AMD Radeon VII Announcement and Discussion

[yuri]

Mkay, they can salvage the broken VEGA 20 chips in this Vega and the MI50. That's fine.

I just gotta giggle at all those "first GAMING 7nm GPU" labels - the main new features of the chip are half-speed FP64, new AI instructions, full-chip ECC and xGMI interface.

The rest is just another Vega... TDP == VEGA 10 and frequency is "up to 1.8GHz".

 

[Esrever]

Radeon VII looks really good for machine learning enthusiasts due to price and 16GB of HBM, even without tensor cores a lot of workflows are very memory limited. I wonder if AMD's efforts on that end have improved to the point of usability. They have a lot of support for machine learning frameworks now but I have no idea if their ease of use is up on par yet.

I will see what reviews say when it's out, I'm sure someone will try to make it work.

 

[Geeforcer]

So safe to say someone got AdoredTV pretty good, right?

 

[Tkumpathenurple]

AMD still seems to be behind in terms of architectural performance. Their 7nm chip is matching a 2080(12nm). Nividia 7nm's chips will probably leap well ahead again...

Maybe so, but does that really matter?

If they can achieve greater performance per dollar, they remain competitive. And they do so using a graphics card that's a 7nm pipe cleaner.

Why rush a new architecture to market when you already have a competitive product?

 

[Deleted member 13524]

For starters, twice the bandwidth.

With twice the ROPs.

The question is wtf does the +62% OpenCL perf means given that the Blender (+27%), Davinci (+15%) and Adobe Premier (+29%) are actually all OpenCL based...

Can OpenCL make use of INT8 calculations? Vega 20 brings quad-rate INT8, so twice that of Vega 10 which AFAIK promotes INT8 to FP16.

If they can achieve greater performance per dollar, they remain competitive. And they do so using a graphics card that's a 7nm pipe cleaner.

Yup the Radeon VII as a consumer card looks more like a pipe cleaner for 7nm. Lisa Su mentioned more GPUs later this year so this is the last we'll see of the Vega architecture.


I think AMD is also positioning Vega VII as a Frontier Edition replacement, as they mentioned its value for content production as well as gaming.
My guess is Radeon VII will have the same access to non-certified driver optimizations for professional applications as the Vega FE. This would position it closer to a Titan Xp, or Titan V without the tensor op hardware.
We also don't know if they're disabling 1:2 FP64 throughput, yet. If not, this card could become a favorite for low-cost compute stations in universities and smaller laboratories.

 

[pcchen]

For this price, this is a great card for people who are doing compute and possibly also AI, if you don't need CUDA.
For gamers it's not that bad, if it has similar performance to a 2080. It has more VRAM, and more bandwidth. It might well be more future proof. Of course, it could be slower running ray tracing, but ray tracing is not something "required" by a game. Also we don't know whether ray tracing in a more mature form is going to be fast enough with current cards. Maybe it's something going to be only really usable with next generation cards.
Personally I only got a 2080 TI because I want to be able to play 4k@60fps with one GPU right now. Ray tracing is something that, if workable, nice to have, not a "must have."

 

[Rootax]

Ah, I didn't see the "twice the rop".

 

[yuri]

Are those 128 ROPs really double confirmed? The MI50/60 spec sheets don't cover ROPs.

Btw just a side-note: even this cut-down ~300W TDP consumer oriented 7nm Vega is not "north of 2GHz".

 

[Ike Turner]

You sure about that?
Without raytracing and DLSS, yes, but one could just aswell list exclusive features AMD has. And even stuff that shouldn't be exclusive but still are (within GPU markets) like OpenCL 2.0, NVIDIA still fails to support it.

They have the "luxury" of having shitty OpenCL support because CUDA is everywhere (and their GPUs are in 90% of the devs/cc machines) . A few OpenCL only software packages had to implement CUDA in the past 2 years just to retain their users because Nvidia doesn't give a damn about OCL2. 0 or anything that isn't their proprietary tech.

 

[no-X]

AMD still seems to be behind in terms of architectural performance. Their 7nm chip is matching a 2080(12nm). Nividia 7nm's chips will probably leap well ahead again...

Pardon? What kind of architectural performance are you comparing? If Vega 20 / Radeon VII (13,2 bil. transisotors) performs close to GTX 2080 (13,6 bil. transistors), than I can't see any deficiency on AMD side.

 

[Tkumpathenurple]

Actually same price as their competitor, NVIDIA just asks $100 premium for Founders Edition

My mistake, I thought the 2080 was closer to a grand.

 

[Love_In_Rio]

Pardon? What kind of architectural performance are you comparing? If Vega 20 / Radeon VII (13,2 bil. transisotors) performs close to GTX 2080 (13,6 bil. transistors), than I can't see any deficiency on AMD side.

except for ray tracing and AI cores and tdp (in a generational node difference)...

 

[no-X]

except for ray tracing and AI cores and tdp (in a generational node difference)...

Well, and there are features on AMD side, which Turing doesn't support, e.g. full-spectrum precision from Int4 to FP64 at high performance, HBCC which allows to use datasets larger than VRAM, PCIe 4.0 etc. Anyway, I asked on "architectural performace", not "architectural features".

(please move to proper thread, it was splitted just before I sent my post)

 

[del42sa]

Are those 128 ROPs really double confirmed? The MI50/60 spec sheets don't cover ROPs.

Btw just a side-note: even this cut-down ~300W TDP consumer oriented 7nm Vega is not "north of 2GHz".

nope, I think Anandtech got this wrong...

 

[yuri]

Extreme data science skills here:

Enjoy

 

[Pressure]

Are those 128 ROPs really double confirmed? The MI50/60 spec sheets don't cover ROPs.

Btw just a side-note: even this cut-down ~300W TDP consumer oriented 7nm Vega is not "north of 2GHz".

Board power is not confirmed.

Ryan says 128 ROPs.

 

[Digidi]

Does Primitive Shaders and NGG now work?

 

[gamervivek]

128ROPs is intriguing, if this clocks like RX Vega with rarely hitting the 1.8Ghz boost clock and overclockable beyond with more juice, it could be a good enthusiast card positioned against 2080Ti.

 

[Megadrive1988]

So there's getting 1 TB/s bandwidth from 16GB HBM2, that's kinda interesting on its own.

 

[Kaotik]

Are those 128 ROPs really double confirmed? The MI50/60 spec sheets don't cover ROPs.

Btw just a side-note: even this cut-down ~300W TDP consumer oriented 7nm Vega is not "north of 2GHz".

Not really, GCN ROPs aren't tied to memory controllers. It could be 128, but just because it has double the memory controllers doesn't mean it has to have 'em.