Pricing Discussions around AMD VEGA *over-flow*

[Deleted member 13524]

I guess in that sense Navi will be more "evolutionary" than revolutionary with regard to Vega.

Yes, I bet Navi will be to Vega what Polaris was to GCN3 (i. e. mostly a Vega shrink to 7nm and some low hanging fruit). Biggest change we'll see is if Navi gets to adopt multi-chip configurations seen as a single GPU.

 

[Rootax]

Excuse me but Navi need to be "revolutionary". Vega is maybe nice from a tech PoW, but it doesn't performs. And if it's too complexe for your driver team, then it's not a good solution in the end. Ffs a oc furyX could take on Vega56 right now.

Navi need to change all that, and performs from day one, not be a headcase for driver team, and while at it, close the performance/watt with nVidia. So, a revolution imo.

 

[itsmydamnation]

Excuse me but Navi need to be "revolutionary". Vega is maybe nice from a tech PoW, but it doesn't performs. And if it's too complexe for your driver team, then it's not a good solution in the end. Ffs a oc furyX could take on Vega56 right now.

Navi need to change all that, and performs from day one, not be a headcase for driver team, and while at it, close the performance/watt with nVidia. So, a revolution imo.

RV770 wasn't revolutionary in any way over R600.

 

[3dilettante]

There was that set of patents concerning variable-width SIMD and high-performance scalar units that were speculated to be part of Vega, ahead of launch.
Those seemed like they'd play well with the varying output of the geometry pipeline and rasterizer, but no indication of those for Vega. If they did show up for Navi, it could be as big a change to the CUs as the primitive shaders and rasterizer were to the geometry engines.

On the compute side, Volta has some notable changes to thread synchronization and divergence handling that AMD may need to respond to. It has certain patents dealing with synchronization and wave compaction that might not be totally the same but would help. Nothing in Vega notes this, but it's something that likely needs to show up if HPC, variable-width SIMDs, or compute are to be a target for Navi or the next generation.
Then there's AMD's various methods for memory coherence, more effective caches, and stronger consistency, particularly if sharing with CPUs, which nothing has been noted for as of yet for Vega.
Then there's dynamic load-balancing with work-stealing and work-sharing that AMD has patented, and again hasn't mentioned for Vega. One unknown is how the geometry engines have implemented some kind of intelligent workload scheduler, and if this is another case where the front end has started implementing features the back ends haven't.

It seems like there are significant parts of GCN that AMD has made a lot of noise about needing to do better that Vega hasn't been documented as doing anything with.

 

[xEx]

So apparently GN will release a video where they disclose the cost of vega and show that AMD is not making much of money with the "mrsp" prices and that the bundles were a way AMD tried to make some money out of it.

 

[digitalwanderer]

Ugh, the more I think about this the more it makes my thinker-thingy in my head hurt!

There is no good way to get these cards to gamers over miners by bundles or "introductory prices", there just ain't. By not fighting the miners they're going to lose the gaming market share no matter what, so screw it! Raise the price, hell jack it up through the roof! The only people who are going to be getting these are miners because they will stop at nothing to get them. Bundles, price; all irrelevant unless something can be done to stop the miner's insatiable demand for it.

Or mebbe I'm just having a bad day, I dunno. <sigh>

I just see it at the very least worst option. If the cards are the best at mining the miners will get them no matter what, so AMD might as well make a profit off them and a generous one at that rather than the retailers who will jack the prices up anyways.

 

[Rootax]

No serious miner want this.

 

[seahawk]

Pascal is old because it's basically Maxwell in 16FF+ and super optimized for higher frequencies.
It's as old and dying as Polaris.

Imho it is not. It is a logical evolution of Maxwell with ASC scheduling improved and other smaller tweaks.

 

[Entropy]

So apparently GN will release a video where they disclose the cost of vega and show that AMD is not making much of money with the "mrsp" prices and that the bundles were a way AMD tried to make some money out of it.

That would be very strange. The money AMD makes out of vega depends on costs and accounting. The costs aren't available even to AMD shareholders and changes over time due to yields, cost of parts (such as memory) and volumes. The accounting is as-you-wish.

We might get a snapshot BOM from taiwan soon.

 

[entity279]

No serious miner want this.

How can you tell? These sites testing mining performance barely know to run the program and do some basic over / under - clocking.

Have they played with memory straps (or custom memory timings), voltage mappings, have they ran through the various power targets? And so on..

Or maybe no miner wants it today, how about tomorrow?

 

[Rootax]

How can you tell? These sites testing mining performance barely know to run the program and do some basic over / under - clocking.

Have they played with memory straps (or custom memory timings), voltage mappings, have they ran through the various power targets? And so on..

Or maybe no miner wants it today, how about tomorrow?

