Gamers Nexus: HBM2 Costs Estimated... Is AMD Profiting much on Vega?

[DocSwag]

Source: http://www.gamersnexus.net/guides/3032-vega-56-cost-of-hbm2-and-necessity-to-use-it

Quote

Variations of “HBM2 is expensive” have floated the web since well before Vega’s launch – since Fiji, really, with the first wave of HBM – without many concrete numbers on that expression. AMD isn’t just using HBM2 because it’s “shiny” and sounds good in marketing, but because Vega architecture is bandwidth starved to a point of HBM being necessary. That’s an expensive necessity, unfortunately, and chews away at margins, but AMD really had no choice in the matter. The company’s standalone MSRP structure for Vega 56 positions it competitively with the GTX 1070, carrying comparable performance, memory capacity, and target retail price, assuming things calm down for the entire GPU market at some point. Given HBM2’s higher cost and Vega 56’s bigger die, that leaves little room for AMD to profit when compared to GDDR5 solutions. That’s what we’re exploring today, alongside why AMD had to use HBM2.

Quote

AMD’s pricing structure for Vega uniquely leans on bundle packs to help improve the company’s value argument in a competitive market. MSRP is $400 on RX Vega 56, $500 on RX Vega 64, and an added $100 upcharge in exchange for two games and some instant discounts. AMD’s intention with this is to offer greater value to gamers, but clearly also will help the company increase margins and move more Ryzen parts, thereby recouping potentially low or negative margins on Vega. This is aided particularly with game bundles, where AIB partners pay AMD about $29 for the game codes, though that is often waived or offered in exchange for MDF. AMD also stated desire to stave off some mining purchases with increased bundle prices, as this would offset the value proposition of the card. Since the bundles are sold as standalone SKUs and can’t be broken by consumers (into parts), it seems that this is potentially an effective solution at keeping miners at bay.

Quote

There are two major costs with a video card: The GPU die and the memory, with follow-up costs comprised of the VRM and, to a lesser extent, the cooler.

 

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. It’d cost a “normal person,” for instance, about $45 to build the VRM on Vega – that’d include the $2.70 per-phase cost of the IRF6894s and IRF6811 hi- and lo-side DirectFETs, about $8.80 for all six of the IR3598 drivers, and roughly $4 on the IR35217 (from public sellers and datasheets). AMD is a large company and would receive volume discounts. Even as individuals, we could order 10,000 of these parts and drive that cost down, so these numbers are strictly to give an idea of what it’d cost you to build the VRM.

Quote

We did speak with numerous folks, including Kanter, on estimated GPU die pricing, but the estimates had a massive range and were ultimately just educated guesses. Without something more concrete to work with, we’re going to just stick to HBM2 and interposer pricing, as that’s the figure we know. GPU cost and yield cost are only really known by AMD and GlobalFoundries, at this point, so no point in working with total speculation.

Quote

The next question is what GDDR5 costs. A recent DigiTimes report pegs GDDR5 at about $6.50 for an 8Gb module, though also shows pricing for August onward at $8.50 per module. With old pricing, that’s around $52 cost for an 8GB card, or $68 with new pricing. We do not presently know GDDR5X cost. This puts us at around 3x the cost for HBM2 which, even without factoring in yields or the large GPU die, shows why AMD’s margins are so thin on Vega. We also know that AMD is passing along its HBM2 cost to partners at roughly a 1:1 rate – they’re not upcharging it, which is what typically happens with GDDR. There’s no room to upcharge the HBM2 with Vega’s price target.

Ignoring GPU cost and cost of less significant components, like the VRM and cooler, we’re at $100-$130 more than 8GB of GDDR5 cost to build. This is also ignoring other costs, like incalculable R&D or packaging costs. Again: We’re just focusing on memory today.

Quote

There’s more to it, though. HBM2 critically allows AMD to run lower power consumption than GDDR5 would enable, given the Vega architecture.

Speaking with Buildzoid, we know that Vega: Frontier Edition’s 16GB HBM2 pulls 20W max, using a DMM to determine this consumption. This ignores the voltage controller’s 3.3v draw, but we’re still at 20W memory, and no more than an additional 10W for the controller – that’s less than 30W for the entire memory system on Vega: Frontier Edition.

