AMD Vega Discussion Thread

[xkm1948]

I am hoping for the best. That is all I am willing to say regarding VEGA now.


You guys remember this before Fury X release? We all know how that release ended in the end. Fury X did match Titan X Maswell after a year, only at 4K resolution though.

 

[john_]

My prediction based upon the latest 3 yrs of AMD GPU release suggests that VEGA might not meet my expectations of a 80Ti killer.

I would suggest you be more cautious GT90, some extreme fanboys might be super insulted right now.

I don't think Vega was meant to be a 1080 Ti killer. AMD started talking about Vega many months ago. Nvidia moved on in this time period, from GTX 1080 to 1080 Ti. While fans expectations can easily jump from a GTX 1080 killer to a GTX 1080 Ti killer in milliseconds, I doubt AMD has the potential to redesign Vega to offer 30% higher performance just in a few months.

 

[xkm1948]

I don't think Vega was meant to be a 1080 Ti killer. AMD started talking about Vega many months ago. Nvidia moved on in this time period, from GTX 1080 to 1080 Ti. While fans expectations can easily jump from a GTX 1080 killer to a GTX 1080 Ti killer in milliseconds, I doubt AMD has the potential to redesign Vega to offer 30% higher performance just in a few months.

1060 did force them to raise 480 performance by pumping extra MHz into it. Although that extra MHz definitely hurted Polaris' power efficiency.

If VEGA is capable, i would expect AMD to pump it as high as they can then slam another AIO on it. I am hoping for 1400~1500. Then it will at least leap over 1080.

 

[FordGT90Concept]

I mean, lets be realistic, when you change so many things in the rendering pipeline of the graphic card and still somehow yield same performance as R9 Fury X, things just don't add up for the leaks.

Bare in mind that if this is what will be Vega Nano, we don't know how many watts they're targeting. If it's only 150 watts and outperforms a water cooled Fury X, that's pretty damn good. It also suggests there's a lot of room for a 250w card. We just don't know yet.

 

[alucasa]

Bare in mind that if this is what will be Vega Nano, we don't know how many watts they're targeting. If it's only 150 watts and outperforms a water cooled Fury X, that's pretty damn good. It also suggests there's a lot of room for a 250w card. We just don't know yet.

Something to note though. I haven't seen AMD winning against the Green mob in energy efficiency for a while.

 

[RejZoR]

Well, the Mi25's are PASSIVE cooled. And while I believe professional cards have the highest binned chips, that's still an indication. There is NO WAY you could run Fury X passive. I don't think even with 14nm production node...

 

[W1zzard]

Well, the Mi25's are PASSIVE cooled

Like all similar datacenter cards they require tons of airflow in the case they are installed in

 

[xkm1948]

Navi will be the first from the ground up new design GPU since Raja's return to AMD from Apple. Fiji, Polaris and part of Vega were all on the roadmap before Raja's return and the spin off of RTG. Raja and his new RTG team did their best to improve Fiji/Polaris/Vega. It is up to Navi to pitch AMD back in the high end game.

Like all similar datacenter cards they require tons of airflow in the rack case they are installed in

Holy shit W1zzard replying to a rumor thread? What is happening? You know something all known prophet?

 

[cdawall]

Well, the Mi25's are PASSIVE cooled. And while I believe professional cards have the highest binned chips, that's still an indication. There is NO WAY you could run Fury X passive. I don't think even with 14nm production node...

Yea like w1z said that is far from passive systems...those coolers are rated for 300+ watts.

 

[RejZoR]

@W1zzard
I'm aware of that, but essentially, that's a blower style design. And we know how far you can push those (hint, not far even at extreme airflows).

 

[FordGT90Concept]

Those fans are high pressure fans (not blowers) meant for ramming a lot of air through tight fins, noisily. Their cooling capacity far exceeds the capabilities of any all-in-one air cooling solution.

I can't tell how big they are from the picture but they look like 80mm or 92mm and likely run north of 6000 RPM. Think Vantec Tornado.

 

[alucasa]

Those fans run at like 9,000 RPM.

 

[Vya Domus]

R9 Nano was a dumb idea , in the sense that it was clearly an afterthought and Fiji did not fit that role. It throttled like mad and the only thing that kept performance afloat was the huge stream processor count . If they do have another Nano in the works I'm hoping that means Vega has proper power efficiency and runs cool enough.

 

[xkm1948]

R9 Nano was a dumb idea , in the sense that it was clearly an afterthought and Fiji did not fit that role. It throttled like mad and the only thing that kept performance afloat was the huge stream processor count . If they do have another Nano in the works I'm hoping that means Vega has proper power efficiency and runs cool enough.

Fiji was never supposed to run ~1050MHz. The sweet spot in terms of power efficiency for 28nm Fiji is actually around 800~900MHz. The release of 980Ti forced AMD's hand so they had to up the game a bit, leaving the FuryX with only 100~150MHz overclock headroom. I have tried this myself. Undervolting Fiji and keep the speed around 1GHz actually reduces power consumption greatly(roughly 25% less power consumption in FireStrike Ultra, 1GHz FijiXT core with -25mV undervolt)

 

[FordGT90Concept]

Something to note though. I haven't seen AMD winning against the Green mob in energy efficiency for a while.

R9 Nano was a dumb idea , in the sense that it was clearly an afterthought and Fiji did not fit that role. It throttled like mad and the only thing that kept performance afloat was the huge stream processor count . If they do have another Nano in the works I'm hoping that means Vega has proper power efficiency and runs cool enough.

