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

[YoloSwag]

As reported previously AMD will be launching the Radeon VII that comes with 16GB of HBM 2 and still uses the VEGA architecture but in 7nm -

Reported by Fudzilla -

Quote

Currently HBM 2 chips cost  around #80 for 4GB. It is very hard for us to estimate the exact price that AMD is paying to its supplier but with this in mind 16 HBM 2 could be as expensive as $320, very close to half of the whole cost. This will be close to what AMD is paying.

^{The "#80" is obviously an error which should have been '$80'}

 

If you think this $80 is overestimating it, GamersNexus did make a report about it last year but using the RX VEGA's in context -

Quote

Let’s start with HBM2 and interposer pricing, as that’s what we’re most confident in. Speaking with David Kanter of Real World Tech, the analyst who broke news on Maxwell’s tile-based rasterization and who previously worked at Microprocessor Report, we received the following estimate: “The HBM2 memory is probably around $150, and the interposer and packaging should be $25.” We later compared this estimate with early rumors of HBM2 pricing and word from four vendors who spoke with GamersNexus independently, all of which were within $5-$10 of each other and Kanter’s estimate. This gave us high confidence in the numbers. Taking his $175 combined HBM2 + interposer figure, we’re nearly half-way to the MSRP of the Vega 56 card, with the rest of costs comprised of the VRM, GPU, and dime-a-dozen electrical components.

For reference, the $175 he's talking about here is for 2x4GB stacks of HBM2. Also, the RX VEGA's had the same issue in which the HBM2 on the card costs almost half the GPU's price.

 

So with this, we could say the HBM2 implementation on the Radeon VII could be costing them around $320-$325($150x2 + $25) but depending on underlying agreements between manufacturers this could be a bit lower.

This is still expensive considering the whole card itself would cost around $699 not taking into account demand as Radeon VII was speculated to be launching in fewer numbers (5,000 units). Not new really as RX VEGA had a similar issue.

 

In relation to this, 3dcenter made an article recently around the cost of GDDR6. Please note that they used a ',' instead of a '.' to indicate decimals.

 

Using this data we could say that 16GB of GDDR6 could cost around $187.04 without talking into account implementation costs (which could be a lot more since you would need to take into account the space it takes up on the PCB).

 

IMHO - I don't know about you but this seems a bit cheaper than HBM2. Of course others will argue that Radeon VII was targeted for compute etc. well it was pitted against the RTX 2080 in AMD's CES presentation so...

For me they're still gonna market this as a high-end gaming card and of course it offers better compute than Nvidia GPUs but come on.

 

How much of a performance hit difference it would be to replace HBM2 with GDDR6 we do not know for sure.

With this said I am kind of thinking there could have been some sort of anti-competitive practices going on here for example GDDR6 was available last year but only Nvidia managed to get GPUs with this. Same with AMD's RX VEGA being the only mainstream GPUs (I don't consider Nvidia's Titan V a mainstream GPU) to get HBM2. EDIT: I meant that these memory chips aren't exclusive, but production-wise there could have been reservations for example SKHynix could only produce 10,000 units but AMD issues buy orders for 12,000 units. Something like that, I mean not really anti-competitive but sort of.

 

But what do you guys think? Tell us in the comments below. (lol)

 

UPDATE EDIT:

 

New thought -

WHAT IF AMD LAUNCHED A RADEON VII WITH ONLY 8GB of HBM2!?

 

- With what I said in the top part, HBM2 costs $80 per 4GB stack. By launching an 8GB model, shouldn't this make the card $100-180 cheaper!?

a $519-$599 model could sell like hotcakes. Who needs 16GB of HBM2 to play Fortnite anyway? (Obviously a joke but you get the idea)

Edited January 27, 2019 by YoloSwag
NEW IDEA

[DrMacintosh]

Yep. Rumor has it that Navi will use traditional memory thankfully. Meaning it will actually make money. 

[Ashleyyyy]

...why use that memory then?

