AMD Moves Vega 56 Embargo Forward, Asks Reviewers to Prioritize Over 64

[IntMD]

Well, at least stock shouldn't be much of an issue, anyway from the look of it. Overclockers apparently have over 1k of the various amounts in stock. Orders begin somewhere between 12pm & 2pm (UK time - most likely about 2pm) on Monday essentially.

 

 

https://forums.overclockers.co.uk/threads/18789296/

[laminutederire]

4 hours ago, LaboonTheWhale said:

Depending on the game and the optimization. Games I have been mainly playing like Rising storm 2: vietnam, Squad, and VR it struggles. Now I'm just being nit picky as it can reach 60 in most, but after investing in a 144hz monitor. I want it to reach it. only 144hz I've experienced so far without LOW graphics is minecraft  Paired with a ryzen 1700 @ 3.8Ghz

At 1440p with a Fury Nitro, I got 144fps in cs go, basically all old source games, in all bioshocks except the last one where it sometimes dips into the 120fps range, in all platform games, as well as alien isolation most of the time. DOOM runs fairly well as well (Overwatch as well, when I tried it (basically all e sports titles))

However, even if it's not optimal, 75-100 fps range with freesync is noticeably better than 60fps locked, even if it underutilize the monitor a bit. You'd have to have a bit more than 1080 ti perf with a more gaming oriented cpu to attain 144fps in all games, so I guess above 75fps in AAA games with freesync is livable until at least Vega 20/Volta.

 

On topic, this may show that AMD finally understood we just want the bloody damn cards and that they should just stop waiting to release them and release them as soon as possible.

[JupiteL]

I don't think I'll be able to buy one anyway since 580 already either twice the price or not available.

[Bananasplit_00]

6 hours ago, dizmo said:

It's a little better though, no?

If it's 20% better, I'm still all over it if it's what the Nano is based on.

If it's par, then the 1070 wins because of it's much lower TDP.

 

I like the Fury X, just can't use it because of the cooler. I considered the R9 Nano, but want more VRAM.

Get a Fury X and a broken RX or R9 GPU for its cooler? 

 

3 hours ago, lots of unexplainable lag said:

Honestly, a Vega 64 at that price is just retarded. 1080 performance at 115W more power draw. It's a repeat of the Fury launch where the Fury was a way more sensible buy over a Fury X.

 

The Vega 56 is a way better buy in every way, shape and form, IF the leaked performance we've seen is actually accurate and it crushes a 1070. Otherwise that should also not exist at $399.

The Fury can be flashed to unlock more CU's and better overclocking making it a better card then the Fury X, though because of binning the Fury X reaches slightly higher maximum core clocks. 

 

3 hours ago, Humbug said:

Yep, it seems like Vega56 is closer to the optimal point for the Vega architecture.

Compared to the Vega64 it takes far less power, but only gives up a little bit of performance.

They also probably cut the HBM voltage seeing as that clock drops quite a bit. 

[Mihle]

Price/Perf will be much better at the 56 anyway. With OC its probably 96% of the pref of the 64.

[dizmo]

12 minutes ago, Bananasplit_00 said:

Get a Fury X and a broken RX or R9 GPU for its cooler?

I'd basically have to find a Nano for it's cooler...it's the only card that's the same size and iirc they're based on the same thing.

[DocSwag]

Vega 56 launches August 28th?

 

Holy crap yes. I'm gonna be on a trip for 2 weeks soon and this means I'll be able to order it first thing without miners getting in the way.

 

Bad for you guys but good for me  

[Coaxialgamer]

I have to say, I'm both impressed and disappointed by vega at the same time. Impressed by the cost of the actual product and the  technology inside but disappointed by the huge 484mm² die that can barely compete with a 314mm2 die on an inferior process, while using 2x the power. 

[Coaxialgamer]

15 hours ago, rrubberr said:

Well, here in the OpenCL compute space, they might as well charge 1080ti prices for it.

Depends on what volta can do. Nvidia's "tensor cores" are just glorified vector processors for mixed precision (mainly fp16 I think) , marketed for AI and deep learning. If used for compute, they could drastically improve performance. 

[DXMember]

11 hours ago, Morgan MLGman said:

What worries me the most is that this might make Nvidia be like Intel = stop innovating. At least so far Nvidia has kept releasing new generations that are an actual improvement over previous ones, let's hope it doesn't change.

