AMD Vega MegaThread! FAQ and Resources

[Martell1977]

My most optimistic feeling when it comes to 7nm was we will see SoC type chip early 2019. Anything bigger probably won't be coming until mid 2019 at the earliest. Meaning 7nm Vega (meant for their machine learning stuff) won't be coming out until H2 2019. Also it depends how good TSMC and GF refining their 7nm process for bigger chip. Worse we could end up seeing 20nm repeat.

Well. that's not exactly a problem. We know AMD plans to release a 12nm Vega and Ryzen lineup later this year, so 7nm chips in 2019 seems about right.

It makes me wonder though where Intel is with their process. I read a while ago that TSMC and GF 7nm is actually closer to Intel's 10nm, when measured by Intel standards. If this is true, them Intel still have a tidy lead.

 

[goldstone77]

My most optimistic feeling when it comes to 7nm was we will see SoC type chip early 2019. Anything bigger probably won't be coming until mid 2019 at the earliest. Meaning 7nm Vega (meant for their machine learning stuff) won't be coming out until H2 2019. Also it depends how good TSMC and GF refining their 7nm process for bigger chip. Worse we could end up seeing 20nm repeat.

Well. that's not exactly a problem. We know AMD plans to release a 12nm Vega and Ryzen lineup later this year, so 7nm chips in 2019 seems about right.

It makes me wonder though where Intel is with their process. I read a while ago that TSMC and GF 7nm is actually closer to Intel's 10nm, when measured by Intel standards. If this is true, them Intel still have a tidy lead.

When looking at the different designs of each foundry they all have everything made up into units called "IP blocks" GF actually has the most dense cells for 7nm SOC. Intel uses single dummy gate, and contact-over-active-gate, which increases their logic density by being able to cram in more IP blocks than GF. I've gone through the estimates and they range from 11%-19.6%. Intel reports 100.8 MTr/mm² for their own process. Estimates from two others show GoFlo's process to be ~86-90.5 MTr/mm². While this shows Intel will have a logic density advantage they have yet to release anything other than marketing material on 10nm, Regardless this will be the closest anyone has come to Intel's density for a long time, and even with the logic advantage this should be the closest comparison of uArch between AMD and Intel that would could have ever hoped to see. Right now Ryzen's transistors built on Samsung's 14nm process more closely resemble 20nm transistors. This will change at 7nm, and I will provide a links where I discuss these topics in more detail so anyone can view them if they choose. Ryan Shrout recently visited GloFlo Malta 8 Fab, and his article suggest that their maybe battle going on between TSMC and GloFlo for AMD's business for CPUs and GPUs. It also suggests GloFlo 7nm SOC will be used for AMDs mobile devices. With AMD waiting till the last minute to make the decision to go with TSMC or GloFlo as per the WSA.

The last point:

Provides sourcing flexibility for certain products to support product roadmap and business plans through 2020.

In electronic design a semiconductor intellectual property core, IP core, or IP block is a reusable unit of logic,cell, or integrated circuit (commonly called a "chip") layout design that is the intellectual property of one party.

Process node discussion
http://www.tomshardware.com/forum/id-3341285/amd-naples-server-cpu-info-rumours/page-22.html#20517313
Ryan Shrout
http://www.tomshardware.com/forum/id-3341285/amd-naples-server-cpu-info-rumours/page-28.html#20687647

Getting Radeon Vega Everywhere: An Exclusive Media Interview at AMD Tech Day, with CEO Dr. Lisa Su
by Ian Cutress on January 24, 2018 8:00 AM EST
https://www.anandtech.com/show/12312/getting-radeon-vega-everywhere-an-exclusive-interview-with-dr-lisa-su-amd-ceo?_ga=2.128746956.1302477664.1518393516-1732219874.1517885173

Q18: With GlobalFoundries 14nm, it was a licensed Samsung process, and 12nm is an advancement of that. 7nm is more of a pure GF design. Is there any change in the relationship as a result?

LS: So in 7nm, we will use both TSMC and GlobalFoundries. We are working closely with both foundry partners, and will have different product lines for each. I am very confident that the process technology will be stable and capable for what we’re trying to do.

Edit: Part of Lisa Su "will have different product lines for each. " does not exclude desktop CPUs as Ryan Shrout's comments lead on to suggest.

 

[renz496]

Hard to say. GF track record is not as stellar as TSMC and Intel when it comes to new process node. There are times we heard that GF will finally catch up to TSMC or Intel on process node but in the end they still end up fall short. Instead of GF and TSMC battling to get AMD business it's more like GF is trying their hardest not to let AMD slip from their grip. TSMC can get away without AMD. i don't know how it is right now but in the past over 50% revenue in fab industry was owned by TSMC.

