TR Forums

Some gerbils may be interested in this -- I would suggest taking this with a heaping spoonful of salt though. No idea if it's legit or FUD. But here goes....

https://www.shacknews.com/article/10936 ... -at-a-loss

The sources said that AMD is simply repurposing Radeon Instinct MI50 cards for the Radeon VII launch and that the company will be losing money on each unit sold at its $699 MSRP.

I added the bold as it seems to be the most significant point.

"But we'll make up for it in volume!"

Sure, apart from the partially disabled die, fused off features, and totally new cooler, it's exactly the same as the MI50. It's not like they're just putting a sticker on the box, it's a different product.

Not sure how much I trust TweakTown, let alone their anonymous sources. How the hell would a supply chain member, even someone who works at an AIB vendor, know about AMD's internal profit margins?

Is AMD losing money if you just look at the BOM?
Probably not although they aren't getting rich either. For all the whining about the RTX 2080, if it was only $500 this card probably wouldn't exist. The part of the rumor about how these are just rebadged Radeon Instinct parts appears to be right.

Is AMD "losing" money in the opportunity-cost sense? Sure assuming they are getting strong demand for the full priced Radeon Instincts.

I don't entirely buy it. If they are die harvests then they couldn't have sold them anyways. That said, there's the 16 GB of HBM on the card that probably isn't cheap.

I can see two possibilities.

1) "We made more M150s than we could sell, so we have a bunch of inventory we'll have to write down. Selling it at a loss beats not selling it at all."

2) The story is complete hooey.

My guess is that the volume will be relatively low because they're going to mostly just sell partially-defective dies that weren't going to work for the MI50. Although why they haven't made an "MI45" or whatever to consume those at a higher price is anyone's guess.

derFunkenstein wrote:

My guess is that the volume will be relatively low because they're going to mostly just sell partially-defective dies that weren't going to work for the MI50. Although why they haven't made an "MI45" or whatever to consume those at a higher price is anyone's guess.

It is because they ate too much power at load for enterprise customers. They are rather picky with power budget in HPC/number crunching builds.

NTMBK wrote:

Sure, apart from the partially disabled die, fused off features, and totally new cooler, it's exactly the same as the MI50. :roll: It's not like they're just putting a sticker on the box, it's a different product.

It's not partially disabled, it's the exact same die. And the FP64 stuff is probably just done in firmware I bet. Or even the driver...

auxy wrote:

NTMBK wrote:

Sure, apart from the partially disabled die, fused off features, and totally new cooler, it's exactly the same as the MI50. It's not like they're just putting a sticker on the box, it's a different product.

It's not partially disabled, it's the exact same die. And the FP64 stuff is probably just done in firmware I bet. Or even the driver...

Are you saying the MI50 and MI60 aren't the same die? As R&D strapped as AMD is I just find that hard to believe.

derFunkenstein wrote:

auxy wrote:

NTMBK wrote:

Sure, apart from the partially disabled die, fused off features, and totally new cooler, it's exactly the same as the MI50. It's not like they're just putting a sticker on the box, it's a different product.

It's not partially disabled, it's the exact same die. And the FP64 stuff is probably just done in firmware I bet. Or even the driver...

Are you saying the MI50 and MI60 aren't the same die? As R&D strapped as AMD is I just find that hard to believe.

I read this as auxy saying that the Radeon Vega VII is exactly the same (die, HBM2, and interposer) config as the Radeon Instinct MI50.

The Radeon Instinct MI60 is the same die, but it is fully-enabled, and it has 4 8 GB stacks of HBM2 not 4 4 GB stacks.

Topinio wrote:

I read this as auxy saying that the Radeon Vega VII is exactly the same (die, HBM2, and interposer) config as the Radeon Instinct MI50.

The Radeon Instinct MI60 is the same die, but it is fully-enabled, and it has 4 8 GB stacks of HBM2 not 4 4 GB stacks.