[ZestyJalapeno]

AMD come on man, use that HBM production capacity to refresh 56 and 64 on 12nm so I can finally upgrade my GPU and not have to deal with this stupid farming malware  GeForce Experience every time I start up a game

[YoloSwag]

1 minute ago, firelighter487 said:

...why use that memory then?

This was a question Raja Koduri (when he was at AMD) failed to answer when he did a Q&A in reddit for the RX VEGA launch.

3 minutes ago, DrMacintosh said:

Yep. Rumor has it that Navi will use traditional memory thankfully. Meaning it will actually make money. 

I literally can't wait for Navi and GDDR6. Finally we get to see a Polaris replacement.

[Drak3]

4 minutes ago, firelighter487 said:

...why use that memory then?

Koduri might have started losing his touch when Vega was designed.

[PopReference]

Considering some rumours were saying the HBM would cost as much as or more (more then the Vega VII's announced price) coming out before the announcement of the card; really, the article could read as " AMD may actually sell vega 7 at above cost!" to be honest.

[leadeater]

1 minute ago, PopReference said:

" AMD may actually sell vega 7 at above cost!" to be honest.

Don't you mean below cost? Doesn't everyone sell above the cost to make?

[YoloSwag]

2 minutes ago, Drak3 said:

Koduri might have started losing his touch when Vega was designed.

VEGA wasn't his baby, so he was basically "that ain't my baby"

 

Then he left AMD...

 

NAVI was his but since he left, the momma (Lisa) was like "I'll raise this up just like I did the last one"

[AlwaysFSX]

9 minutes ago, firelighter487 said:

...why use that memory then?

It's lower power, you can imagine why that might be necessary.

[Drak3]

1 minute ago, YoloSwag said:

VEGA wasn't his baby

He had a hand in VEGA as well.

[Jurrunio]

Seeing that Radeon VII is basically a repurposed Radeon Instinct makes more sense, more of a respond to Nvidia than to make money, I'm not surprised. The full workstation card could easily cost companies a lot more even in bulk

[YoloSwag]

Just now, Drak3 said:

He had a hand in VEGA as well.

Yeah but it was basically set in stone. He was in the "raising this child as my own" part but it really wasn't his. Good man tho.

[PopReference]

6 minutes ago, leadeater said:

Don't you mean below cost? Doesn't everyone sell above the cost to make?

there were something going around saying the HBM2 would cost $750, that's just what I remember hearing, but that would be what I meant.

[Mihle]

11 minutes ago, leadeater said:

Don't you mean below cost? Doesn't everyone sell above the cost to make?

Rumor has it that Vega VII was more of a thing they did for "marketing" in the way that they didn't want it to go too long between GPU releases and then having something that had that performance.

Previously I have also heard they would sell at cost, if I remember right.

But that was rumors.

[Mihle]

10 minutes ago, PopReference said:

there were something going around saying the HBM2 would cost $750, that's just what I remember hearing, but that would be what I meant.

That's not correct, what someone said was that the hole card cost 750 $ to make. I think it was someone previously in AMD, but I would have to check that.

Maybe when I have time if I remember.

[GoldenLag]

21 minutes ago, PopReference said:

there were something going around saying the HBM2 would cost $750, that's just what I remember hearing, but that would be what I meant.

That was in reference to the cost of the whole card. Not just the memmory.

 

36 minutes ago, firelighter487 said:

...why use that memory then?

Power and memmory bandwidth. Those are the main reasons why.

[bcredeur97]

I think the VII is more of a showcase card of the new possibilities... 7nm, lots of HBM2 etc. they aren't intending to sell a lot or really make money from them I don't think. almost like a concept car... bunch of new ideas all crammed together.

then some of those ideas will be refined and will go into a final end product where the money making intentions actually are.

[KaitouX]

10 minutes ago, Mihle said:

That's not correct, what someone said was that the hole card cost 750 $ to make. I think it was someone previously in AMD, but I would have to check that.

Maybe when I have time if I remember.

