AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

silent_guy said:
http://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-1080-and-1070-founders-edition-review/10

For compute, they preempt at instruction level or thread level. For graphics, it's pixel level.
Click to expand...
I mis-remembered the diagram then, with in-flight pixels needing to reach export before preemption can occur.

Anarchist4000 said:
There will always be tradeoffs. The QoS measures should be easily disabled if there was no need. Reserve CUs for guaranteed latency, but risk them being idle. Same with reserving waves, but you only lower occupancy as opposed to idling hardware. It's ultimately up to the developer to decide which seems the proper route. Options never hurt.
Click to expand...
My position is that there are trade-offs and that the developer must decide. That doesn't leave much room for the HWS to infer what the tradeoffs are.

Yes, and full preemption would always be undesirable. That said, there are cases where it's the best choice. It could be compute preemption as well.
Click to expand...
Full preemption of the graphics context is very costly, and it's generally the best choice when there aren't alternatives.
Compute is able to preempt at a far smaller granularity, which is where discussing this often requires stating what is being preempted, and at what granularity. There are vastly different things operating under the label of preemption.

I'm not saying CU reservation isn't a good option for some cases. It definitely has it's uses, but will likely come at the cost of overall throughput. Throwing the entire GPU at a problem with prioritization and some contention would likely outperform a 20% reservation with guarantee. Predictable that takes longer to execute isn't ideal. Regardless, all of these options exist so a developer can decide what is best.
Click to expand...
That's called QoS, and predictable with overall performance being poorer is often what real-time entails.
It is not about outperforming a solution that is highly unresponsive, it is about making the hardware acceptable for task types that will not tolerate typical GPU latencies or safe from having the system reset the device due to timeouts.

We haven't seen it, but gating off execution units in general isn't exactly new.
Click to expand...
Gating off whole SIMDs is not new, while going down to the lane level has not come up for architectures in general discussion.
Doing per-lane voltage adjustment on top of that also hasn't come up for current architectures.
 
The 375 watt figure is really odd. Not only is it easily the largest GPU TDP I've ever heard of (how many power connectors is that?) but it also, bizarrely, signifies Vega would go down significantly in perf/watt figures from Polaris. Or at least the Frontier Editions would.

For reference Polaris gets about .046 teraflops per watt (at 11.5 teraflops for 250 watts). Under the assumed TDP it'd go down to .034. I can't really see the newer GPU losing 26% of its efficiency. Now one possibility is that the Frontier Editions are basically uselessly bad early cards AMD is hoping to make money off of anyway. It's not like they haven't done new types of chip binning recently (Nano). And the right sort of customer might want to take an order of 13.1 teraflop GPUs regardless of tdp, it being at least near Nvidia's not deliverable for several months/way too expensive 15 teraflop Volta stuff.

But the other possibility is that it's just some typo someone from AMD made and it's supposed to be 275 watts. A 10% increase in TDP from the usual 250 watt high end, for a roughly 5% increase in clockspeed from the earlier performance figures of 12.5 teraflops seems entirely reasonable. Especially for a "Gold plated/high end bin" ridiculous kind of GPU like the $1800 one is supposed to be. Don't know which of these is true but I'd bet it's one or the other at the very least.
 
So, in the preview, the AMD official clearly states the FE card is for professionals, it may not be using certified drivers, but they are definitely using some kind of pro drivers. They are even using a Radeon Pro renderer API in the CAD application. The guy actually outright admits it, the FE card has the advantage of workstation drivers, compared to TitanXp which lacks them. Will be curious to know about the difference between the FE edition and the WX line, now that they are very similar.

Also they did gaming comparisons, but with no fps counter, which is a curious omission!
 
Last edited:
DavidGraham said:
Also they did gaming comparisons, but with no fps counter, which is a curious omission!
Click to expand...
The positive thing is that gaming performance was apparently indistinguishable from Titan Xp, which sounds good for RX-version which is supposed to do better in games
(though gametests were limited to 3 games)
 
Kaotik said:
The positive thing is that gaming performance was apparently indistinguishable from Titan Xp, which sounds good for RX-version which is supposed to do better in games
Click to expand...
Maybe, but you can't really distinguish 65fps from 75fps, or a 100fps from 90fps with the naked eye. You need fps counters fort that, which was curiously lacking here.
Kaotik said:
(though gametests were limited to 3 games)
Click to expand...
This is actually suspicious, the fact that AMD can't get a clear performance advantage in 2 AMD optimized games using AMD optimized Vulkan and DX12 paths is clearly telling. Fjij was faster in Sniper Elite 4 than 1070 and 980Ti, despite being slower overall. I expected Vega to at least show a clear advantage in this title. FuryX was also faster than 1070 and remains faster than 980Ti in Doom, before NV drivers uplifted the 1070 performance above the FuryX.
 
