similar performance to a GTX 1080 Ti and the RX VEGA 64 did not compete against the GTX 1080 Ti but instead against the non-Ti in which it didn't win much against. And I doubt the extra 8GB of HBM2 would increase the performance to GTX 1080 Ti levels.

 

What I was saying is that if they cut the memory by half, they are effectively cutting the cost of the card by a significant amount making it more attractive to buy. The R9 390X was definitely the same as had slightly better performance than an R9 290X although they are not the same architecture (but they do have similar core and transistor count, TMUs, etc) and the increased memory made it more attractive but I don't think it's the same case with Radeon VII and RX VEGA 64 as we do see a good increase in performance at least in AMD's slides from not just a node shrink but having more transistors (13.2bn vs 12.5bn) despite having lesser cores (3,840 vs 4,096).

 

7 hours ago, Stefan Payne said:

RX590 is 12nm, wich is basically an improved 14nm process with some slight changes, wich allows for higher clockrates and/or better power consumption.

 

Only Radeon 7 is 7nm.

Yeah, sorry about that. Will update the post thanks.

[YoloSwag]

5 hours ago, huilun02 said:

As such it is now established as fact that the memory of the Radeon VII is half the cost of the product

Despite having no official info from AMD

They're not gonna answer anything regarding actual production costs if that's what you wanted to know.

 

The purpose of investigations done by other publications was to reveal shadows of objects and not the actual object.

I understand that the title sounds a bit misleading but the research done by other publications aren't pure speculation. Again we're only gonna see shadows unless AMD itself sheds light on the subject.

[Doobeedoo]

Hm kinda odd though, HBM2 and that tech in general is supposed to be lower in price than initially show. But we know they're repurposed Instict cards to show something new to more close the gap until new flagship. 

[CarlBar]

1 hour ago, Doobeedoo said:

Hm kinda odd though, HBM2 and that tech in general is supposed to be lower in price than initially show. But we know they're repurposed Instict cards to show something new to more close the gap until new flagship. 

 

To be fair AMD are just seriously ahead of the curve ATM with HBM. Give it another few memory generations and standard GDDR just isn't going to be able to scale capacity wise without some new innovation. If and more probably when that happens, HBM or a similar technology is going to be the only option to keep scaling vram capacity.

[Doobeedoo]

2 minutes ago, CarlBar said:

 

To be fair AMD are just seriously ahead of the curve ATM with HBM. Give it another few memory generations and standard GDDR just isn't going to be able to scale capacity wise without some new innovation. If and more probably when that happens, HBM or a similar technology is going to be the only option to keep scaling vram capacity.

That's true, price aside, it is a better memory tech.

[Stefan Payne]

2 hours ago, YoloSwag said:

Wouldn't say it's pointless though.

I would as that would essentially be a VEGA 56, shrinked into 7nm.

How much of the Performance is due to memory bandwith and how much due to other improvements is not (yet) known.

But my guess would be that most of it is due to 4096bit...

 

2 hours ago, YoloSwag said:

Since they compared it to an RTX 2080 which had similar performance to a GTX 1080 Ti and the RX VEGA 64 did not compete against the GTX 1080 Ti

Well, if I would, I could show you that it is possible for the VEGA 64 to compete against the 1080ti

I can name 2 games:
Forza Horizon 4 (both beat 1080ti, 56 more or less, 64 without any doubt) 

Ni No Kuni 2 (VEGA 64 beats 1080ti slightly)

Source: GameGPU.ru

 

Both run exceptionally on AMD Hardware (for obvious reasons)...

2 hours ago, YoloSwag said:

but instead against the non-Ti in which it didn't win much against.

Depends on the game and how gameworks infested or incompetent the (engine) developer is....

So that opens a vast variety for the performance, from way 

 

2 hours ago, YoloSwag said:

And I doubt the extra 8GB of HBM2 would increase the performance to GTX 1080 Ti levels.

Its not the 8GiB, its the additional 2048bit Memory and ~500GiB/sec Bandwith.

 

2 hours ago, YoloSwag said:

What I was saying is that if they cut the memory by half, they are effectively cutting the cost of the card by a significant amount making it more attractive to buy.

Yes, but two Problems:

a) halve the bandwith, wich means loss in performance (and probably all the advantages of VEGA 7 over 56)

b) new supply chain for 2hi HBM Modules

 

Both doesn't make too much sense. a) defeats the Purpose of Radeon 7 and b) causes unnecessary stock of parts you don't need or want. 

And you still have the same problems with manufacturing than 4hi...

So what if the most significant cost of the manufacturing is the assembly of all 6 Dies to one Chip and on one Package??

 

And the Difference between 2hi and 4hi negligable (because nobody makes 2Hi/half Capacity Stacks).

See the Problem with that Statement?

