Vega/Navi Rumors (Updated)

[Sweepr]

Vega 10
- 64 CUs / 4096 SPs
- 24 TFLOPs 16-bit (VideoCardz) / 10 TFLOPs single-precision (Techfrag)
- Not as huge (6000 SPs+) as some expected, but will have a 300W dual GPU version to go up against NVIDIA's best next year

Vega 11
- Smaller chip, could replace some Polaris 10 products

Vega 20
- Same number of stream processors as Vega 10 (4096 SPs)
- Same architecture (GFX9) at 7nm
- Lower TDP at 150W

Navi 10/11
- Delayed to 2019

10 TFLOPs (Techfrag) is 16% better than Fury X, 12 TFLOPs (based on 24 TFLOPs 16-bit from VCZ) would be 39%.

Vega 10 will be released in first quarter of 2017, it has 64 Compute Units and 24TF 16-bit computing power. Vega 10 is based on 14nm GFX9 architecture. It comes with 16GB of HBM2 memory with a bandwidth of 512 Gbps. The TBP is currently expected at around 225W. Meanwhile, dual Vega 10 will be released in second quarter of 2017 and TBP should be around 300W.

Now here’s something interesting that we didn’t know yet. Vega20 will use 7nm GFX9 architecture. It will feature 32GB of HBM2 memory with 1 TB/s of bandwidth. TBP is around 150W and it will support PCI-Express 4.0. It will also have 64 Compute Units.

http://techfrag.com/2016/09/16/amd-vega-gpu-ship-h1-2017
http://videocardz.com/63700/exclusive-first-details-about-amd-vega10-and-vega20


January 6, 2017 Update: http://videocardz.com/65521/amd-vega-10-and-vega-20-slides-revealed


February 14, 2017 Update:

The mainstream and performance tier cards will be filled by a second Vega chip. This chip is known as Vega 11 and is the smallest of the two. I am personally more excited about this as it will be aimed at a wider audience due to its cost and design. The Vega 11 graphics core will be replacing the Polaris 10 graphics core as AMD’s mainstream chip so we can expect pricing to be under $300 US.

...Based on the performance tier of this graphics card, we can expect it to go the the traditional memory route. AMD may end up using GDDR5/X memory if they don’t find the prices of HBM2 too feasible or we can see a single stack design. As with Vega 10 which features 8 GB HBM2 on two stacks, Vega 11 might come with a single HBM2 stack that offers 4 GB VRAM at speeds close to 256 GB/s, along a 1024-bit interface. AMD will have multiple options open to them and a single die stack would mean higher yields and less chances of error during production.

http://wccftech.com/amd-radeon-rx-500-vega-10-vega-11-gpus


June/July 2017 Update:

- PCPerspective Radeon Vega Frontier Edition Live Benchmarking

- Reddit Live Stream of Benchmarking

Live testing summary:

• Time Spy Graphics = 6785 pts
• Fire Strike Ultra Graphics = 5091 pts
• Doom Vulkan 4K Ultra @ 55-65 FPS
• The Witcher 3 @ 28-35 FPS at 4K Hairworks on / 41-42 FPS with Hairworks off (80 C running the game)
• Troubles OCing past 1650 MHz with the blower
• Tried to OC HBM, starts at 945 MHz, managed 960 MHz stable and at 980 MHz it got too hot
• Tester says you won't be able to OC without additional cooling, thermal throttling (80-85 C)
• Tester put 2x 80 mm fans to help (open air case) and it was down to 75 C
• Tester doesn't think it will touch the 1080 Ti performance-wise, only 1070-1080s
• Mining: 30-35 MH/s
• Cinebench R15 OpenGL: 97.39 FPS
• Initial testing done using Pro drivers, then switched to the Gaming drivers and performance didn't change
• Card initially operate at 1348-1528 MHz in gaming mode running The Witcher 3 (37-42 FPS), then after retesting it was back to 1650 MHz max (and delivered the same 42 FPS as Pro mode)
• >300W for the card
• EVGA P2 1200W Platinum Power Supply + Ryzen R7 1800X CPU

https://www.reddit.com/r/Amd/comments/6k4jwh/live_vega_testing


- AMD Radeon Vega Frontier pictured and tested

https://videocardz.com/70541/amd-radeon-vega-frontier-lands-in-the-hands-of-the-first-owner

 

[Vaporizer]

Navi delayed to 2019 is surprising since it is a 12 month delay.