Yes it was a rumor that it would cost 750 for the entire card, here is one of the sources: https://wccftech.com/exclusive-mike-rayfield-amd-retires/

Apparently some other AMD workers said similar things but it's all rumors.

Quote

He presented suggestions that werent really feasible such as ‘Radeon VII ‘ which was to be a Vega 20 based consumer facing part that cost $750 to build and would barely tie in with an [NVIDIA] GTX 1080 Ti.

At the point he made the suggestion AMD was probably expecting Turing to have similar performance improvements compared to Pascal, as Pascal had over Maxwell, so making a $700 card that would have the same performance as the supposed new $350 card would be stupid, but Nvidia Turing cards had basically the same performance as the Pascal ones which probably caused the Radeon VII to be a possibility to keep AMD name around until they can show Navi.

[mariushm]

If you put the memory chips so close to the gpu chip, you don't need a lot of power to send the data between chips, so instead of wasting let's say 20-40 watts in the memory controller (added for both chips), you could spend only 5-10w and use those 30 watts or so to increase the power budget of the whole chip. 

You also get very high bandwidth, each hbm2 chips is 1024 bit wide, so you get a 4096 bit interface to memory.  Compare that to GDDR memory where each chip is 32bit wide, so you'd need 8 chips to have 256 bit bus.

If a game or application is really designed all the way to work with 4KB transfers to optimize the throughput, there could be performance increases from using hbm memory.

You could also get smaller latency by placing the chips so close together.

 

You're all kind of forgetting that AMD also has some higher end cards which have option for local storage on the card, like you can attach a 1 TB SSD directly on the video card and that acts as extra buffer for the card, for applications that work with huge sets of data. So instead of constantly shuffling data through the pci-e slot, the card can simply move data from hbm to ssd on the card instead, or read directly from ssd at up to 3-4 GB/s as needed.

 

[Taf the Ghost]

46 minutes ago, Drak3 said:

He had a hand in VEGA as well.

Main thing is that Raja was in charge during the Production Stage of the architecture, and RTG was something of a wreck for that. That's the unsaid aspect of all of the problems with AMD's GPU division. It's one thing not to push too hard with a new design, but it's quite different when things are just outright busted on the silicon.

[schwellmo92]

Something to keep in mind is the HBM price estimates were from when flash memory was in short demand and prices inflated. Supply has caught up and prices have dropped quite a bit, mind you I’m not sure how much of the flash price drop makes it through to HBM.

[PopReference]

41 minutes ago, Mihle said:

That's not correct, what someone said was that the hole card cost 750 $ to make. I think it was someone previously in AMD, but I would have to check that.

Maybe when I have time if I remember.

 

30 minutes ago, GoldenLag said:

That was in reference to the cost of the whole card. Not just the memmory.

That makes more sense, the only specifics I recall is that the HBM by itself cost the most if not more then the amount mentioned in the OP.

[Taf the Ghost]

As a note, HBM wasn't the incorrect choice for AMD. The market and the use of HBM just ended up going in a different direction. FPGAs eat up most of the supply and pay top dollar. Notable, Nvidia hasn't rolled out anything but Volta using it.

[YoloSwag]

35 minutes ago, bcredeur97 said:

I think the VII is more of a showcase card of the new possibilities... 7nm, lots of HBM2 etc. they aren't intending to sell a lot or really make money from them I don't think. almost like a concept car... bunch of new ideas all crammed together.

then some of those ideas will be refined and will go into a final end product where the money making intentions actually are.

This is basically what was always at the back of my head but didn't know how to say. Thank you so much.

 

 

To all -

AMD doesn't have a high-end card. They've been launching proof-of-concept products in the high-end segment for a long time now.

The last good high-end stuff was the R9 290/290X (also the 390/390X but it's basically the same thing).

 

The most effort AMD has made so far was in Polaris. Soon Navi. Because this was the segment which gives them notoriety.

Having the best stuff at the mid-range and low-end market was always their strategy (Nvidia always had the "If we always win the high-end segment, we win in all segments" kind of strategy which does work).