Right, and so there are some defective dies.

edit: Oh, wait. I get it now. nvm.

auxy wrote:

NTMBK wrote:

Sure, apart from the partially disabled die, fused off features, and totally new cooler, it's exactly the same as the MI50. It's not like they're just putting a sticker on the box, it's a different product.

It's not partially disabled, it's the exact same die. And the FP64 stuff is probably just done in firmware I bet. Or even the driver...

Good point, I got the MI50 and MI60 mixed up- thought the MI50 was a 64CU part! Whoops.

So with that established, it does sort of seem like they're selling the exact same product with a different cooler and box. They would have had to staple on some video outs but you are kidding yourself if you think the Instinct boards didn't already have the pads for them. (・へ・;)

I wonder if whoever they got this story from wasn't simply told that AMD was losing money on them in the sense of not being able to sell these dies/cards as Instincts. There is no way the bill of materials is over $700... Σ(´ー｀；)

Is anyone even buying/using Instinct cards? I know a couple guys who do CFD stuff for big petrol companies and they're all in on NV. According to them, everyone in that industry is.

auxy wrote:

Is anyone even buying/using Instinct cards? I know a couple guys who do CFD stuff for big petrol companies and they're all in on NV. According to them, everyone in that industry is.

That may well be why the Radeon VII exists at all...

K-L-Waster wrote:

auxy wrote:

Is anyone even buying/using Instinct cards? I know a couple guys who do CFD stuff for big petrol companies and they're all in on NV. According to them, everyone in that industry is.

That may well be why the Radeon VII exists at all...

Between the CAD/Catia stuff I've personally done, and research in college (CFD/particle sim stuff)--it's always been Nvidia. They just have all the industry support and plugins. AMD had and still pretty much has nothing relative to what Nvidia does with third parties.

auxy wrote:

So with that established, it does sort of seem like they're selling the exact same product with a different cooler and box. They would have had to staple on some video outs but you are kidding yourself if you think the Instinct boards didn't already have the pads for them. (・へ・;)

I wonder if whoever they got this story from wasn't simply told that AMD was losing money on them in the sense of not being able to sell these dies/cards as Instincts. There is no way the bill of materials is over $700... Σ(´ー｀；)

Is anyone even buying/using Instinct cards? I know a couple guys who do CFD stuff for big petrol companies and they're all in on NV. According to them, everyone in that industry is.

It's a different PCB, per the videos on AMD's site. There are the outputs for video, power inputs on a different side, and possibly dropped PCIe gen. 4.

This Vega VII will have a different heatsink too, 3 axial fans, and obviously a completely different shroud. The firmware is different, probably dual and will disable ECC and have 1/16 rather than 1/2 the FP32 rate for FP64; also the profiles will be for slightly higher clocks.

The MI50 is not even available until late Q1.

And yes, almost everyone professionally uses NVIDIA not AMD.

K-L-Waster wrote:

auxy wrote:

Is anyone even buying/using Instinct cards? I know a couple guys who do CFD stuff for big petrol companies and they're all in on NV. According to them, everyone in that industry is.

That may well be why the Radeon VII exists at all...

Interesting theory. Make it cheap enough that someone who would otherwise not be interested tries it out? I think the problem afterwards is the Instinct cards with proper certification and ECC cost just as much as their Tesla/Quadro counterparts. I've seen quotes for MI25 ranging from $5000 to $10,500, depending on hardware validation level.

I think AMD fabricated a whole bunch of boards for MI50, hoping that the server vendors would lap them up, but then had a load of excess stock because it's REALLY REALLY HARD to actually buy a server with MI50.

All the vendors prefer to scale out with MI25 instead - probably better thermals, more flexible layouts and power delivery requirements. Either way, MI50 aren't selling because they're not something that anyone seems willing to actually put on sale!

Chrispy_ wrote:

I think AMD fabricated a whole bunch of boards for MI50, hoping that the server vendors would lap them up, but then had a load of excess stock because it's REALLY REALLY HARD to actually buy a server with MI50.