 

[JoeRambo]

Looks like a shrink of Fiji on 14nm @ ~1.25Ghz clock with updated GCN architecture from Polaris and 16GB of HBM2 memory?

 

[showb1z]

16GB HBM2 with 512GB/s bandwidth. Right.
Not much effort went into this rumor. D-

 

[Alqoxzt]

Looks like a shrink of Fiji on 14nm @ ~1.25Ghz clock with updated GCN architecture from Polaris and 16GB of HBM2 memory?

it's gfx IP9 while Fiji is ip8 Polaris is ip8 decimal something
Ip9 might bring full features DX12 up to full tiers (up to 3rd) as well as IPC increase.

 

[antihelten]

16GB HBM2 with 512GB/s bandwidth. Right.
Not much effort went into this rumor. D-

Those numbers aren't entirely impossible, they would however require 8Hi stacks*, which I don't think SK Hynix is currently offering (they only offer 4Hi at the moment).

*2 8Hi stacks of HBM2 running at 2Gbps would equal 16GB and 512GB/s.

 

[Alqoxzt]

And dx overhead reduction as well...... just wishful thinking

 

[raghu78]

16GB HBM2 with 512GB/s bandwidth. Right.
Not much effort went into this rumor. D-

Well said. Hynix has just started making 2 Gbps 4 Hi HBM2 chips with 4 GB capacity and 256 GB/s bandwidth. So we are likely to see 8GB and 512 GB/s or 16 GB and 1TB/s.

 

[Unreal123]

So AMD is telling to people that you should buy Titan X or GTX 1080 because VEGA will be around Titan X level of pricing ,however, it will be slower then Titan X and with 1 year late. I hope for the sake of AMD that Volta do not launch in 2017 and if it does then their will be no hope for AMD to get even 1% of high GPU market.


Trolling is not allowed
Markfw900
Anandtech Moderator

 

[Alqoxzt]

Well said. Hynix has just started making 2 Gbps 4 Hi HBM2 chips with 4 GB capacity and 256 GB/s bandwidth. So we are likely to see 8GB and 512 GB/s or 16 GB and 1TB/s.

Wasn't there some cheap Hbm2 news ?

 

[Qwertilot]

Navi delayed to 2019 is surprising since it is a 12 month delay.

If it has to be GloFo (for the well known, if unfortunate, economic reasons) and GloFo have now lost their 10nm process then getting a big/new chip out on their next die shrink probably would be something around 2019.

Whatever happens in 2017, NV will give their range a decent %age price/performance bump. Either Volta or dropping the GP102 stuff down the stack a bit and reorganising. They have to to keep their upgrade business going.

 

[witeken]

So GPU were stuck at 28nm for 4 years, and suddenly we're already talking about 7nm ?

 

[antihelten]

So GPU were stuck at 28nm for 4 years, and suddenly we're already talking about 7nm ?

Well the rumor in question (the videocardz one) covers a period from 2017 to 2019, so it's not really that weird.

 

[Bryf50]

So AMD is telling to people that you should buy Titan X or GTX 1080 because VEGA will be around Titan X level of pricing ,however, it will be slower then Titan X and with 1 year late. I hope for the sake of AMD that Volta do not launch in 2017 and if it does then their will be no hope for AMD to get even 1% of high GPU market.

For better or worse, AMD hasn't told people anything.

 

[Arachnotronic]

For better or worse, AMD hasn't told people anything.

Oh really?

Mark Papermaster

We have shown some slides talking about this progression of the CPU roadmap, but we made a decision four years ago as I said to focus on both CPU and graphics, and you've seen the fruits of those labors; so the eight points of gain is prior actually to launch of our FinFET based Polaris products. Why is that? It wasn't -- the focus when in addition on the new CPU -- on the new graphics engines that Polaris represents was also one the software capability, we've improved the quality and the frequency of our entire software stack around our Radeon products and we've improved our marketing to show where Radeon plays, to show the excellence, and our competitiveness and experience that we provide. So, we build on that eight points of gain that we just experienced with the new -- Polaris products which are ramping nicely again, we've actually sold out, in the beginning it was not a supply constraint, it was frankly sold out based on demand we've now responded to that, we've got Polaris restocked. And so there's the disruptive play with Polaris on that performance end of the market. VEGA is on track as we said for first half of '17 to attack in that high-end of the market, in that highly performance enthusiast end of the market.