Thing is AMD doesn't want to source another, pretty useless DRAM for a product with an expected lifetime of 3-6 Months...

At Worst a Year...

 

But its said that Navi is performing better than expected and cost like 350€ or so for the higher end model.

2 hours ago, YoloSwag said:

The R9 390X was definitely the same as had slightly better performance than an R9 290X although they are not the same architecture (but they do have similar core and transistor count, TMUs, etc) and the increased memory made it more attractive but I don't think it's the same case with Radeon VII and RX VEGA 64 as we do see a good increase in performance at least in AMD's slides from not just a node shrink but having more transistors (13.2bn vs 12.5bn) despite having lesser cores (3,840 vs 4,096).

The 290X and 390X are literally the same chip, only difference beeing some Optimizations on the chip, so that it allows for higher frequency. 

The Hawaii XT had 1GHz Core and 1250MHz Memory.

The Grenada XT had 1050MHz Core and 1500MHz Memory.

 

Its similar to RX480 -> 580.

2 hours ago, YoloSwag said:

Yeah, sorry about that. Will update the post thanks.

Nothing to be sorry about, mistakes happens

[cj09beira]

11 hours ago, leadeater said:

And Nvidia's memory compression is much better than AMD's so they need less of it comparatively.

its been quite a bit since nvidia made a push to better memory/color compression, i question if they haven't done it because it would affect their medical applications where color quality is very important.

30 minutes ago, Doobeedoo said:

That's true, price aside, it is a better memory tech.

hbm needs cheap products using it to allow for higher volume, for now its hot stuff used by expensive gear=high price.

one day  

[Doobeedoo]

52 minutes ago, cj09beira said:

hbm needs cheap products using it to allow for higher volume, for now its hot stuff used by expensive gear=high price.

one day  

Well yeah, really currently mainly can benefit flagships so it can make sense to use it for them.

[CarlBar]

2 hours ago, cj09beira said:

its been quite a bit since nvidia made a push to better memory/color compression, i question if they haven't done it because it would affect their medical applications where color quality is very important.

hbm needs cheap products using it to allow for higher volume, for now its hot stuff used by expensive gear=high price.

one day  

 

The real issue is that only AMD is using it in the graphics card space. Get NVIDIA in thee and that will change fast as volumes needed go up and production ramps to meet demand.

[YoloSwag]

16 hours ago, Stefan Payne said:

-snip-

Yeah was gonna say that as you saw the strikethroughs. The R9 290X and R9 390X are the same but some might argue they are different just because of the codenames and optimizations. I think it was all microcode or something else though since the rest of the specs of the chip didn't change.

 

Also, I am looking forward to Navi. Like I said in other threads and here, I think AMD's winner would always be in the mid-range segment and hopefully this Navi chip does well.

[JoostinOnline]

On 1/26/2019 at 5:58 PM, firelighter487 said:

...why use that memory then?

I don't know if it was true, but the explanation I heard (I think by Linus actually) was that HBM2 had the potential to be cheaper than GDDR5 back when cryptomining was huge. Unfortunately, because it's a niche product the price won't go down unless other companies start putting HBM2 in their products as well. Less demand doesn't necessarily mean lower prices.

 

I think a more accurate question is why include so much memory in Radeon VII? It's being marketed as a mainstream gaming card, but nothing I'm aware of comes close to using that much memory.

[JoostinOnline]

22 hours ago, CarlBar said:

 

Highly unlikely. NAVI should be here by then.

Source? Navi keeps getting delayed and I didn't think there was anything but speculation on its release.

 

It's painful seeing their GPU side fail so hard. They're killing Intel on the CPU side but NVIDIA has gotten Intel-level lazy since AMD stopped offering any competitive GPU's.

 

[Taf the Ghost]

From everyone's favorite benchmark parser:

 

Quote

AMD Radeon VII

Graphics Score

FS - 27400

FSE - 13400

FSU - 6800

TS - 8700

Pulling some launch review numbers of the RX Vega 64. (Using Guru3D here.) 

 

RX Vega 64:

FireStrike: 19200

FireStrike Extreme: 9609 (From Bit-Tech)

FireStrike Ultra: 5415

TimeSpy: 7243

 

Depending on the task type, it's between 20-40% faster than a RX Vega 64. At least these testing numbers. We'll see what the official spec versions come in at.

[Stefan Payne]

...with +50% Bandwith compared to VEGA 64, though the Comparisation with VEGA 56 is more interesting...

[Taf the Ghost]

1 hour ago, Stefan Payne said:

...with +50% Bandwith compared to VEGA 64, though the Comparisation with VEGA 56 is more interesting...

Main comparison is going to be "what did they fix?" and "how much better is the node?". GloFo's 14nm wasn't a good high-end GPU node. Great server & mainstream GPU node, though, which is why GloFo is making a lot of money off the node.