Pricing Discussions around AMD VEGA *over-flow*

[Heinrich4]

Do they have any evidence to back up any of those claims? Their HBM costs article was bogus comparing capacity without accounting for bandwidth. With drivers and feature enablement, the issue would be getting the card performing as expected, not lowering the price to create a $300 Ti in the near future. If AMD we're losing money, they wouldn't be pushing out cards to miners or providing rebates.

Maybe you right. However for me the signs are clear, Raja's "sabbatical" retreat plus a series of problems (logistics, drivers, development for various markets ),plus the fact that vega 10/64 may need to change their MSRP, probably because of bill of materials.
AMD due to its current size and financies can not afford to subsidize large amounts of gpus, perhaps also so there is relatively little supply of vega gpu compared to the counterparts of nvidia,in addition off course to known problems with vendors/suplies (HBM,mounting interposers etc).

Edited.

 

[DmitryKo]

It's December now but nothing really changed - Rx 500/Vega cards are still priced at a 2x premium, and you still can't buy them even at these inflated prices, except for the lowest-end parts like Rx 550/560.

Even used Radeon cards sell at the same 2x premium - not only old GCN cards like HD7000-series and Rx 200/300/400-series, but even the now-unsupported HD5000/6000 series. And of course all used NVidia cards now command the same 2x premium as well! I don't even want to know what the fuck is this Etherium network and what it is good for, I just sincerely wish them to die in a horrible accident...


I can't see how AMD would overcome this shortage situation without losing relevance in the PC graphics market. If they can bump the production of Vega and/or Navi generation, miners will be happy to consume all additional cards wholesale - unless AMD increases the MSRP to match and exceed Nvidia cards with similar mining performance, which would only make matters worse. It's a dead fucking end for AMD gamers...

Even if Raja Koduri will be able to come with a ground-up redesign of Intel GPU and compute architecture, it won't be available until 2021 at the earliest, and competition like Nvidia Volta, Xeon Phi, and Qualcomm Centriq etc. does not come close to offering similar price-performance ratio.

 

[Bondrewd]

I can't see how AMD would overcome this shortage situation without losing relevance in the PC graphics market.

Move GPUs back to TSMC and order some more wafers ffs.

If they can bump the production of Vega and/or Navi generation

Lisa literally fears building the GPU inventory. For a good reason.

Xeon Phi

Phi is dead.

 

[Grall]

Move GPUs back to TSMC and order some more wafers ffs.

Ok, so spend unknown time and resources to re-engineer designs for TSMC process, then (assuming there is sufficient spare capacity available) wait additional ~4 months for them to be diffused, tested, hope no respin needed, packaged, assembled into graphics cards, shipped, then hope mining bubble hasn't popped by then...

 

[Bondrewd]

Ok, so spend unknown time and resources to re-engineer designs for TSMC process, then (assuming there is sufficient spare capacity available) wait additional ~4 months for them to be diffused, tested, hope no respin needed, packaged, assembled into graphics cards, shipped, then hope mining bubble hasn't popped by then...

He did mention Navi.
Navi is 146% N7 TSMC.

 

[DmitryKo]

Move GPUs back to TSMC

AMD has already contracted TSMC for Navi production. The interesting questions are: 1) will they be able (or willing) to ramp up production to satisfy the increased demand, and 2) if yes, will Etherium miners still constitute the entirety of retail sales, assuming price/performance will stay the same?

Phi is dead.

Dead or alive, it's not competitive for Etherium mining.

 

[Bondrewd]

1) will they be able (or willing) to ramp up production to satisfy the increased demand

They are most definetly not willing to pander to crypto market (at least right now).

2) if yes, will Etherium miners still constitute the entirety of retail sales, assuming price/performance will stay the same?

Does anyone mine ETH anymore? I thought XMR is the new hotness.

 

[entity279]

XMR just happens to run very efficiently on Vega.

everything is being mined right now, ETH included (and at recored rates)

 

[DmitryKo]

They are most definetly not willing to pander to crypto market

I don't think there is a feasible way to differentiate miners from gamers and system builders - shortages in the retail would affect all these categories, as all of them can order box or OEM package and small or large quantities.

The problem is, shortages would further reduce AMD's installed base, as the only guaranteed way to get a new mid- and top-range Radeon graphics card would be ordering an expensive custom-built PC from a large OEM builder. Then developers will start defecting to the NVidia camp, first making runtime optimizations for GeForce hardware then switching to proprietary CUDA and PhysX platforms.

I thought XMR is the new hotness.

Whatever cool new algorithm currently allows you to earn a couple extra bucks with your GPU until a faster dedicated ASIC is developed.