They cut out double precision to keep down the power consumption so that they can clock higher once

then they changed the process node and clocked the core higher again

and then they said single SLI connecter is not strong enough, so they raised the clock speed of the connector, then said it's still not strong enough and then they used two of them and said - oh and btw, no Quad-SLI

What innovation are you talking about? So far the only thing they have innovated on is "Tensor" cores on V100, at which point they might as well be producing ASIC's...

 

XDMA, Bringing HBM to consumers, implementing HBCC, not to mention SSG on the professional front - that's innovation

[captain cactus]

Article is offline now. Either AMD's not happy and forced them to or they found the info was wrong and they pulled the article from the side themselves.

[Morgan MLGman]

1 hour ago, DXMember said:

They cut out double precision to keep down the power consumption so that they can clock higher once

then they changed the process node and clocked the core higher again

and then they said single SLI connecter is not strong enough, so they raised the clock speed of the connector, then said it's still not strong enough and then they used two of them and said - oh and btw, no Quad-SLI

What innovation are you talking about? So far the only thing they have innovated on is "Tensor" cores on V100, at which point they might as well be producing ASIC's...

 

XDMA, Bringing HBM to consumers, implementing HBCC, not to mention SSG on the professional front - that's innovation

Hmmmm, maybe I used a wrong word here -> they at least have significant performance gains generation per generation (pricing is another matter, it's retarded) instead of doing what happened with Haswell->Broadwell and then Skylake->Kaby Lake

[Humbug]

Reviewers have just got the proper drivers with additional features like TBR, power savings etc enabled. Below is Vega64 (not 56).

 

Performance now is looking comparable to the aftermarket GTX 1080 designs.

 

 

So Vega FE owners can expect a handy little boost once Vega RX launches and they receive these drivers too. Currently Vega FE gaming performance is lower than GTX 1080 founders edition.

[DXMember]

7 hours ago, Morgan MLGman said:

Hmmmm, maybe I used a wrong word here -> they at least have significant performance gains generation per generation (pricing is another matter, it's retarded) instead of doing what happened with Haswell->Broadwell and then Skylake->Kaby Lake

Well intel does pretty much the same, they bump clock speeds - they just don't have a headroom to do it by 50% every time, and all things considered, nVidia is running out of that headroom for SIMD, or at least they should be pretty soon, well at least in theory...

the IPC is pretty much negligible past Sandy Bridge with the exception of some instruction sets in certain situations where Haswell has a notable difference.

The only major thing Intel did is move from Ring Bus to Mesh on Skylake-X and redesigning cache architecture, but that's hardly innovation...

[DXMember]

1 hour ago, Humbug said:

Reviewers have just got the proper drivers with additional features like TBR, power savings etc enabled. Below is Vega64 (not 56).

 

Performance now is looking comparable to the aftermarket GTX 1080 designs.

 

So Vega FE owners can expect a handy little boost once Vega RX launches and they receive these drivers too. Currently Vega FE gaming performance is lower than GTX 1080 founders edition.

damn, if we can unlock all or part of the shaders on 56 and if it clocks to those sick 1750 with like a Sapphire Tri-X cooler, that'd be dope yo!

 

though the results seem interesting... at least judging by the Fury X results posted on:

Fire Strike is only 25.7% gain for Vega 64 with 55.2% clock advantage

something real wrong is some of that - how do you get such a bad scaling?

Could it be the lower memory throughput, higher memory latency?

[Agosto]

12 hours ago, Coaxialgamer said:

I have to say, I'm both impressed and disappointed by vega at the same time. Impressed by the cost of the actual product and the  technology inside but disappointed by the huge 484mm² die that can barely compete with a 314mm2 die on an inferior process, while using 2x the power. 

TSMC's 16nm is NOT inferior to GloFo's 14nm, it's the opposite. GloFo has a little bit denser process but 16nm finfet from TSMC is better for high power applications.

[SteveGrabowski0]

On 8/11/2017 at 0:17 AM, dizmo said:

It's a little better though, no?

If it's 20% better, I'm still all over it if it's what the Nano is based on.

If it's par, then the 1070 wins because of it's much lower TDP.