Last edited:
DavidGraham said:
So, in the preview, the AMD official clearly states the FE card is for professionals, it may not be using certified drivers, but they are definitely using some kind of pro drivers. They are even using a Radeon Pro renderer API in the CAD application. The guy actually outright admits it, the FE card has the advantage of workstation drivers, compared to TitanXp which lacks them. Will be curious to know about the difference between the FE edition and the WX line, now that they are very similar.

Also they did gaming comparisons, but with no fps counter, which is a curious omission!
Click to expand...

I always found the whole "But it's not a pro card!" line of thought somewhat puzzling, considering:

http://pro.radeon.com/en-us/frontier/
 
So these FE:s seems to have "semi-pro-drivers" that isn't as optimized as the real "Radeon Pro" Vegas will be that comes later, and "semi-gaming-drivers" that isn't as optimized as the "Radeon RX Vegas will be that also comes later?

Strange product launch. I don't get it?
 
Geeforcer said:
I always found the whole "But it's not a pro card!" line of thought somewhat puzzling, considering:

http://pro.radeon.com/en-us/frontier/
Click to expand...
Here’s the disruptive thing
That brings us to the disruptive part of AMD’s push: Although powerful, Titan Xp is limited to consumer drivers. To approach the performance of the Radeon Vega Frontier Edition with an Nvidia product, you’d have to step up to at least a $2,000 Quadro P5000. A P6000 card, which is kind of the professional equivalent of a Titan Xp, is nearly $6,000.

AMD blames these high prices for driving many prosumers to order workstation-class systems with lower-end professional cards so they can replace them with faster consumer GPUs. AMD believes the Frontier Edition can end this madness and give the company a triumphant return to the professional workstation card game.

Note that the Frontier Edition will do so with pro-optimized software support, though the drivers for the Radeon Vega Frontier Edition won’t actually be fully certified. AMD contends the new GPU design and the HBM2 memory matter as much as the drivers in making Radeon relevant for workstations again.
http://www.pcworld.com/article/3202...vega-frontier-edition-vs-nvidia-titan-xp.html
Click to expand...
It seemed fairly clear in the article. They've been comparing to consumer Titan Xp because that is the targeted professional market for FE.

DavidGraham said:
This is actually suspicious, the fact that AMD can't get a clear performance advantage in 2 AMD optimized games using AMD optimized Vulkan and DX12 paths is clearly telling. Fjij was faster in Sniper Elite 4 than 1070 and 980Ti, despite being slower overall. I expected Vega to at least show a clear advantage in this title. FuryX was also faster than 1070 and remains faster than 980Ti in Doom, before NV drivers uplifted the 1070 performance above the FuryX.
Click to expand...
Why is that at all suspicious? It's exactly what would be expected given what we know so far and the PC World article explained as much.

Those AMD "optimized" titles wouldn't have been well optimized for an unreleased architecture running incomplete drivers. Why would anyone have expected AMD to present numbers that aren't final, likely to increase significantly, and aren't on the gaming clocked part? They simply ballparked current performance against Titan Xp as that appears to be Vega's target. That's been known for a while based on theoretical specs.
 
Geeforcer said:
What seems clear is that they have a professional card that they absolutely positively refuse to compare to competing professional cards at any price point.
Click to expand...
It's probably the same hardware, but unfinished drivers with no certification targeting a price sensitive market. They even stated as the use case Xp wouldn't have had pro drivers because of cost. The WX line would be the comparison for Nvidia's pro line and we haven't seen that yet.
 
Anarchist4000 said:
It seemed fairly clear in the article. They've been comparing to consumer Titan Xp because that is the targeted professional market for FE.
Click to expand...
And Yet an 800$ castrated GP104 Quadro is equal or better than Vega FE.
Why would anyone have expected AMD to present numbers that aren't final, likely to increase significantly, and aren't on the gaming clocked part?
Click to expand...
Because they have done so in the past? and multiple times already? To just not do it now in the presence of the competitor is a huge omission on AMD's part, one that signifies total inconfidence in their product.
Anarchist4000 said:
but unfinished drivers with no certification targeting a price sensitive market
Click to expand...
Who said they are unfinished drivers? they are just uncertified, meaning direct support from AMD for these pro apps is not available. Also please stop quoting the article, AMD statements in the video are clear enough on their own
 
You must log in or register to reply here.
Back
Top