Will AMD release any high performance GPU's in Q2 or Q3?

[Zenny]

And yet Pascal can still only throw everything through it's pipelines serially.
Where as up to and including Polaris (I don't know the one of Vega ... yet) is fully "native" parallel processing capable. (Polaris has 2 HWS f.ex.)

In terms of pure compute performance AMD has always had more power but never being exploited because of eco systems, this has been changing and still is.

That is my point, it's going from 1 philosophy and design to the complete opposite whilst that part is starting to improve considerably.

What? Pascal doesn't do workloads serially, where did you get that info? I can only assume you are getting mixed up with async compute but that's nothing to do with how parallel a GPU is. Once again just check Nvidia white papers and marketing material from as way back as 2008 (myth busters even did a video about it with them), everything is about parallel computing. That's what fundamentally CUDA is about, GPGPU. Jeez, Nvidia GPUs have been doing this since the GeForce 8 days. https://www.cs.umn.edu/activity/talk...-geforce-8-gpu

[Evildeffy]

What? Pascal doesn't do workloads serially, where did you get that info? I can only assume you are getting mixed up with async compute but that's nothing to do with how parallel a GPU is. Once again just check Nvidia white papers and marketing material from as way back as 2008 (myth busters even did a video about it with them), everything is about parallel computing. That's what fundamentally CUDA is about, GPGPU. Jeez, Nvidia GPUs have been doing this since the GeForce 8 days. https://www.cs.umn.edu/activity/talk...-geforce-8-gpu

Not Async Compute, though that falls under it as well, Pascal is architecturally designed as such as it is a direct derivative of Maxwell which does the same thing.

If you read about how Pascal operates it will tell you it is fed all commands for it's queue serially and via the driver instead of having said hardware to distribute it.
This means the driver schedules and interrupts the GPU for commands through 1 list, Maxwell used to HAVE to finish their previous operation before moving on to the next operation where Pascal introduced firmware level changes where it allowed to halt/pause it's operation, interject another operation and continue with the previous operation.

This can only be done serially because nVidia designed their cards to have 1 giant pipeline for this.

AMD's designs (like I said though .. for me an unknown for Vega) has hardware schedulers placed as well as multiple of pipelines to accept and execute all commands in parallel, eliminating the driver middle-man (and therefore latency) entirely.

So to clarify it a little bit more:
AMD's architecture allows for immediate parallel command processing and thereby a full low level access.
nVidia's architecture requires the command to be given from the driver to start something parallel and get data first then it can tell it's cuda cores what to do.

nVidia has over time optimized it as far as software optimizations can go, and they did so pretty well at that.

This architectural difference in design is what often gives AMD a pure compute edge but also is a cause for the increased power consumption over their nVidia counterparts, the hardware that allows to do this is "expensive" in terms of power allocation and die size.

This operation was exposed during the Async Compute hype days indeed but it is how the GPUs operate regardless of load.
I specifically stated parallel processing and not parallel execution, in every which way in Pascal this doesn't alter.
Volta maybe it will ... but Pascal (and Maxwell but surprisingly not Kepler since that does actually have more hardware to parallel process) is as it is.

[Thunderball]

Not even sure that will even remotely help. Volta is around the corner too. Vega is already pretty bad vs Pascal, Volta will just obliterate it.

That depends on if AMD can significantly drop prices. Larger chips (like Vega) typically heavily benefit from being in production for a long period of time, improving yields (which is a main problem for larger chips' pricing) and maybe somewhat improving silicon quality (allowing for higher clocks/lower voltage>less power consumption). This has been done multiple times before - Nvidia has been doing that with top Ti cards for 3 gens now: by releasing top of the line chip via professional/Titan cards and releasing a mainstream card based on the same chip later when production matures.

[Evildeffy]

Just as a sidenote:

Vega Reviews are out, check your local YouTube/Websites for their reviews.

Vega 56 seems to be a winner although like always ... don't go for reference.