 

We also know that an RX 480 uses 40-50W for its 8GB, which is already a significant increase in power consumption per-GB over Vega: FE. The RX 480 also has a memory bandwidth of 256GB/s with 8GB GDDR5, versus Vega 64’s 484GB/s. The result is increased bandwidth, the same capacity, and lower power consumption, but at higher cost to build. In order for an RX 480 to hypothetically reach similar bandwidth, power consumption would increase significantly. Buildzoid calculates that a hypothetical 384-bit GDDR5 bus on Polaris architecture would push 60-75W, and an imaginary 512-bit bus would do 80-100W. For this reason alone, HBM2 saves AMD from high power budget that would otherwise be spent solely on memory. This comes down to architectural decisions made years ago by AMD, which are most readily solved for with HBM2, as HBM2 provides greater bandwidth per watt than GDDR5. HBM is effectively a necessity to make Vega at least somewhat power efficient while keeping the higher memory bandwidth. Imagine Vega 56, 64, or FE drawing an additional 70-100W – the world wouldn’t have it, and it’d be among the hottest cards since the GTX 480 or R9 290X.

 

The Vega architecture is clearly starved by memory bandwidth, too: Overclocking HBM2 alone shows this, as its gains are greater than just core clock increases. AMD didn’t have another choice but to go with HBM2, even though costs would be roughly one-third on the memory. GDDR5 might be possible, but not without blowing power consumption through the roof or losing on performance by limiting bandwidth.

So basically, HBM2 alone on Vega is costing about $150, with the interposer costing another $25 which pegs the cost of the memory system in Vega at $175. This is really, really high, especially considering that on Vega 56 this would comprise nearly HALF the cost of the GPU. This also probably explains AMD's choice to go for 2 HBM stacks instead of 4 on Vega; 4 probably would've made the memory alone cost almost $400. The VRM setup probably costs another $30-$40 ish, which would make the cost of the video card excluding the gpu already a little over $200. Say the GPU die costs $75, and AMD really isn't making a huge profit on 56 or 64.

 

However, AMD really had no choice and had to go with HBM. HBM leads to lower power consumption which in turn lets AMD push clock speeds a bit higher, and when AMD is at a significant performance/watt deficit compared to Nvidia, they need everything they can get to catch up.

 

At the same time, it has to be questioned whether this really was worth doing considering that going HBM2 costs AMD approximately three times as much compared to GDDR5. Is it really worth paying $100 more just to get 20W more headroom which may lead to 4-5% more performance? 

[Drak3]

God damn, that's a pretty penny. Hopefully, RTG learned something from the whole ordeal, and will launch GDDR and HBM variations of NAVI.

[DrMacintosh]

Well if you watched the Gamers Nexus video on this you would know the answer to that. 

[LeapFrogMasterRace]

Just now, DrMacintosh said:

Well if you watched the Gamers Nexus video on this you would know the answer to that. 

Please be nice

[marldorthegreat]

13 minutes ago, Drak3 said:

God damn, that's a pretty penny. Hopefully, RTG learned something from the whole ordeal, and will launch GDDR and HBM variations of NAVI.

Navi hopefully will be like epyc. They design 1 smaller chip, with roughly the same performance as the rx 580, at 80-100 watts. (rumored to be on 7nm) And "Glue" the dies together on a substrate. Which would allow them to make a gpu like vega without running into issues with large gpu manufacturing. It could mean that we see for the first time since the HD 7000 series a full lineup of NEW gpus.

[Drak3]

Just now, marldorthegreat said:

Navi hopefully will be like epyc. They design 1 smaller chip, with roughly the same performance as the rx 580, at 80 watts. (rumored to be on 7nm) And "Glue" the dies together on a substrate. Which would allow them to make a gpu like vega without running into issues with large gpu manufacturing. 

We already know that Navi is going to be an MCM at the top. The only forseeable issue with Navi as of right now, is that it requires too many watts at stock, cost as much as the Volta equivalent, and doesn't outperform it by any meaningful margin.

 

I hope that isn't the case, but it is a possibility.

[Bcat00]

Navi 600watts ahoy!!!!!

[Taf the Ghost]

Answer: No.  That + drivers were likely the biggest reason for the delay. But AMD will sell you an awesome Workstation card for a lot more money.