 

[goldstone77]

Hard to say. GF track record is not as stellar as TSMC and Intel when it comes to new process node. There are times we heard that GF will finally catch up to TSMC or Intel on process node but in the end they still end up fall short. Instead of GF and TSMC battling to get AMD business it's more like GF is trying their hardest not to let AMD slip from their grip. TSMC can get away without AMD. i don't know how it is right now but in the past over 50% revenue in fab industry was owned by TSMC.

TrendForce Reports Top 10 Ranking of Global Semiconductor Foundries of 2017, TSMC Ranks First with Market Share of 55.9%
Wednesday , 11 / 29 / 2017
https://press.trendforce.com/press/20171129-3029.html

 

[-Fran-]

https://www.anandtech.com/show/12552/amd-announces-real-time-ray-tracing-for-prorender-and-radeon-gpu-profiler-12

AMD adding some Ray-tracing stuff into their Pro lines (I think?).

I wonder if Navi will come with "full" support for this new RT stuff?

Cheers!

 

[Martell1977]

https://www.anandtech.com/show/12552/amd-announces-real-time-ray-tracing-for-prorender-and-radeon-gpu-profiler-12

AMD adding some Ray-tracing stuff into their Pro lines (I think?).

I wonder if Navi will come with "full" support for this new RT stuff?

Cheers!

I'll likely be doing a full build about the time Navi drops. Really looking forward to see what AMD has to offer. Might end up being my first AMD CPU since my K6-2 300 with a Navi GPU.

Lets hope the GPU prices come back to sanity by then.

 

[renz496]

https://www.anandtech.com/show/12552/amd-announces-real-time-ray-tracing-for-prorender-and-radeon-gpu-profiler-12

AMD adding some Ray-tracing stuff into their Pro lines (I think?).

I wonder if Navi will come with "full" support for this new RT stuff?

Cheers!

hard to say. sometimes AMD also slow to support DX feature. before we thought that polaris will definitely going to support FL12_1 but in the end AMD only add the hardware on Vega generation. but before supporting more feature AMD might need to do something about GCN first.

 

[Martell1977]

AMD might need to do something about GCN first.

It will be interesting to see if Navi is something new or another GCN tweak.

 

[-Fran-]

AMD might need to do something about GCN first.

It will be interesting to see if Navi is something new or another GCN tweak.

Uhm... Chances are low for them to re-build something similar to GCN from scratch given how deep their feet are already in the mud.

I hope for, let's call them, "big tweaks" to GCN with Navi. Raja's influence might be lingering for a long time, or at least after Navi's generation. This is a very convoluted way to say that I would expect a harder cut of GCN after Navi.

Cheers!

 

[Kulasko]

It would have been very interesting to see Raja roaming as free as Keller. GCN still really suffers from insufficient development of the whole architecture and too many patch-up fixes.

 

[cdrkf]

https://www.anandtech.com/show/12552/amd-announces-real-time-ray-tracing-for-prorender-and-radeon-gpu-profiler-12

AMD adding some Ray-tracing stuff into their Pro lines (I think?).

I wonder if Navi will come with "full" support for this new RT stuff?

Cheers!

I'll likely be doing a full build about the time Navi drops. Really looking forward to see what AMD has to offer. Might end up being my first AMD CPU since my K6-2 300 with a Navi GPU.

Lets hope the GPU prices come back to sanity by then.

K6 - 2.... I remember those, the Super Socket 7 Platform lasted for a really long time thanks to AMD!

 

[goldstone77]

Rumor, 9Gbps RAM multi GPU video cards based on Vega coming later this year.
[video="https://www.youtube.com/watch?v=ePsBmCU8sww"]https://www.youtube.com/watch?v=ePsBmCU8sww[/video]
HW News - AMD 9Gbps GDDR5 GPUs, Memory 'Gouging'
Gamers Nexus
Published on Apr 23, 2018

 

[goldstone77]

by Brandon Hill — Thursday, April 26, 2018
AMD 7nm Vega Radeon Instinct GPU AI Accelerators Enter Lab Testing

"I'm also happy to report that our next-generation 7-nanometer Radeon Instinct product, optimized for machine learning workloads, is running in our labs," said Dr. Su. "We remain on track to provide samples to customers later this year."