[Thunderball]

What? Pascal doesn't do workloads serially, where did you get that info? I can only assume you are getting mixed up with async compute but that's nothing to do with how parallel a GPU is. Once again just check Nvidia white papers and marketing material from as way back as 2008 (myth busters even did a video about it with them), everything is about parallel computing. That's what fundamentally CUDA is about, GPGPU. Jeez, Nvidia GPUs have been doing this since the GeForce 8 days. https://www.cs.umn.edu/activity/talk...-geforce-8-gpu

Chill, the guy just likes his AMD stuff. Nvidia card generally have a lot shorter pipelines compared to AMD cards so you can consider AMD cards as ones doing more work in parallel, but that doesnt mean anything, every approach has it's pros and cons. Longer pipelines mean a larger chip that consumes more power, shorter pipeline means having to rely on a better software optimization and higher clocks (which Nvidia manages fine as we know).

[Gaidax]

Yes likely driver optimizations will help as will decent third party models. Problem is by the time all that comes it will hardly be relevant.

[Thunderball]

So the reviews are out. Vega 56 (which people should only care about) looks like 3-5% ahead of 1070 with $50 higher MSRP and 125W higher TDP, oh and it's stuck at 300W max power draw, so no overclocking (at least GPU overclocking).

[Lathais]

So the reviews are out. Vega 56 (which people should only care about) looks like 3-5% ahead of 1070 with $50 higher MSRP and 125W higher TDP, oh and it's stuck at 300W max power draw, so no overclocking (at least GPU overclocking).

Now is that only because reference design and throttling? Of course, the third party boards will likely be more expensive so it's kinda moot, just curious if they are only that 3-5% ahead because of throttling.

- - - Updated - - -

Yes likely driver optimizations will help as will decent third party models. Problem is by the time all that comes it will hardly be relevant.

Yeah, that's the big issue. Too late to the party.

[Life-Binder]

Vega 56 isnt that bad from what Im reading, the only real drawback compared to 1070 seems to be power consumption (but its still better then Vega 64s perf/watt) .. its nothing amazing though, just a small amount ahead of a stock 1070 .. final prices can be anything these days due to gouging and miners

can be worth a pickup if you have Freesync


Vega 64 though is just ugh .. and the $699 price tag on the liquid cooled 64 version should not exist, period



the most dissapointing thing (for me) is that there is no 1080ti contender, even the liquid cooled 64 is not close to a stock Ti, much less an AIB Ti

wait for Navi (c)




I expect Vega 56 might settle somewhere toward the bottom of steam charts .. assuming the miners dont buy up every last unit

dont expect to ever see Vega 64 there, its pretty close to being DoA as far as gaming segment goes

- - - Updated - - -

[Evildeffy]

Chill, the guy just likes his AMD stuff. Nvidia card generally have a lot shorter pipelines compared to AMD cards so you can consider AMD cards as ones doing more work in parallel, but that doesnt mean anything, every approach has it's pros and cons. Longer pipelines mean a larger chip that consumes more power, shorter pipeline means having to rely on a better software optimization and higher clocks (which Nvidia manages fine as we know).

I like both, what I don't like is nVidia's consumer "friendlyness" ... but I give credit where it's due.

And I dare you to prove me wrong as I mentioned only technical details which are correct.

Random sidenote:
The "longer" pipelines you mention don't really matter in GPUs as branch prediction does in CPUs.
The "larger amount" of pipelines however do as it means there's a HardWare Scheduler in place, the difference between an nVidia Maxwell/Pascal and any AMD from HD79X0 series and up (funnily enough both HD5000 and HD6000 series were designed as nVidia is now and vice versa).

Like I said prior the pro is that with AMD's approach they have a better compute capability in general as programmes are better programmed to handle as such whilst gaming, thanks to Microsoft's half-arsed approach of DX9/11 and OpenGL's notorious difficulty and failed documentation, are not.

And I like my general technical knowledge stuff, I cannot help it if people cannot accurately read what's written.