 

I like the Fury X, just can't use it because of the cooler. I considered the R9 Nano, but want more VRAM.

If it's 20% better then it's a 210W 1080 for $400. It seems pretty unrealistic considering Vega64 is rumored to trade blows with the 1080. So how does the cutdown 210W card basically equal a 1080? Techpowerup has the 1080 as 21% more powerful than the 1070 at 2560x1440 in their big testsuite of games.

[Doobeedoo]

Odd but more interested to see 64 one. 

[The Benjamins]

 

 

 

 

[Coaxialgamer]

16 hours ago, Agost said:

TSMC's 16nm is NOT inferior to GloFo's 14nm, it's the opposite. GloFo has a little bit denser process but 16nm finfet from TSMC is better for high power applications.

Tsmc's process is less dense, which is (arguably) the most important thing for gpus. Also, the whole "high power" thing is BS. Both 16FF and 14nm LPP were developed as mobile-first process technologies, and GF's process has been seen in very high performance products, such as ryzen ( which is capable of 4ghz). 

Tsmc's tech hasn't proven itself as a high performance process because nobody has made >3ghz parts out of it. 

And given that neither company publishes transistor performance data, I have a hard time imagining how you'd know which process performs better. 

[leadeater]

5 hours ago, Coaxialgamer said:

Tsmc's process is less dense, which is (arguably) the most important thing for gpus. Also, the whole "high power" thing is BS. Both 16FF and 14nm LPP were developed as mobile-first process technologies, and GF's process has been seen in very high performance products, such as ryzen ( which is capable of 4ghz). 

Tsmc's tech hasn't proven itself as a high performance process because nobody has made >3ghz parts out of it. 

And given that neither company publishes transistor performance data, I have a hard time imagining how you'd know which process performs better. 

Couple of things though:

• High performance doesn't mean high clock rate, math/compute wise look no further than CPU vs GPU
• Ryzen or more correctly Zen has a very obvious fabrication process limitation with clock rate scaling to power draw and voltage requirement, unlike Intel's process.
• Actual GPU design has more to do with it than what the initial fabrication process was designed for, it is a factor though

 

It's almost impossible to say who's process is better without manufacturing the same GPU design with any modifications that are required for each fabrication process.

 

High clock rates can come from having more stages in an instruction pipeline and having each stage be less complex circuitry wise which can enable a CPU to run at a higher clock rate. CPUs have more stages than a GPU does (I think) along with various different instruction sets to accelerate certain code paths which is one of the reasons why CPUs clock higher. Having more stages does have down sides though, it doesn't reduce the instruction latency and may in fact increase it due to overhead of more stages and pipeline stalls.

 

We do however need to be very careful in comparing CPUs to GPUs since they are very different in design and architecturally.

 

Here is a nice chart of Intel's pipeline stage count over architecture generations though.

 

Quote

Year	Microarchitecture	Pipeline stages	max. Clock
1989	486 (80486)	3	100 MHz
1993	P5 (Pentium)	5	300 MHz
1995	P6 (Pentium Pro; later Pentium II)	14 (17 with load & store/retire) [further explanation needed]	450 MHz
1999	P6 (Pentium III)	12 (15 with load & store/retire)	450–1400 MHz
2000	NetBurst (Pentium 4)	20 unified with branch prediction	800–3466 MHz
2003	Pentium M	10 (12 with fetch/retire) [further explanation needed]	400–2133 MHz
2004	Prescott	31 unified with branch prediction	4000 MHz
2006	Intel Core	12 (14 with fetch/retire)	3333 MHz
2008	Nehalem	20 unified (14 without miss prediction)	3600 MHz
2008	Bonnell	16 (20 with prediction miss)	2100 MHz
2011	Sandy Bridge	14 (16 with fetch/retire)	4000 MHz
2013	Silvermont	14-17 (16-19 with fetch/retire)	2670 MHz
2013	Haswell	14 (16 with fetch/retire)	4400 MHz
2015	Skylake	14 (16 with fetch/retire)	4200 MHz
2016	Goldmont	20 unified with branch prediction	3500 MHz
2016	Kaby Lake	14 (16 with fetch/retire)	4500 MHz
2017	Coffee Lake	14	5000 MHz
2017	Cannonlake	14	5000 MHz

 https://en.wikipedia.org/wiki/List_of_Intel_CPU_microarchitectures

[cj09beira]

On 8/12/2017 at 7:26 AM, DXMember said:

damn, if we can unlock all or part of the shaders on 56 and if it clocks to those sick 1750 with like a Sapphire Tri-X cooler, that'd be dope yo!