[ScratchCat]

This puts Vegas power efficiency Even lower than before. Given that the 1080 has a TDP of 180w and a memory TDP of about 50W that gives a GPU TDP of 130W. Vega has a TDP of 350W for the water cooled version and with a memory TDP of 20W a power consumption for te GPU of 330W!

TDP ~ power consumption , goo enough for a comparison 

[Terryv]

5 hours ago, Bcat00 said:

Navi 600watts ahoy!!!!!

Not likely. As die size goes down, power efficiency goes up. Same for clock speeds. As it stands, Vega is pushed way past peak efficiency to achieve the performance they need.

 

Navi running multiple smaller die's (I'd say the sweet spot would be 2048 SP's per die IMO) coupled with more reasonable clock speeds could be very power efficient. This doesn't take into account any potential die shrinks either.

[Coaxialgamer]

2 hours ago, ScratchCat said:

This puts Vegas power efficiency Even lower than before. Given that the 1080 has a TDP of 180w and a memory TDP of about 50W that gives a GPU TDP of 130W. Vega has a TDP of 350W for the water cooled version and with a memory TDP of 20W a power consumption for te GPU of 330W!

TDP ~ power consumption , goo enough for a comparison 

There's no mention of GDDR5X at all. Consensus is that the 1080 memory pulls 20w minus the memory subsystem. 

 

Also, it doesn't work like that. You have a power budget, and you balance it between memory and gpu. It's an integral part of power efficiency. 

[Trixanity]

8 hours ago, DocSwag said:

At the same time, it has to be questioned whether this really was worth doing considering that going HBM2 costs AMD approximately three times as much compared to GDDR5. Is it really worth paying $100 more just to get 20W more headroom which may lead to 4-5% more performance? 

As your own quote says: Vega is bottlenecked by memory, so reducing bandwidth by going GDDR would have made it perform been significantly worse.

 

What AMD needed was higher clocked HBM but suppliers failed to deliver. The only other option would have been increasing costs by going 4 stacks instead of 2 and therefore doubling bandwidth. It would probably have increased performance significantly but would (I suppose) add another $150.

[Coaxialgamer]

17 minutes ago, Terryv said:

Not likely. As die size goes down, power efficiency goes up. Same for clock speeds. As it stands, Vega is pushed way past peak efficiency to achieve the performance they need.

 

Navi running multiple smaller die's (I'd say the sweet spot would be 2048 SP's per die IMO) coupled with more reasonable clock speeds could be very power efficient. This doesn't take into account any potential die shrinks either.

AMD needs those clock speeds. They can't hope to get anywhere close to nvidia without them. Lower clocks require more IPC for a given amount of performance, which requires more hardware, thus silicon area. MCM's can help cost, but they aren't magic. 

For a long time, I'd argued that vega is a necessary step. GCN in the form of polaris ultimately cannot compete against nvidia. 

[tsk]

AMD bet on HBM too soon with Fiji and probably had no point of return with Vega, an expensive mistake indeed.

[Terryv]

8 minutes ago, Coaxialgamer said:

AMD needs those clock speeds. They can't hope to get anywhere close to nvidia without them. Lower clocks require more IPC for a given amount of performance, which requires more hardware, thus silicon area. MCM's can help cost, but they aren't magic. 

For a long time, I'd argued that vega is a necessary step. GCN in the form of polaris ultimately cannot compete against nvidia. 

I agree that vega is a transition. As AMD add modules in their future MCM design, clock speed becomes less important. They can just add modules to get whatever performance they need and keep clock speeds in their best efficiency range.

[Dabombinable]

Considering Vega is memory starved, wouldn't it have been better to go with 4x smaller HBM2 modules as opposed to 2? That way Vega would have the wide memory bus of Fiji, along with HBM2's bandwidth improvements.

[Coaxialgamer]

3 minutes ago, Dabombinable said:

Considering Vega is memory starved, wouldn't it have been better to go with 4x smaller HBM2 modules as opposed to 2? That way Vega would have the wide memory bus of Fiji, along with HBM2's bandwidth improvements.

What i think happened was that AMD 's was planning on using SK-HYNIX' s 1ghz hbm2 modules, which were delayed. 

Using 4 stacks of hbm2 would have likely been too expensive anyway, given the significant price of 2*4hi hbm2 foubd of vega. 