[Life-Binder]

http://www.tweaktown.com/articles/83...e/index12.html

http://www.tweaktown.com/articles/83...e/index13.html


damn, tweaktown completely slammed vega


guru3D and some others were less harsh

[Thunderball]

Now is that only because reference design and throttling? Of course, the third party boards will likely be more expensive so it's kinda moot, just curious if they are only that 3-5% ahead because of throttling.

It's power throttling, better cooling is not going to change much. Lower operating temperature is going to make Vcore VRM more efficient but that's not much of a difference.

[Lathais]

It's power throttling, better cooling is not going to change much. Lower operating temperature is going to make Vcore VRM more efficient but that's not much of a difference.

So it's thermal throttling and you don't think better cooling will help? Price aside, better cooling will make it better than a 1070. If we also have third party boards being less expensive than reference, like we saw with nvidia this last generation, they could end up having a place on the market, however briefly. I agree though, for the most part, these cards are DoA and should not really be purchased.

[Thunderball]

So it's thermal throttling and you don't think better cooling will help? Price aside, better cooling will make it better than a 1070. If we also have third party boards being less expensive than reference, like we saw with nvidia this last generation, they could end up having a place on the market, however briefly. I agree though, for the most part, these cards are DoA and should not really be purchased.

Power throttling. AMD has to distinguish Vega 56 from Vega 64 so max power limit is going to heavily limit it's performance, and effectively remove any overclocking headroom while we're at it.

[Life-Binder]

[Lathais]

Power throttling. AMD has to distinguish Vega 56 from Vega 64 so max power limit is going to heavily limit it's performance, and effectively remove any overclocking headroom while we're at it.

Ah, ok, if it's power throttling, then yeah, that's entirely different and yeah, not gonna get much more out of it. My mistake.

[Shakadam]

So it's thermal throttling and you don't think better cooling will help? Price aside, better cooling will make it better than a 1070. If we also have third party boards being less expensive than reference, like we saw with nvidia this last generation, they could end up having a place on the market, however briefly. I agree though, for the most part, these cards are DoA and should not really be purchased.

Vega 56 is already better in price/performance than the 1070 if they're available somewhere close to MSRP. The performance in reviews varies according to which games are tested but on average the stock Vega 56 is a little above the GTX 1070, and it also seems to have higher OC potential (especially once 3rd party cards are available). Sweclockers got their OC'ed Vega 56 to perform very similarly to a stock Vega 64/stock GTX 1080

http://www.sweclockers.com/test/2424...-56/26#content

The trade off as expected is higher power draw compared to the Nvidia competition, but when undervolting (+OC'ing!) the Vega 56 is not as catastrophic anymore as seen from the Gamers Nexus video. Higher still than the GTX 1070 yes but the performance is also higher.



Imo the Vega 56 is looking as a pretty good option at its MSRP, especially if you already have a Freesync monitor. Still waiting to see what the 3rd party offers will actually retail at but I'm cautiously optimistic. The fact that it's already ahead of its main competition regarding performance and that gap will likely only expand with maturing drivers should certainly not be dismissed either.

The Vega 64 is much less interesting though. Trading blows with the GTX 1080 at much higher power draw and with what seems like very limited OC potential isn't very appealing imo.

[Life-Binder]

1070 is soon 1.5 years old


Volta 2070 in Q1 2018, wam bam thank you Mam

[Kagthul]

Dont forget that for a lot of people, youll have to factor in nit just the cost of the GPU, but also a new PSU to run it. So the card may carry a 399$ MSRP, but factor in a 70-120$ for a new PSU.

Add that to these cards, and their value peoposition takes a hefty dive.

[Gaidax]

The power consumption is unreal, who is watercooled version for? Enthusiasts? Worse than 1080Ti in every way imaginable.

The only viable card there is Vega 56, rest are simply waste of time.

Meanwhike Nvidia are as cool as a cucumber, I bet they won't even bother trying to drop prices, except for maybe 1070 where Vega makes at least some lick of a sense.