AMD AMA - Starts at 12PM EST on Wednesday May 9

[jsimenhoff]

Nope, it's open for 24 hours. Its been 20.

Correct, the AMA is open until 12pm EDT

 

[ksec]

DDR5, PCI-E 5.0, 5Gbps Ethernet, along with Mobile 5G, I have always thought 2020 is going to be there era of 5G, would we see these 3 things aligned? Given PCI-E is pretty much short lived.

Cant wait until Zen 2 release.

 

[caveman-jim]

Amazing AMA overall, James. One final question from me, please!

I have always craved to learn a basically overview of all the steps involved in getting a CPU released to market. For instance, around 2 years of chip design, around 1 year of validation, and so on. I would love to have the name, basic description, and average duration of all the major steps. It’s truly a pity I’ve never seen a full article about this — just vague comments about “tape in”, “tape out”, etc.

Broad strokes, the phases look something like this:

• Architecture development
• Product outline definition
• Product modeling and synthesis
• First tape out
• Bring up validation
• Refinement and improvement for performance, yield, cost, features
• Final Tape out
• Platform Validation
• Mass Production

The timing varies depending on a lot of factors, but overall that's a 3-5 year process, depending on a number of factors.

 

[caveman-jim]

Thanks James for the AMA

Is it possible to get a short timeline of the Zen from inception to release just so we can get a general idea on CPU development works (at least from AMD side).

When I joined in mid-2013, Zen had been under development for a year or a little more.

 

[caveman-jim]

Dear James san

Thank you for your prompt reply.
We are looking forward to seeing your new informations.

I'm sorry for that I would like to add a question.
I have a Photon 2 now, so I want a next tablet powered by Ryzen Mobile.
Do you know any successor of something?

Best Regards,

Hi A-Key-san, unfortunately I have no knowledge in that area.

 

[bearmoo]

Broad strokes, the phases look something like this:

• Architecture development
• Product outline definition
• Product modeling and synthesis
• First tape out
• Bring up validation
• Refinement and improvement for performance, yield, cost, features
• Final Tape out
• Platform Validation
• Mass Production

The timing varies depending on a lot of factors, but overall that's a 3-5 year process, depending on a number of factors.

So, am I right to assume that the Vega 7nm part running in the labs has come out of final tape out and is in the Platform Validation stage?

 

[caveman-jim]

Thank you James for answering our questions.

Is Raven Ridge still on GloFo's original 14nm process that Summit Ridge was produced with, or run with some / all of the "12nm" optimizations, or totally different optimizations than the "12nm" process has? I have seen Raven Ridge being referred to as "14nm+", and Raven Ridge A0 (CPUID 00810F00h) ES parts seemed to be in testing phase around the same time as Summit Ridge A0 (00800F00h) ES parts (late 2016), based on CPU support information contained in AMD platform BIOSes, though the final stepping and retail product for Raven Ridge (B0, 00810F10h) were released much later.

It seems that Raven Ridge, even including the desktop lineup, consists of very low leakage parts, and hit a voltage-frequency wall earlier, whereas Pinnacle Ridge, at least in the case of the 2700X, have higher leakage current than Summit Ridge.

It also seems that while Pinnacle Ridge CPUs exhibit higher temperature / power consumption at a given voltage, it is completely offset (and then some) by a shift in voltage-frequency curve with lower voltage required for given frequency, and higher frequencies achievable.

(1)... depending on what the case really is, could you clarify the differences between this "14nm+" and "12nm"? Are the observations above accurate abstractions?

(2) A lot of enthusiasts are overclocking their Ryzen CPUs, with many users hitting 4 GHz and above for all-core overclocks. For Summit Ridge, Robert Hallock mentioned that for long-term use, it would be advisable to stay under a maximum of 1.425v, and mentioned that 1.1v SoC / uncore voltage would be helpful for memory overclocks. Are you able to give guidance on the voltage tolerance of Pinnacle Ridge, considering its different leakage characteristics?

1 - 14nm+ is an improvement over the original 14nm used for 1st Gen Ryzen. Ryzen Desktop with Radeon Vega graphics ("Raven Ridge") went to final production with 14nm+; the benefit was a lower voltage for same frequency curve, vs 1st Gen Ryzen. This allowed us to put the first implementation of Precision Boost 2 into the processor. Precision Boost 2, and the improved L1/L2/L3 and memory latencies are the common factors of Ryzen 2000 series desktop processors. For 12nm, we took advantage of the faster transistors but didn't use the new libraries to execute a die shrink of the design vs 1st gen Ryzen.