If you recall, Radeon Instinct is AMD's product line for machine intelligences and deep learning accelerators. The current lineup features a mixture of Polaris- and Vega-based GPUs and could be considered competitors for NVIDIA's Tesla family of products.

Dr. Su went on to confirm that it has two partners with regards to 7nm production for its next-generation GPUs. "So, our foundry strategy is to use both TSMC and GlobalFoundries on the first 7-nanometer product. We are using TSMC for that product and we have a very strong relationship with them. And so, we do see a good momentum on it from what we see, and I'm not concerned about capacity."

The comments by AMD's CEO come weeks after we came across references to Vega 20 in a Linux driver package. Half a dozen Vega 20 SKUs were identified, although no other details were revealed. Vega 20 is rumored to have 64 Compute Units and 32GB of HBM2 with 1TB/s of memory bandwidth. It will also boast PCI-Express 4.0 support and have a TDP of 150W.

According to commentary from AMD at this year's CES, 7nm Vega products for mobile along with the 7nm Radeon Instinct accelerators will ship during the latter half of 2018.

https://hothardware.com/news/amd-7nm-finfet-vega-20-testing

 

[goldstone77]

Radeon Instinct
Super early benchmark result of Vega 20 demonstrates questionable improvement over Vega 10.

AMD Vega 20
A die-shrinked Vega 10 has appeared in AMD labs last week, as revealed by the company on social media. Luckily for us, one of the employees was not careful enough and forgot to disable online result validation. This gave us the first result of yet unreleased Vega 20 graphics card.

The card used in this test is clearly an engineering sample running on early drivers. Officially AMD Vega 20 only goes into 7nm Radeon Instinct, which is a graphics card for machine learning, a card which even lacks display connectors. It shouldn’t really matter for us since Frontier, RX 64 and FirePro all share the same board design. The Vega 20 card that was tested by AMD could be a successor to any of those cards.

The Vega 20 graphics card will feature 32 GB of HBM2 memory, as confirmed by 3DMark listing and by our slides from last year. The device ID of 66A0 belongs to Vega 20, as confirmed by Linux display driver patches.

The clock speeds were obviously read incorrectly, but the same issue could be observed for Vega 10 before it was released (so yes, it does not mean it was running at 1000 MHz). The Vega 20 testing platform features Ryzen 7 1700.

Remember those are early results. We could be seeing much better scores in the coming days, or these could be the last scores we will see in the next few months. Impossible to say.

Vega Frontier vs Vega 20 3DMark11 score:

https://videocardz.com/76076/early-amd-vega-20-3dmark11-benchmark-result-emerges

 

[-Fran-]

Uhm... If Vega 20 will have around 150W TDP, then that might mean they'll push it as a Polaris replacement? Maybe they moved forward Navi's release? (maybe half a year?)

If they put current Vega performance at Polaris level, that should be good enough for now, but that leaves me wondering about the "high end". That would leave Navi or Vega 20+ or something.

Thoughts?

Cheers!

 

[renz496]

still depending on how viable the yield is? but even this Vega 20 won't be available on the market until 2019 time frame.

 

[jaymc]

Uhm... If Vega 20 will have around 150W TDP, then that might mean they'll push it as a Polaris replacement? Maybe they moved forward Navi's release? (maybe half a year?)

If they put current Vega performance at Polaris level, that should be good enough for now, but that leaves me wondering about the "high end". That would leave Navi or Vega 20+ or something.

Thoughts?

Cheers!

Well there's Rumours of Multi GPU Vega around a long time and again now recently (2 and 3 GPU initially for server market but maybe consumer versions), this could solve the high end problem if they release a Consumer mGPU Vega on 7nm.... lol sounds pretty beastly actually

Tech Rumour Mill - Vega, Epyc, DDR5 and Nvidia 1180:
https://youtu.be/79zASO7CTWU

Edit: Btw Dr Lisa Su just moved 3 of the Zen design engineers over to Radeon (officially there meant to be working on power efficiency, but Multi GPU an Infinity Fabric could be anothey reason) so yeah look's like something is definitely going on over there alright..

 

[goldstone77]

12nm Zen is not a die shrink. She could have Zen resources working on power/performance improvements like with Zen+/12nm. But, yes they could be working on infinity fabric to tie in multiple processors as well.

Auguste Richard - Northland Securities

Yes. Thanks for taking the question. Real, real quickly, can you just tell me if the 12-nanometer is a dye shrink or just performance improvement?

Lisa T. Su - Advanced Micro Devices, Inc.

