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

Compared to the V64, there quite a bit more V56 for the lowest advertised price without the bundle. And over the days following the launch, there were reinforcements shipping in, although demand probably still vastly outstripped supply. Additionally, with V56 it was already clear that there was limited supply of them - which AMD did not make clear before the V64 launch.

Some people still seem to be frustrated about that, but I think we should put this case to rest, hoping that everyone will be acting a bit smarter in the future.
 
Last edited:
Sure. But if you're an enthusiast looking for a good value the purchasing experience around Vega is frustrating and that's NOT good for AMD and, even though they at least have a product offering now, they are still effectively absent from this market for many people. If I had confidence that there would be custom Vega 56s readily available @ $399 within a reasonable period of time I would have considered waiting to purchase one of those over the custom GTX 1080 I actually purchased @ $499.

Clearly, there are people who will take any and all opportunities to present AMD's efforts in a negative light (and the reverse, of course), but this situation is not great for them, even if it does benefit them financially in the short term.

Vega's biggest problem isn't performance, its that its a year later then it should be. That was caused by not funding RTG enough ( looks to be the right decision given how zen turned out) , for too long AMD milked GCN ( Tahiti, Hawaii , Fiji) without really driving the overall uarch forward, especially when you consider what NV did during those generations. The tech level of Fury should have been polaris and the tech level of polaris should have been Vega.

Given that Vega seems to require quite a lot of diver team work this just adds insult to injury because its not even getting shown in an ideal light on top of being 1 year after the competition. Now over the next 1-2 years as AMD take cpu market share its time for AMD to give GPU uarch more of a fair share.
 
Compared to the V64, there quite a bit more V56 for the lowest advertised price without the bundle. And over the days following the launch, there were reinforcement shipping in, although demand probably still vastly outstripped supply. Additionally, with V56 it was already clear that there was limited supply of them - which AMD did not make clear before the V64 launch.

Some people still seem to be frustrated about that, but I think we should put this case to rest, hoping that everyone will be acting a bit smarter in the future.

Agreed. And I do believe more effort was made to enable V56 to be available @ MSRP for a longer period of time. The situation is what it is.
 
(The hardware review thread is getting too cluttered so decided to post this here.)

So after a week of Vega 56 ownership and tinkering with the wattman settings and looking around the forums, my takeaway is that Vega has very good potential for AMD. As in, it can hit the clockspeeds with more volts while Fiji just guzzled more power and again unlike Fiji it's not spending its time near the boost clocks given in Wattman at stock, which means that the card can hit those clocks with tweaking and has lots of headroom to work with.

In firestrike, the card does 19-19.5k(graphics) at stock balanced setting, I can get 2k over it, while some people are doing almost 24k(!) with the 64bios. So as I said the potential is there, AMD can do a far better job of silicon revisions here than the hopeless 480->580 transition, combine that with faster HBM2 and driver maturity and I can see the Vega rerelease on 14nm+ doing 20% or so better than what AMD have managed with current Vega. But how well and quickly can AMD execute this? Not much time till Volta and the new Vega might have some arch. changes as well if the 1/2DP rumor pans out.
 
If I had confidence that there would be custom Vega 56s readily available @ $399 within a reasonable period of time I would have considered waiting to purchase one of those over the custom GTX 1080 I actually purchased @ $499.
Yeah, I've been waiting for custom vegas to appear since mid-june, and it's now three months later with no news whatsoever as to when they'll appear. If I wasn't such a damn stubborn nerd I would just have just bought a 1080 and been a happy camper with super performance at a low power price. However, I'll hold out, and I'll keep waiting and hoping for that magic driver that will finally turn on the MIA bits of Vega's die and turn it into a 1080Ti competitor, the damn fool that I am...
 
Oddly enough, when binning was added for Vega10 in Linux drivers they did state it was tuned for Raven. Which is odd being that Vega10 is out and Raven not currently.
Has anyone compiled versions of the driver with and without binning included and run some performance tests?
 
Oddly enough, when binning was added for Vega10 in Linux drivers they did state it was tuned for Raven. Which is odd being that Vega10 is out and Raven not currently.
It was stated elsewhere that the DSBR would tend to benefit resource-constrained implementations more. Vega 10 has a surplus of resources relative to an APU, and with clear signs of not utilizing what it has.
The marketing gives it modest performance and efficiency benefit, and measurable bandwidth savings.

It's going to matter a lot more for Raven Ridge, and aside from other scenarios as to why Vega 10 has had teething pains it is also the case that Raven Ridge is very, very important for AMD in the near term in ways Vega 10 is not.
 
Has anyone compiled versions of the driver with and without binning included and run some performance tests?
Not that I'm aware as it's in the master branch. There are quite a few TODOs remaining. Been keeping an eye out for tests. If I manage to find a Vega I'd test it.

It was stated elsewhere that the DSBR would tend to benefit resource-constrained implementations more. Vega 10 has a surplus of resources relative to an APU, and with clear signs of not utilizing what it has.
If that resource is a bottleneck it would apply. At the very least it's less power and cache thrashing. I'd expect cache to have a significant impact and DSBR should improve that. I'd agree Vega isn't fully utilizing something and DSBR should help even if indirectly.

It's going to matter a lot more for Raven Ridge, and aside from other scenarios as to why Vega 10 has had teething pains it is also the case that Raven Ridge is very, very important for AMD in the near term in ways Vega 10 is not.
I'd agree Raven is more important for graphics. Vega10 still more significant for compute and Raven graphics aren't here yet. That appears to be how AMD prioritized development.
 