[Agosto]

11 minutes ago, Terryv said:

I agree that vega is a transition. As AMD add modules in their future MCM design, clock speed becomes less important. They can just add modules to get whatever performance they need and keep clock speeds in their best efficiency range.

Not unless they fix SP scaling in the architecture. Look at Fiji and Vega, 4096 cores SKUs should have about 14% higher performance clock for clock than 3584 ones, but the difference is actually much smaller (something like 5%)

[leadeater]

26 minutes ago, Dabombinable said:

Considering Vega is memory starved, wouldn't it have been better to go with 4x smaller HBM2 modules as opposed to 2? That way Vega would have the wide memory bus of Fiji, along with HBM2's bandwidth improvements.

 

20 minutes ago, Coaxialgamer said:

What i think happened was that AMD 's was planning on using SK-HYNIX' s 1ghz hbm2 modules, which were delayed. 

Using 4 stacks of hbm2 would have likely been too expensive anyway, given the significant price of 2*4hi hbm2 foubd of vega. 

That does bring up the question could they not have used 4 HBM1 modules? Fiji actually has more raw memory bandwidth than Vega, is there anything special about HBM2 other than higher bandwidth per module? Is 4 HBM1 stacks of today cheaper than 2 HBM2 stacks?

[Agosto]

10 minutes ago, leadeater said:

 

That does bring up the question could they not have used 4 HBM1 modules? Fiji actually has more raw memory bandwidth than Vega, is there anything special about HBM2 other than higher bandwidth per module? Is 4 HBM1 stacks of today cheaper than 2 HBM2 stacks?

HBM is limited at 1GB per stack.

[Terryv]

25 minutes ago, Agost said:

Not unless they fix SP scaling in the architecture. Look at Fiji and Vega, 4096 cores SKUs should have about 14% higher performance clock for clock than 3584 ones, but the difference is actually much smaller (something like 5%)

I don't thing there's a scaling issue in the arch. Like others have said, vega 64 is likely starved for memory bandwidth. If they fix that with 4 modules instead of 2 they instantly double memory bandwidth.

[leadeater]

7 minutes ago, Agost said:

HBM is limited at 1GB per stack.

Well then 8 of them is the perfect solution, I'm a genius! .

 

RIP AMD's wallet and any PCB space saving, that would be a monster sized GPU package lol.

[Prysin]

9 hours ago, DocSwag said:

Source: http://www.gamersnexus.net/guides/3032-vega-56-cost-of-hbm2-and-necessity-to-use-it

So basically, HBM2 alone on Vega is costing about $150, with the interposer costing another $25 which pegs the cost of the memory system in Vega at $175. This is really, really high, especially considering that on Vega 56 this would comprise nearly HALF the cost of the GPU. This also probably explains AMD's choice to go for 2 HBM stacks instead of 4 on Vega; 4 probably would've made the memory alone cost almost $400. The VRM setup probably costs another $30-$40 ish, which would make the cost of the video card excluding the gpu already a little over $200. Say the GPU die costs $75, and AMD really isn't making a huge profit on 56 or 64.

 

However, AMD really had no choice and had to go with HBM. HBM leads to lower power consumption which in turn lets AMD push clock speeds a bit higher, and when AMD is at a significant performance/watt deficit compared to Nvidia, they need everything they can get to catch up.

 

At the same time, it has to be questioned whether this really was worth doing considering that going HBM2 costs AMD approximately three times as much compared to GDDR5. Is it really worth paying $100 more just to get 20W more headroom which may lead to 4-5% more performance? 

meanwhile Nvidia pays around 12-15$ per GDDR5X 1GB stack on the 1080Ti...

[Terryv]

2 minutes ago, Prysin said:

meanwhile Nvidia pays around 12-15$ per GDDR5X 1GB stack on the 1080Ti...

I believe that's the price of regular GDDR5. I'm not sure the price of the X is known

[Coaxialgamer]

22 minutes ago, leadeater said:

 

That does bring up the question could they not have used 4 HBM1 modules? Fiji actually has more raw memory bandwidth than Vega, is there anything special about HBM2 other than higher bandwidth per module? Is 4 HBM1 stacks of today cheaper than 2 HBM2 stacks?

HBM1 would have limited them to 4GB VRAM maximum.