Ross Seymore

Is that a completely new architecture, somewhat analogous to Zen or it is in more of an iteration on Polaris?

Mark Papermaster

No, you'll see yet again significant performance per watt changes of improvement, now that's what we will have, we have shown a long range roadmap for our Radeon graphics products and at each generation we will significantly improve the performance per watt. That's what the workloads demand across -- not only gaming but these emerging opportunities I described a moment ago and that's what we'll provide with the roadmap.

http://seekingalpha.com/article/400...e-bank-2016-technology-conference?part=single

 

[Maverick177]

I can't find anything about Vega 10 having 10TF DP in the techfrag article.

 

[Bryf50]

What does that have to do with the specs, price or performance of Vega? You missed my point. Which is obviously that these numbers are rumors and not information from AMD.

 

[Glo.]

So AMD is telling to people that you should buy Titan X or GTX 1080 because VEGA will be around Titan X level of pricing ,however, it will be slower then Titan X and with 1 year late. I hope for the sake of AMD that Volta do not launch in 2017 and if it does then their will be no hope for AMD to get even 1% of high GPU market.

AMD is not telling this to anyone. Its you who made all of this based on rumors. Nothing more, nothing less.

Your post, as a whole should be considered only as trolling, because it has zero-to-none value in terms of substance. Its just all made up, based on nothing, at all.

 

[Sweepr]

AMD is not telling this to anyone. Its you who made all of this based on rumors.

Your post, as a whole should be considered only as trolling, because it has zero-to-none value in terms of substance. Its just all made up, based on nothing, at all.

Confused?

 

[antihelten]

I can't find anything about Vega 10 having 10TF DP in the techfrag article.

No one said anything about 10 TFLOPS DP, the OP lists it as 10 TFLOPS SP.

The only mention of this in the Techfrag article is in the title where it simply says 10 TFLOPS (with no mention of single/double precision). Given a total of 4096 shaders, the 10 TFLOPS number can almost certainly only be single precision (would require a clock frequency of ~1200MHz) .

 

[Headfoot]

Given how bottlenecked Fury X seemed to be with 4096 shaders, a new card with the same amount but with all the plumbing changes of GCN4 (and more Vega specific ones even) will likely have a much different performance profile.

 

[Glo.]

Confused?

There are rumors that Aliens built Egyptian pyramids. Using your logic, the rumors confirm that.

 

[Sweepr]

No, you're just contradicting yourself. If it's ok to come up with very optimistic predictions regarding Polaris/Vega based on 'rumors', different opinions should be allowed.

Back to topic, according to the rumors big Vega launches first, interesting (and exciting if true). Q1-17 as well.

 

[swilli89]

The biggest takeaway I have is this:

We have also learned that AMD is planning to replace Polaris 10 with Vega 11 next year.

If AMD is planning on replacing the Radeon 470/480 not even a year after I think they may have something special with Vega's architetcture. We know Vega will be on the same 14nm process, so if this is indeed true it sounds like Vega may bring another good step down in power efficiency. Why else replace a fairly new SKU on the same process unless the uArch is improved?

I've seen plenty of threads on reddit where an undervolted (maintains stock frequency) 480 matches a 1060 in power efficiency even in DX 11 titles so I bet AMD is really focusing on optimizing around this metric in both uArch and software/firmware tuning.

Other than that I would take the Vega 10 specs with a grain of salt, something doesn't seem right about those memory numbers, OTOH it is a good sign we are hearing Q1 now instead of H1. I imagine a lot of buyers out there would like to try something new with a Zen/Vega combo build!

 

[Headfoot]

My money is on that the increase in efficiency is basically due to HBM, like on Fury, and maybe some due to a slight lowering of clock speed. That and as you increase die size less of the power budget as a percentage goes to non-scalable uncore and other fixed power pieces.

I'd be pleasantly surprised if there is an increase in power efficiency due to even further architectural improvement but I wouldn't bet on it myself.