I can tell from being a (small) one, and talking with a lot of them. It's not rocket science, the performance/power consumption is not good, especially with the eth situation. And it requires bigger power supply. So, I don't base what I said on the sites testing mining performances, but from the mining crowd. Eh, I'm sure some of them will by some, but it's not a good solution right now, and I doubt it will ever be (it's won't do 36/37mh/s under 120-130w...). And I'm not even talking about the price of the card...

 

[Deleted member 13524]

I can tell from being a (small) one, and talking with a lot of them. It's not rocket science, the performance/power consumption is not good, especially with the eth situation. And it requires bigger power supply. So, I don't base what I said on the sites testing mining performances, but from the mining crowd. Eh, I'm sure some of them will by some, but it's not a good solution right now, and I doubt it will ever be (it's won't do 36/37mh/s under 120-130w...)

That's great to know.
Keep telling your friends to stay away from Vega then. Be sure to spread that message as much as you possibly can.

 

[Rodéric]

nVidia sells the Founder Editions up to two per customer, assuming a customer is name + address it should seriously limit parasites from getting them...

 

[leoneazzurro]

I can tell from being a (small) one, and talking with a lot of them. It's not rocket science, the performance/power consumption is not good, especially with the eth situation. And it requires bigger power supply. So, I don't base what I said on the sites testing mining performances, but from the mining crowd. Eh, I'm sure some of them will by some, but it's not a good solution right now, and I doubt it will ever be (it's won't do 36/37mh/s under 120-130w...). And I'm not even talking about the price of the card...

In reality it depends. I know some miners that purchased RX Vega, as they planned to undervolt the core and overclocking only the memory. They obtained a 37MH/h value, that is higher than everything Nvidia has in the consumer space, even if optimized (blockchain driver but no BIOS mod). That could or could not be more profitable, depending on the price of electricity, which varies not only from nation to nation but also from individual to individual (as I stated before in this thread). If those 7 MH/h (assuming an average 30MH/h for the competition) are worth more than the price of energy plus the eventual purchase price difference, then Vega is more profitable than competition, even for AMD's own cards. And when undevolted and underclocked, Vega gains miles in the power efficiency department, as some reviews found, so the difference is not so high as at stock settings.

 

[CarstenS]

With a Titan XP/Xp (whether or not you're counting that as consumer space - they run with regular GF drivers and Vega FE is not substiantally less expensive), you can also get to that level of 37 MH/s with memory Overclocking. But I've seen Vega with some more OC reach 40 MH/s, so there's probably a bit of headroom left, especially when you run on very good conditions for electricity and are space constrained, i.e. need as much MH/s per card as possible.

Sadly, still no reliable undervolting apart from the Power Target slider (don't know how it is called in the english version of the driver, in german it's Leistungsgrenze, Power Limit) for me.

 

[leoneazzurro]

Yes, I think there is also some headroom left, because i.e. for Polaris cards there are some modifications "avalilable" allowing them to be more optimized and thus achieving higher than usual MH rate. At the moment those are not yet available for Vega, AFAIK. Moreover, looking at a pure bandwidth perspective (as for Eth it should be memory constrained) , we have a theoretical 65% advantage of Vega Vs. RX480/580 (and latency should be also lower, not counting even the new hashing instructions in Vega), which we have not seen yet in practice. Modded 480/580 are said to reach a little below 30MH/h:

https://forum.ethereum.org/discussion/12292/sapphire-rx-580-pulse-8gb-hash-rates

 

[CarstenS]

The new instruction is intended for accelerating SHA256, as stated in the Shader ISA:

499
V_XAD_U32
D.u32 = (S0.u32 ^ S1.u32) + S2.u32.
No carryin/carryout and no saturation. This opcode exists to
accelerate the SHA256 hash algorithm.

Ethereum does to my knowledge not use this, but instead the similar Keccak-256.

ETH hashing is not only about raw bandwidth - with that, Fiji would be king, but it is not. Even GDDR5X is much slower than GDDR5 (1070 outperforming 1080, 1080 Ti or Titan XP/p only ~20% faster than 1070 instead of 50).

 

[leoneazzurro]

Yes, I know that, but I'm also taking in account that Vega has ~ 2 times the raw compute power than a 580 and Latencies should be better in HBM2 that GDDR5. I'm not saying that we'll see a doubling of the performance, but it's likely that with some mods Vega could break the 40MH/h barrier.

 

[CarstenS]

Just coming from Fiji experience. Compute Power means very little on ETH hashing. Also, counting latencies in time required instead of clock cycles should give a clearer picture when comparing HBM, HBM gen2, G5 and G5X.

 

[Deleted member 2197]

This is assuming the primary audience of the RX Vega is gamers.

AMD needed limit one card per purchaser for the first 3 months at the introductory price of $499, or until they had sufficient supply. Any more than one card purchased would double the price for each card.

And they should have not released specialized mining logic in their drivers if the primary target of the cards was gamers.

No one is to blame but greedy AMD .....