If that resource is a bottleneck it would apply. At the very least it's less power and cache thrashing. I'd expect cache to have a significant impact and DSBR should improve that. I'd agree Vega isn't fully utilizing something and DSBR should help even if indirectly.

In Gamers Nexus interview, Mike Mantor confirmed DSBR will improve cache latencies and effective bandwidth. See it at 9m29s.
 
If that resource is a bottleneck it would apply. At the very least it's less power and cache thrashing. I'd expect cache to have a significant impact and DSBR should improve that. I'd agree Vega isn't fully utilizing something and DSBR should help even if indirectly.
Vega 10 has much more in an absolute sense, be it shader engines, clocks, CUs, bandwidth, cache, power budget, better host system, etc. It has more to compensate for problems, and more slack for things like power or sub-optimal software.

Raven Ridge is likely to face crippling bandwidth and power limits and interference from a CPU. It's operating 1-2 orders of magnitude down on many things Vega 10 takes for granted.

I'd agree Raven is more important for graphics. Vega10 still more significant for compute and Raven graphics aren't here yet. That appears to be how AMD prioritized development.
APUs are strictly more important for AMD in general. AMD's non-presence in compute means Vega is a growth project.
Mobile x86 volumes and revenues are very important for AMD overall, probably more so than all discretes.
If the GPU caused AMD to miss the cycle again for that market, that is something that would get executives in trouble.
 
APUs are strictly more important for AMD in general. AMD's non-presence in compute means Vega is a growth project.
That's what I'm saying. Discrete Vega for deep learning, HPC, and SSG would still be high margin markets to focus on with graphics secondary. The APU needed to sell CPUs, however they always seem to be significantly lagging. We just saw Bristol Ridge arrive after Vega was out.

If the GPU caused AMD to miss the cycle again for that market, that is something that would get executives in trouble.
It would, but as Bristol just released, shouldn't that have been the APU focus? Seems late if Raven is coming, so inline with past schedules. As DSBR was tuned for Raven, it doesn't appear it's that far behind schedule. Especially as it's tied to Zen2/Zen+ or whatever variation that worked out to. More likely the CPU side was dragging a bit, but we may lack insight to those schedules. Still seems strange to launch Bristol with Raven shortly thereafter.
 
It would, but as Bristol just released, shouldn't that have been the APU focus?
Bristol has been available to OEMs since last year. This year's release is a launch for retail, and those are not in the form factor or the TDP range AMD has a serious interest in.
AMD needed something for the release cycle.

As DSBR was tuned for Raven, it doesn't appear it's that far behind schedule. Especially as it's tied to Zen2/Zen+ or whatever variation that worked out to. More likely the CPU side was dragging a bit, but we may lack insight to those schedules.
Ryzen had issues, but it ironed out most of them in a few months while being almost totally functional and sufficient for client computing. The x86 side should be generally established, and it had work going into it even before manufacturing.
The GPU is the one going over half a year out from rumored samples with significant chunks of the design document being MIA.

Still seems strange to launch Bristol with Raven shortly thereafter.
AMD has to cater to the market it sells to, and that means "launching" what they have available based on the deadlines set by the market.
 
Ryzen had issues, but it ironed out most of them in a few months while being almost totally functional and sufficient for client computing. The x86 side should be generally established, and it had work going into it even before manufacturing.
The GPU is the one going over half a year out from rumored samples with significant chunks of the design document being MIA.
I'm saying Ryzen seems based on a refined Zen and not Vega core. Raven being Zen+ with whatever improved. Probably something with Infinity design that Vega also incorporated. The smaller Vegas are coming later and they may have adopted the changes transparently.
 
I'm saying Ryzen seems based on a refined Zen and not Vega core. Raven being Zen+ with whatever improved. Probably something with Infinity design that Vega also incorporated. The smaller Vegas are coming later and they may have adopted the changes transparently.
Probly they improve zen performance with lower freq. memories like they will have in APUs.
 
I'm saying Ryzen seems based on a refined Zen and not Vega core. Raven being Zen+ with whatever improved. Probably something with Infinity design that Vega also incorporated. The smaller Vegas are coming later and they may have adopted the changes transparently.
AMD's latest projections have Zen, Zen 2, and Zen 3. The Zen 2 block would be uncomfortably far in the future.

Zen+, or what AMD used to call it, doesn't seem to rate as a significantly different core. Zen seems to count for 14nm and 14nm+, whatever that means, but the tweaks in that case may have more to do with implementation details for the device envelope and process sharing with a GPU, or a "refresh" composed of process maturation and firmware tweaks.

I would say that Ryzen and EPYC in particular have utilized the Infinity Fabric in more clear ways than Vega 10 has.
Raven Ridge might be served by an APU-specific crossbar or blocks specific to APU usage, but what about Vega 10's implementation seems useful?
 
Raven Ridge might be served by an APU-specific crossbar or blocks specific to APU usage, but what about Vega 10's implementation seems useful?
I'm not sure with Vega10. Maybe an external path between CUs that doesn't get used for graphics? Shared cache or registers. A dual Vega or extra lanes for onboard controllers for pro products taking up space perhaps?

I'd still think 32 PCIe lanes makes sense. Given SSG with 4x 512GB NVMe drives. On a consumer part that would be useless without boards with huge slots. Maybe embedded or APUs benefit more. That probably costs some die size that isn't generally useful. Helps compute, not consumer at least.
 
Back
Top