Vega is mainly about efficiency. AMD is losing mindshare because of the reputation for being hot and loud. Getting thermals under control was a higher priority for Vega than taking the performance crown. A lot of the architectural changes in Vega are expressly for the purpose of improving efficiency.

 

[alucasa]

I hope so. Though from what I am reading around here, Vega seems to be about beating Green Titan? Or is it just wishful thinking from fanboyz?

The one thing I learned from RX 500 series was that AMD couldn't care less about power efficiency.

 

[xkm1948]

I hope so. Though from what I am reading around here, Vega seems to be about beating Green Titan? Or is it just wishful thinking from fanboyz?

The one thing I learned from RX 500 series was that AMD couldn't care less about power efficiency.

They can't emphasize on efficiency when performance is not so good.

 

[john_]

1060 did force them to raise 480 performance by pumping extra MHz into it. Although that extra MHz definitely hurted Polaris' power efficiency.

If VEGA is capable, i would expect AMD to pump it as high as they can then slam another AIO on it. I am hoping for 1400~1500. Then it will at least leap over 1080.

AMD is behind in performance, so they push their cards as much as they can. This isn't something new. I can think of Cyrix processors many many years ago being factory overclocked, trying to catch up with Intel's first Pentiums. But pushing a design to the max, does destroy any power efficiency.

As for 480, AMD wasn't trying to catch up with GTX 1060, but beat GTX 970. And also they wanted to push the card as much as necessary to get the "VR Ready" sticker. Without that sticker, all their marketing efforts, about the cheapest VR Ready card, would be meaningless.

 

[eidairaman1]

Like all similar datacenter cards they require tons of airflow in the case they are installed in

Do they make universal fan mounts for ATX cases like the ones aft?

 

[FordGT90Concept]

The one thing I learned from RX 500 series was that AMD couldn't care less about power efficiency.

Polaris will never match Maxwell/Pascal in efficiency because it lacks tiled rasterizing capability. RX 580 is about exploiting improvements at GloFo.

 

[alucasa]

I am no longer tied up with CUDA, so AMD GPU becomes an option finally. The last ATI/AMD GPU I had was IceQ 4770.

So, yeah, it's been a while. Though, since my 1060 isn't even a year old and I tend to skip a gen, my next GPU will be after a while.

 

[dorsetknob]

This isn't something new. I can think of Cyrix processors many many years ago being factory overclocked, trying to catch up with Intel's first Pentiums.

In the world of x86 poker Cyrix lost out when Intel went pentium
They Struggled to compensate for FPU with Software ( Intel would not licence the IP for hardware inclusion).
Increase in Pentium Clockspeed with inc FPU ment Cyrix could not compete
AMD could and did Compete as they had IP they could Cross licence with Intel
So No Seat at the x86-64 Poker table for cyrix

 

[xkm1948]

OpenCL is great once you have adjusted to the learning curve. Surprisingly adapting CUDA based programs to OpenCL is not as hard as I first thought. I must say Nvidia offers excellent tech support for their CUDA.

When it works, the FijiXT core with 4096SP is truly a monster for scientific computing. I used it for some Protein sequence homolgy search(Adapted from NCBI BLASTP suite) and boy that thing destroyed the pair of Tesla K40 we have on our University's HPC. That was a good day, I can still remember our HPC mananger's facial expression after I demoed how a consumer grade FuryX beat a pair of his expensive K40.

Anyway I will be skipping Vega. This FijiXT will have to work until at least Navi/Volta is around.

 

[HD64G]

OpenCL is great once you have adjusted to the learning curve. Surprisingly adapting CUDA based programs to OpenCL is not as hard as I first thought. I must say Nvidia offers excellent tech support for their CUDA.

When it works, the FijiXT core with 4096SP is truly a monster for scientific computing. I used it for some Protein sequence homolgy search(Adapted from NCBI BLASTP suite) and boy that thing destroyed the pair of Tesla K40 we have on our University's HPC. That was a good day, I can still remember our HPC mananger's facial expression after I demoed how a consumer grade FuryX beat a pair of his expensive K40.

Anyway I will be skipping Vega. This FijiXT will have to work until at least Navi/Volta is around.

What you said there means Fiji is a monster in raw power. That monster was mainly held back by software and by its arch not being best for DX9-11 and lower res than 4K due to single CPU thread usage. Agreed. By that logic, if Vega's arch is changed from Fiji's in order not to depend its being used properly by the software should be performing very well. Especially in 4K gaming that shows raw power more clearly as Fiji's position in 4K chart shows.

 

[FordGT90Concept]

Fury X is 8.6 TFLOPs versus Vega's ~12.5. Tesla K40 is 4.29 TFLOP. GCN has mostly had an emphasis on compute performance (think Bitmining) which is why AMD's high TFLOP numbers don't lie. I think Vega will continue that trend but the new features emphasize gaming efficiency.

Navi...I think Navi is about GPU virtualization. GPU workloads are generally scalable by nature. There's no reason why SLI/Crossfire has to exist. Simply install a number of GPUs and through hardware abstraction, make it look like a single big GPU with lots of compute units to software. I think that's what Navi does. It not only works for multiple GPUs but it also works for customizing APUs and custom solutions (e.g. compute-only cards). I think Navi will take GCN and add a networked scheduler on top. If that networked schedule finds other network schedules, they buddy up. NVIDIA implemented something similar with their NVLINK.