Yes, Gus, it is primarily performance improvement, and it is some design improvements. So you see some of the latency improvements and so on. It is not a dye shrink.

https://seekingalpha.com/article/4165998-advanced-micro-devices-amd-q1-2018-results-earnings-call-transcript?part=single

 

[Sakkura]

Edit: Btw Dr Lisa Su just moved 3 of the Zen design engineers over to Radeon (officially there meant to be working on power efficiency, but Multi GPU an Infinity Fabric could be anothey reason) so yeah look's like something is definitely going on over there alright..

Anything Vega-based would already be done with the design process by now. They may be put to work on Navi though, and the roadmaps have always indicated Navi would focus on "scalability" - possibly with multiple dies.

 

[goldstone77]

I found the original article, and it's about Navi.
AMD RX 600 Series GPU Project “Zen” Detailed – Radeon on Steroids to Amp Clock Speeds & Efficiency
By Khalid Moammer
Apr 15

The two key areas that the team is focused on is significantly improving the clock speeds of AMD’s GPU designs and pushing them to be more power efficient. We’re told that this project was put in motion under the direct orders of CEO Su, who expects results in as early as this year. Although the team is also involved with engineering efforts around the Navi architecture which is expected in 2019 as well as the company’s brand new post-GCN architecture that’s expected in 2020.

We’ve been hearing a lot of chatter in private channels about how the company’s post-GCN, all new architecture is going to be Radeon’s Zen so to speak. But if what we’ve been hearing about Radeon Project Zen is also accurate, we might see some of the new team’s results with the Radeon RX 600 series, much earlier than 2020.

https://wccftech.com/amds-secret-radeon-project-zen-to-boost-gpu-clocks-efficiency-in-2018-navi-in-2019-beyond/

 

[Martell1977]

Edit: Btw Dr Lisa Su just moved 3 of the Zen design engineers over to Radeon (officially there meant to be working on power efficiency, but Multi GPU an Infinity Fabric could be anothey reason) so yeah look's like something is definitely going on over there alright..

Anything Vega-based would already be done with the design process by now. They may be put to work on Navi though, and the roadmaps have always indicated Navi would focus on "scalability" - possibly with multiple dies.

From past articles about what they mean by "scalability", they were referring to memory, talking about adding high speed cache to GPU's. But IIRC is was mostly just for the machine learning and other productivity things where memory can be a limitation. They even spoke of connecting SSD's on to the board to add a high speed cache.

 

[Sakkura]

From past articles about what they mean by "scalability", they were referring to memory, talking about adding high speed cache to GPU's. But IIRC is was mostly just for the machine learning and other productivity things where memory can be a limitation. They even spoke of connecting SSD's on to the board to add a high speed cache.

They've already had cards with SSDs on board, and they specifically touted the high bandwidth cache on Vega.

 

[goldstone77]

FTC makes clear 'warranty void if removed' stickers are illegal. ... Under the 1975 Magnuson-Moss Warranty Act
https://www.engadget.com/2018/04/11/ftc-warranty-warning/

Just in case anyone did not know a lot of the times these types of warranties are put on video card screws to void warranty. FTC is making sure everyone knows it's illegal to use warranty voiding stickers!

 

[goldstone77]

AMD Likely To Announce Vega 20 with 32 GB HBM2 Radeon Instinct at Computex
by Hilbert Hagedoorn on: 06/04/2018 08:41 AM

From the looks of it, AMD will be announcing Vega 20 on Computex. Vega20 is a die shrink of the current model, and should be a 7nm part. It's mentioned that the new Radeon Instinct SKU (machine- deep-learning / enterprise GPU) would get a whopping 32 GB of HBM2 memory.

AMD itself has mentioned the existence of VEGA0 multiple ties away and earlier on mentioning that it would be launching in 218, according to their plans. According to a source of tweak town who spread this information, Vega 20 initially would just be a refresh of Vega 10 for the professional market. It is expected to get the same 64CUs as in Vega 10 and is fabbed at 7nm. Early this year it was already spotted in driver entries as well as a benchmark entry back in April, which also shows 32GB of graphics memory.

There's no word on anything for the consumer (gaming) side. For consumers, 7nm video cards base on Navi seem more viable. AMD has a press-conference this Wednesday.

http://www.guru3d.com/news-story/amd-looks-announce-vega-20-with-32-gb-hbm2-radeon-instinct-at-computex.html

 

[-Fran-]

Just a die shrink of Vega sounds boring as hell. I really wish they give us some information about potential improvements to the GCN uArch, or at least a couple of additional tidbits.

I guess Navi is going to be the GPU to look forward to from AMD?

Cheers!