 

[BRiT]

I don't think there is a feasible way to differentiate miners from gamers and system builders - shortages in the retail would affect all these categories, as all of them can order box or OEM package and small or large quantities.

Sure you can, just put it in your EULA., but that's an entirely different discussion...

 

[Deleted member 13524]

Sure you can, just put it in your EULA., but that's an entirely different discussion...

And/or use some really short-lived fans that will last a demanding gamer's lifetime but will break after 3 months of use if it's spinning at 3500 RPM 24/7. That would teach them a lesson!

Dead or alive, it's not competitive for Etherium mining.

The Xeon Phi 7210 co-processor card seems to be very competitive for cryptonight though. I don't know if it's just from using the MCDRAM as L3 (It seems to be just some kind of low-clocked 8-Hi HBM1) or the miner app is making use of those total 32MB L2 that are connected through an inter-core mesh, but it seems to be a lot more power-efficient than Vega. And you can buy one at an e-tailer for the same $1500 as a Vega 56

Whatever cool new algorithm currently allows you to earn a couple extra bucks with your GPU until a faster dedicated ASIC is developed.

These algorithms are made to be extremely dependent on very high bandwidths, meaning they're only really effective on GDDR5 in wide buses, HBM1/2 and L3 cache. Dedicated ASICs will find a really hard time competing with GPUs and CPUs on this. They can't just make a bunch of fixed-function ALUs for SHA-256 and slap a 16bit DDR3 chip in there for some occasion I/O, like they did with all those ASICs before.

And if ordering custom ASICs with DDR3 chips is expensive (because of low quantitites), doing the same with wide and expensive memory would be prohibitive for these small IHVs.

Plus, there's absolutely no resale value for these cheaps, whereas a miner can always sell a graphics card back to a gamer.

 

[Grall]

(It seems to be just some kind of low-clocked 8-Hi HBM1)

IIRC, Phi uses hyper memory cube (hyper cube memory? well, whatever...) for its stacked DRAM.

And you can buy one at an e-tailer for the same $1500 as a Vega 56

Holy crab, surely typo, yeah? Because vega 56 CAN'T cost $1500, that's crazy.

 

[DmitryKo]

Sure you can, just put it in your EULA

What exactly should you put into the EULA - "do not use our drivers for some arbitrary computing tasks which we don't really like"?

The purpose of a software license is primarily 1) to protect the code by reserving full ownership and restricting the rights to modify, reverse engineer, redistribute, etc. 2) to limit the warranty and indemnify the developer from any liabilities arising from its use.

And did EULAs really stop anyone from modding or reverse engineering the drivers?

use some really short-lived fans that will last a demanding gamer's lifetime but will break after 3 months of use if it's spinning at 3500 RPM 24/7

Or initiate a DRM system where approved developers would get a cypher key to enable high-performance mode, while everything else will be slowed down to a crawl! This DRM should of course require permanent Internet connection.

 

[DmitryKo]

The Xeon Phi 7210 co-processor card seems to be very competitive for cryptonight... it seems to be a lot more power-efficient than Vega. And you can buy one at an e-tailer for the same $1500 as a Vega 56

Intel cancelled Xeon Phi 7200 coprocessors (PCIe add-on cards) - only LGA3647 socket processors are currently available. You would need a rack PC hosting several very expensive LGA3647 add-on boards, or a dedicated tower case with an expensive multi-socket LGA3647 motherboard - that's certainly over $800-1000 which Vega 64 currently retails for.

there's absolutely no resale value for these cheaps, whereas a miner can always sell a graphics card back to a gamer.

This is why I see no easy solution. Until datacenters start making it faster/cheaper using dedicated hardware, shortages of high end Radeon cards will remain, and AMD will slowly lose their position in the gaming ecosystem.

ordering custom ASICs with DDR3 chips is expensive (because of low quantitites), doing the same with wide and expensive memory would be prohibitive for these small IHVs

There are HBM/HBM2 memory controller IP solutions recently made available for a full range of TSMC, Samsung and GlobaFoundries process technologies, their claimed client base includes cloud computing ASICs.

https://www.esilicon.com/company/news-events/press-releases/esilicon-tapes-deep-learning-asic/
https://news.samsung.com/us/samsung...network-processor-rambus-28g-serdes-solution/
https://www.rambus.com/blogs/globalfoundries-demonstrates-2-5d-high-bandwidth-memory-hbm-solution/
https://www.globalfoundries.com/new...ates-25d-high-bandwidth-memory-solution-data/
http://www.open-silicon.com/open-si...ive-ip-subsystem-solution-for-high-end-netwo/
http://www.globalunichip.com/en-global/news/pressDetail/HBM2_Total_Solution/