All the vendors prefer to scale out with MI25 instead - probably better thermals, more flexible layouts and power delivery requirements. Either way, MI50 aren't selling because they're not something that anyone seems willing to actually put on sale!

According to Topinio above, the MI50 isn't even available until later this quarter.

Someone is using AMD GPUs, because ROCm development continues to move forward.

But I don't know of anyone who uses them. My only guess is that AMD is using the GPUs internally as part of their EDA simulations or something. Fundamental electronics can be modeled as a linear algebra / matrix problem, and maybe other problems in the electronic-design space can benefit from GPUs.

My other guess is that the entertainment industry uses GPUs for rendering. But most of that seems to be NVidia these days (especially with the Raytracing feature / BVH acceleration on chip now... development seems to be moving towards NVidia). Maybe some video editing / compositing groups remain AMD. I know there are groups that stay on OpenCL1.2 for compatibility between the two platforms for sure.

Chrispy_ wrote:

I think AMD fabricated a whole bunch of boards for MI50, hoping that the server vendors would lap them up, but then had a load of excess stock because it's REALLY REALLY HARD to actually buy a server with MI50.

All the vendors prefer to scale out with MI25 instead - probably better thermals, more flexible layouts and power delivery requirements. Either way, MI50 aren't selling because they're not something that anyone seems willing to actually put on sale!

MI25 is Vega 10 and was announced December '16, and shipped summer '17, of course that's what you can buy from anyone, it's qualified and stocked; MI60 was only announced in November and the PR said first shipment to datacentre customers of MI60 was expected 'by the end of 2018' and MI50 was 'expected to begin shipping to data center customers by the end of Q1 2019'.

Though demand is probably really low compared to that for Tesla's, I'd bet it's the opposeite problem to having made too many -- I don't expect there to be much in the way of opportunity to buy Vega VII for a while, either.

just brew it! wrote:

"But we'll make up for it in volume!"

Checks out. They just have to sell negative units.

Topinio wrote:

This Vega VII will have a different heatsink too, 3 axial fans, and obviously a completely different shroud. The firmware is different, probably dual and will disable ECC and have 1/16 rather than 1/2 the FP32 rate for FP64; also the profiles will be for slightly higher clocks.

Correction: per Ryan Smith, the Radeon VII version has FP64 at 1/8 of its FP32, 1.73 TFLOPS double precision.

Edit: I think this is behind only the $3000 TITAN V, $3000 TITAN Z, and $1000 Radeon HD 7990 in FP64 compute, so for those in the niche of a niche who can use that capability, this is a potentially rather decent card at $700

It's also apparently 64 ROPs (unlike the previously reported 128). That's a pretty big improvement between 7nm and the massive memory bandwidth increase.

https://twitter.com/RyanSmithAT/status/ ... 2238201857

Topinio wrote:

Topinio wrote:

This Vega VII will have a different heatsink too, 3 axial fans, and obviously a completely different shroud. The firmware is different, probably dual and will disable ECC and have 1/16 rather than 1/2 the FP32 rate for FP64; also the profiles will be for slightly higher clocks.

Correction: per Ryan Smith, the Radeon VII version has FP64 at 1/8 of its FP32, 1.73 TFLOPS double precision.

It is still one heck of a steal for the general compute crowd. Nothing comes close in that price range.

Krogoth wrote:

It is still one heck of a steal for the general compute crowd. Nothing comes close in that price range.

Yep, particularly if you're in the FP64 crowd, cheapest > 1.5 TFLOPS ever and 4th 6th highest non-pro card ever, 2nd highest AMD.

So if this is true, would they die cut features or block it through firmware?

Be interesting if you could BIOS flash it back to an MI.

While I really don't have a good reason to buy one...for some reason I really want one. It'd be a huge step up in performance, a sidestep in power consumption (not that I really care), and feel good for some reason.

Waco wrote:

While I really don't have a good reason to buy one...for some reason I really want one. It'd be a huge step up in performance, a sidestep in power consumption (not that I really care), and feel good for some reason.

And maybe Freesync on your TV/Monitor?