2 - That's still a good rough ballpark, although you might consider getting a program like HWInfo and watching the natural voltage maximum applied (make sure you have all BIOS performance enhancers and load line calibration disabled). For myself, I've seen a lot of memory overclocking only need around 1.05 volts for vSOC, although you might need to push to 1.15v - 1.25v for 3600+, especially with tight timings from Ryzen DRAM calculator. On the core voltage side, you're battling the vDroop with your settings a little bit so you might need 1.475 for all core stability at 4.2-4.3Ghz; that'll need some robust cooling and low ambients.

 

[dlerious]

What would you consider the max for temps and voltage for 2nd gen Ryzen (non sub-ambient)?

 

[caveman-jim]

Hey James,

I would like to ask couple of questions about the design of the Zeppelin die.

First of all architecturally is it possible to add 1 more CCX to the CPU die or increase the CCX Core count to lets say 6 without breaking something on the architecture?

Second Question: Are you considering to move into a single CCX design for Desktop products on the future iterations of Zen architecture?

Third: Global Foundries has 2 nodes for 7nm. One is ultra high performance with more power consumption(Namely IBM node). It can work on 5+ Ghz easily, but it consumes a lot of power. The other is extremely power efficient which lets a processor like 2700U to run around 5W TDP. Are you considering to use different nodes for different product groups in the future? Or will the "one size fits all" strategy continue?

Fourth and last: Zen core does not have an AVX512 unit, for that reason, your IPC is vastly inferior to the Intel's architecture on this area. I know it uses a lot of power. Was lowering the power consumption of the cores your main goal for excluding AVX512 unit from the core itself(since GLOFO's 14nm is not able to match Intel's 14nm in terms of perf/w) and is there a possibility for you to include an AVX512 unit on the next iterations of Zen core?

1 - The "Zen" modularity is at the CCX level, which is designed as a four core with l3 cache unit (read more in the ISSCC papers).

2 - Ryzen Desktop with Radeon Graphics is a single CCX design.

3 - For foundry process we'll select the node that offers the best mix of performance, efficiency, manufacturability, and price that supports our business model.

4 - We're always evaluating new instruction extensions. Area efficiency, cost, power, all play parts in architecture design choices, as well as projected adoption rates and target use cases.

 

[raghu78]

Hi Jim
Great work from AMD on Ryzen 2000 series. I am a huge AMD supporter back from the Athlon K7 days. It was really heartening to see AMD deliver on Zen after so many dark years. I hope the team at AMD does not repeat the mistakes of the past like in 2005-2006 by becoming complacent and arrogant. I hope there is a maniacal focus on execution and delivering to the customers, both consumer and enterprise. Here are my questions even though they are not directly related to Ryzen

1. I foresee AMD running into fab capacity issues in the 2020 timeframe since I think Zen 2 will allow AMD to make huge share gains in enterprise and consumer. Is AMD planning with its foundry partners ,especially GF, for much higher wafer allocation and (if need be capacity addition by the foundry partner) to be able to meet very high future demand in the 2020 timeframe ?
2. When can we expect the next socket from AMD incorporating DDR5 and PCIE 4.0 ? Is it 2020 or 2021 .

 

[caveman-jim]

Thanks for all the answers, James.

It's great that AMD promised AM4 socket compatibility until at least 2020, but the BIOS situation is problematic, and even though the boot kit solution did work for me, it's still somewhat of a hassle on both the user and AMD's side.

Some motherboards offer the option to flash BIOS without a CPU, but that's a rare feature available only at the high end.

Has AMD given any thought to having some base functionality which will enable future processors to at least be able to flash a BIOS via USB, even if the rest of the chip and interfaces require an update to work? Using the ARM core, having a standard USB controller, ... I don't know, I'm not familiar with what's needed to achieve this, or what prevents new chips from working with an older BIOS, but surely it's not impossible.

We're looking at ways to make it simpler, for sure. It's mostly a cost and alignment problem, which means it's solvable with time.