 

though the results seem interesting... at least judging by the Fury X results posted on:

Fire Strike is only 25.7% gain for Vega 64 with 55.2% clock advantage

something real wrong is some of that - how do you get such a bad scaling?

Could it be the lower memory throughput, higher memory latency?

ever since vega specs were leaked i though that it would be a bit memory limited, and according to buildzoid it is, (he said for 3% mem oc you got 1% perf, but he didn't say it that was at stock or if it was while the core was overclocked), ever since the 290x amd hasn't increased the ROP count which is a bit weird, as since then amd has gone from 2800 CUs to 4096, which means that fury x's, and more so vega's perf are very dependent on compute effects, thats why on most games fury vs fury x is almost the same perf (around 5% delta) except that on doom the delta increases to 8.5% at 1080p and 9,3% at 4k which seems to support my thesis (in raise of the tomb raider the difference goes down to 1-3%)

[Coaxialgamer]

9 hours ago, leadeater said:

Couple of things though:

• High performance doesn't mean high clock rate, math/compute wise look no further than CPU vs GPU
• Ryzen or more correctly Zen has a very obvious fabrication process limitation with clock rate scaling to power draw and voltage requirement, unlike Intel's process.
• Actual GPU design has more to do with it than what the initial fabrication process was designed for, it is a factor though

 

It's almost impossible to say who's process is better without manufacturing the same GPU design with any modifications that are required for each fabrication process.

 

High clock rates can come from having more stages in an instruction pipeline and having each stage be less complex circuitry wise which can enable a CPU to run at a higher clock rate. CPUs have more stages than a GPU does (I think) along with various different instruction sets to accelerate certain code paths which is one of the reasons why CPUs clock higher. Having more stages does have down sides though, it doesn't reduce the instruction latency and may in fact increase it due to overhead of more stages and pipeline stalls.

 

We do however need to be very careful in comparing CPUs to GPUs since they are very different in design and architecturally.

 

Here is a nice chart of Intel's pipeline stage count over architecture generations though.

 

 https://en.wikipedia.org/wiki/List_of_Intel_CPU_microarchitectures

I know this, but generally "performance" when referring to process technology means transistor switching speed, which is not known for either tech. 

And because we do not know the pipe depth of pascal (maxwell has 6 stages, i don't know if that is everything or just execution though) and polaris, along with the number of transistor states per cycle, we cannot make assumptions about which tech is faster. 

 

I wouldn't even be certain that zen's clock speed wall is a result of process limitations. It could be a result of the high transistor density of that design ( nearly 5 billion transistors in a sub 200mm² die) 

 

However, we do have the gtx 1050 series. 

Unlike the rest of pascal, gp107 was made using Samsung’s 14nm LPP tech, the same as polaris and zen. 

And we find that it pretty much reaches very similar clocks, when considering that it likely has a different voltage curve ( to stay within TDP , which is more important than on the 1060 and up) 

and has a very small die. 

[handymanshandle]

On 8/11/2017 at 0:00 AM, mach said:

Just put them on the shelves already.....

It's not like they are releasing anything groundbreaking that deserves all the teasing

 

I had some hopes for Vega but even what AMD is doing doesnt seem to strike confidence in Vega being anything really all that amazing. I mean the paper specs looked promising but that's only part of the story of course.

 

I don't think it's going to be a shitty card but I think it's going to be rough for people to maybe justify getting either version if the price isnt right for the performance. If they aren't confident in their more expensive card that is suppose to slot in between the 1080 and 1080ti as being a good middle ground then that does say something about how they felt about this release, possibly rushing it. Here's to hoping maybe Navi will have the spirit of Ryzen and put AMD ahead of Nvidia. Now they did do what they said they wanted to do with Ryzen. They made Intel have to re-evaluate stuff. And honestly the i5 X and i7 X and i9s they launched was a knee jerk reaction.

[Guest]

The fact they believe that the Vega 56 is the only card thats gonna captivate the most audience is so disappointing.