Despite the hassles, I think end users prefer to have some kind of cross-compatibility than a forced purchase of new platforms. How we set the expectations was what tripped us up.

 

[zir_blazer]

1 - No new news, sorry.

2 - Can't speculate on future products, sorry.

3 - The interest levels on AMD reference platforms being available in the marketplace has been very low, plus our motherboard partners are not sure they want us to compete - and we currently have not only very good relationships with them, but also a pretty good set of products in the market today. There's been a few references to few SFF boards being available, but that will change as Ryzen becomes more popular and the market drives the need for the less popular (but still very necessary) form factors. There's also the resouce angle to consider - do you want AMD engineers building products that are already in market, or do you want them focused on the next generation of products? Or, if you imagined hiring new staff, do you want those new staff members reinventing products already on sale, instead of being people that deliver the new IP the drives the computing market forward? I think most people pick innovation and new products. But there's certainly still requests AMD can make to the existing partners to see if there's something else they can do.

Thanks for replying.

Sad that there aren't news regarding open source Firmware efforts.


I understand that from AMD point of view, spending resources on doing what only AMD can do makes more sense that spending in things that the Motherboard manufacturers can design by themselves. Where I see the problem, is that while AMD DOES provides the core product that I want, I can't get it in the platform that I want. I don't know if a Motherboard manufacturer will be interesed in catering to my audience, but since currently no one does, I would prefer to get AMD itself to see if it worth pushing.

For reference, my user profile is that of a sort of old school, grumpy power user that got tired of seeing "gaming" motherboards, and would prefer a plain and dull but high quality and solid Motherboard as a go-to product instead of having to sort among a ton of variety just to see the breakpoint where you can get a good Motherboard without spending on gimmicks that I'm not going to use, nor that skimped in quality for cost corner-cutting.
What I'm looking for would be similar in nature to the old Intel branded Motherboards that were discontinued. Meanwhile, I have found my niche in Workstation and Server level Motherboards, but AMD doesn't offer currently anything that isn't the full EPYC, and I have to sacrifice any tinkering with the Firmware options, so no overclocking/undervolting or anything out-of-spec. The closest thing that comes to what I would like is the announced Embedded Ryzen products, of which there is a Raven Ridge version (Some guy already mentioned it, the V1000) that looks spectacular in the hands of a consumer, and may be potentially affordable even for entry level computers.

The comment about mATX isn't because I favour mATX, is because after some theorycrafting sessions, that is the "sweet spot" for Zen IO capabilities as a standalone SoC with no Chipset and avoiding to use onboard third party controllers whenever possible. An embedded version of Raven Ridge with its 24 PCIe Lanes can provide 8x PCIe 8x slot (In 16x), 4 SATAs, 2 10 GBit NICs, one M.2/U.2, two PCIe 1x slots, and two free lanes for an integrated WiFi, and/or maybe an USB Controller to augment its built-in 4 USB 3 Ports. Heck, it can even support ECC RAM. You can take it more as a product petition than an actual question
At least I'm happy that I'm not the only one interesed in such product.

 

[.vodka]

Despite the hassles, I think end users prefer to have some kind of cross-compatibility than a forced purchase of new platforms. How we set the expectations was what tripped us up.

This is the correct approach, thank you for taking it and sticking to it.

There is no justification for forced motherboard/platform "upgrades".

 

[caveman-jim]

Hi James

Heterogeneous computing was considered an important element in the days of Bulldozer especially around the KaverI launch. With the advent of Zen is heterogeneous computing still considered important and if so then what kind of future are we looking at?

We support this through our efforts on OpenCL and other open source software.

 

[eva02langley]

Any Freesync technology on TVs in the future? Is AMD promoting the feature to display manufacturer?

 

[caveman-jim]

Maybe a quick follow up question: can you divulge what silivon process Vega Mobile is made in? I believe eatlier roadmap presented by Mark Papermaster put Vega in the 2018 12nm refresh category so I am wondering if Vega Mobile is the one?

You know, I'm not sure but I'm pretty sure it's 14nm or maybe 14nm+.

 

[caveman-jim]

I too am very curious about ROCm for Windows ...

I believe ROCm support for Windows was supposed to be delivered during last year at the end of the last quarter according to an AMD developer which leaked that detail in a comment in their developer blog ... (we could see Windows support coincide with the release of ROCm 2.0 either by the end of this quarter or conservatively by the end of the next quarter)

That being said, I think it's hinted that James specializes in technical marketing for their CPU team so I am extremely interested in knowing if their engineers are considering hardware extensions such as AVX-512 and transnational memory support such as Intel TSX or if there are patent roadblocks due to the end of their cross-licensing agreement with Intel that could potentially prevent their implementation ...

I'm not in tech marketing, but we do evaluate different instruction extensions to help improve our products, based on area, efficiency, adoption rate, implementation cost, etc. I have no info on RoCM, sorry.

 

[caveman-jim]

Anandtech just did a deliding article on the 2400G. What is the cost of using solder over paste?

It seems to be a negligible price for the benefit.

We don't break out cost at that level, sorry. The decision to use a non-soldered TIM was based on market segmentation, time to market, as well as overall cost profile. It's a decision we can re-examine for future products based on market feedback and success.

 

[caveman-jim]

So, am I right to assume that the Vega 7nm part running in the labs has come out of final tape out and is in the Platform Validation stage?

I can't comment on your speculation. Don't use my simplistic breakdown to inform you for large scale decisions like investing, or market outlook for AMD.

 

[caveman-jim]

What would you consider the max for temps and voltage for 2nd gen Ryzen (non sub-ambient)?

95C is your shutdown temp. So you can drive the frequency/voltage to keep inside of that. Personally I'm comfortable with max temps in the upper 80s and voltages around 1.45v max.

 

[caveman-jim]

Hi Jim
Great work from AMD on Ryzen 2000 series. I am a huge AMD supporter back from the Athlon K7 days. It was really heartening to see AMD deliver on Zen after so many dark years. I hope the team at AMD does not repeat the mistakes of the past like in 2005-2006 by becoming complacent and arrogant. I hope there is a maniacal focus on execution and delivering to the customers, both consumer and enterprise. Here are my questions even though they are not directly related to Ryzen

1. I foresee AMD running into fab capacity issues in the 2020 timeframe since I think Zen 2 will allow AMD to make huge share gains in enterprise and consumer. Is AMD planning with its foundry partners ,especially GF, for much higher wafer allocation and (if need be capacity addition by the foundry partner) to be able to meet very high future demand in the 2020 timeframe ?
2. When can we expect the next socket from AMD incorporating DDR5 and PCIE 4.0 ? Is it 2020 or 2021 .

Thanks for the kind words, and you're right - we are keeping focused on execution and customers. It's a point our leadership make in every all-hands.

Running out of capacity would be a nice problem to have. We have great relationships with our foundry partners so we'll manage as necessary to support our success.

For next socket, we're not setting timelines or specifications yet.

 

[caveman-jim]

Any Freesync technology on TVs in the future? Is AMD promoting the feature to display manufacturer?

Keep an eye on the public news from our social channels - AMD can't announce anything for another company without their permission, so if there is a FreeSync capable TV, I can't tell you any details.

 

[neblogai]

Hello again,

I'm interested- why Raven Ridge got such staggered launch schedule: first laptops with 2500U were released back in October 2017, then desktop G-chips and more mobile SKUs were released 3 months later, 2300U- half a year later, and now there are leaks (and expectations) of more RR SKUs with higher-TDP mobile market. It seems- launch window for different SKUs of the same chip takes a very long time- could be 9 months or so. Were there problems with ramping up Raven Ridge production, or other reasons?

 

[caveman-jim]

Hello again,

I'm interested- why Raven Ridge got such staggered launch schedule: first laptops with 2500U were released back in October 2017, then desktop G-chips and more mobile SKUs were released 3 months later, 2300U- half a year later, and now there are leaks (and expectations) of more RR SKUs with higher-TDP mobile market. It seems- launch window for different SKUs of the same chip takes a very long time- could be 9 months or so. Were there problems with ramping up Raven Ridge production, or other reasons?

The launches are aligned to the global OEM schedule for availability.

 

[wahdangun]

I hope not. Mining is a stupid idea so hopefully AMD won't be encouraging it.

Why, it's not all about mining, but block chain tech is quite revolutionary.

So if AMD have lead in this sector maybe in 5 years, you will see AMD cpu/gpu in all data center.