Speculation: Ryzen 4000 series/Zen 3

[DrMrLordX]

With Zen 3 first launching October 2020, a launch in Q1 2022 is a standard product cycle of 15-18 months.

Sorry, but I'm just sick of people thinking that 15-18 months is an acceptable standard product cycle. The cadence is supposed to be "12-18" months which really means 18 months, or maybe a little earlier if you're lucky, kthx have a nice day! If that's okay for you great, but I'm frankly getting tired of it.

What makes you think waiting until Q1 2022 is AMD milking their dominant position over Intel? I'm sure they are much smarter than that.

If it gets AMD more money in the short term and pleases shareholders, what makes you think it's stupid? It may not be great for people that have been supporting them for years, but it's definitely good for their short-term bottom line.

They know they need to maintain their cadence to continue tightening the noose.

They could have broken Intel's back by next year, and now it's basically impossible, so oh well. Now things are kind of muddy.

There is no fixing anymore

I want to see more data on the MT clocks for the 5900X in particular. Those CPU-z MT numbers are fishy. We need more data, though. In any case, Matisse gained some considerable MT clocks on some systems with 1.0.0.3ABBA. I would not be surprised if Vermeer gets similar AGESA updates.

From the few articles that I have read on the subject, it sounds like the cycle time for processing a 7 nm wafer is on the order of 80 days. So AMD had to be manufacturing Zen 3 die more than 3 months ago to launch in October. It was probably more than 3 months ago since the packaging and testing process takes time and they need to stockpile supply.

I'm sure AMD started stockpiling as far back as May and probably had significant supplies available as far back as August. It's hard to tell since it doesn't seem like they've repeated the mistake they made last year of allowing mobo OEMs to offer file archives of months-old UEFI revisions.

They can’t try to fill current demand by increasing wafer starts now

They won't. Remember that the 3900X in particular sold out in July 2019 rapidly, and significant supplies took months to reach market. I recall it being about three months before AMD managed to alleviate the problem.

They can’t launch with low supply to gauge demand and then increase production due to long lead times.

They've done it before. And I think you're misinterpreting some of what I suggested.

Let's say AMD is getting x wafers per month, and they have to bin EPYC and Ryzen dice from those wafers. Automatically a certain n% goes to EPYC, since it's a higher margin product. AMD will gladly cannibalize any amount of desktop sales for more EPYC sales. But 100-n% dice don't bin properly for EPYC, so desktop Ryzen gets those.

Scenario one: I bin Y number of Ryzen dice from the first x wafers I get in one month of binning/packaging. I release CPU now and start selling them in limited numbers, watching sales and knowing that I already am getting x wafers/month anyway, so over the next 3-4 months, I'm going to get a steady supply of Ryzen dice no matter what future wafer orders I try to make in the interim. Initial sales can inform whether or not I need to accelerate my wafer orders. I can gather and utilize this data almost immediately, and even if it takes 3-4 months to get a larger quantity of wafers from TSMC, I'm still getting my x wafers/month anyway.

Scenario two: I bin 4Y number of Ryzen dice over the next four months, launch, and then gauge demand and maybe order wafers at a higher rate (I'm still taking x wafers/month anyway to keep up my future supply of Ryzen and EPYC, and I'll probably continue that for another 4-6 months minimum after the initial stockpile)

I can't see any particular reason why I would go with scenario two.

It will take a lot longer for Intel’s monopoly to dissolve. Although, at this point, I think both companies are going to have trouble against ARM unless Nvidia really screws it up.

Any people wonder why I think AMD should be hitting with Genoa in 2021, ASAP. Sheesh.

 

[Gideon]

Sorry, but I'm just sick of people thinking that 15-18 months is an acceptable standard product cycle. The cadence is supposed to be "12-18" months which really means 18 months, or maybe a little earlier if you're lucky, kthx have a nice day! If that's okay for you great, but I'm frankly getting tired of it.

I understand your frustration and I do hope they find ways to speed it up in the future to something more like a yearly cadence.

But you can't just make the problem go away by yelling at AMD to "be faster!!!". A large number of people are working overtime to deliver these products as fast as they can. Speeding up product design as complex as a CPU isn't easy, even with unlimited resources there are many things that won't really benefit from more people or money thrown at the problem (When making babies you can improve the bandwidth by using more women, but you can't improve the latency, which is always 9 months minimum)

I remember Lisa Su stating in an interview, that they believe they can deliver more meaningful upgrades with the current cadence. While yearly delivery would be nice, it would end up with much more incremental upgrades (with features artificially being postponed an extra product-cycle for not being "on time").

Besides two more points:

1. They are actually speeding up compared where they were: Zen 2 was released 29 months after Zen 1. Zen 3 comes just 15 months after Zen 2. Zen + doesn't count. It was really just some minor fixes that just weren't ready for Zen 1: cache latency fixes already in later Threadripper revisions, fixed Turbo, minimal MC tweaks coupled with a small node bump that required no rework (new masks). Hopefully they manage to shave some extra months off with Zen 4 as well.
2. Do you remember the state of BIOS'es when 2700x released? Or the turbo fiasco when Zen 2 launched? I remember plenty of people on reddit blaming AMD for releasing "an unfinished product". You'd have way more problems like every cycle when speeding up the releases.

IMO you vastly underestimating the increasing complexities companies are facing with each new product generation. It's not like AMD could deliver new generations every 6 months but "are just lazy".

I do understand that they need to find a way to speed things up as ARM competition is looking fierce and with the continiuing of the current state of affair will just steamroll them, but it's often easier said than done.

 

[Richie Rich]

Not sure whether this legit, but CPU Monkey has uploaded some Zen 3 results in CBR20 for the 5950x, 5900x, 5800x and 5600x.

Cinebench R20 (Single-Core) CPU benchmark list (cpu-monkey.com) *edit* fixed link!

Assuming this is accurate, looks like 15% IPC increase for certain after you consider clockspeed enhancements. But doesn't CB use floating point? So I guess the rumors of the 50% IPC increase for floating point must be incorrect.

15% IPC is good, but I was hoping for around 20% or so.

I suppose when you consider that Zen 3 is using an enhanced version of the same process node as Zen 2, then 15% IPC is quite good. But if rumors are true, Intel looks like they are targeting much larger IPC increases with Golden Cove and its successor. Zen 3 is inevitably going to face off against Golden Cove next year, unless AMD can get Zen 4 out by late next year as well.

I agree. Under normal circumstances 12-15% IPC would be decent performance uplift. But from "completely new architecture" I would expect a much more. It looks like a fake results to me because I expect +40% in FPU heavy load from Zen 3. We have indirectly confirmed AVX512 for Zen 3 so there must be some nice FPU uplift.

I hope official Zen 3 benchmarks will show higher performance. If not then Golden Cove and Sapphire Rapids with new AMX SIMD and DDR5 will be pretty tough competition.

And SMT4 of course.

 

[DrMrLordX]

But you can't just make the problem go away by yelling at AMD to "be faster!!!".

That is true. It would require people voting with their wallets. For my part, if Zen4 isn't available in 2021, then I guess I'll just camp out on Matisse for awhile longer.

I remember Lisa Su stating in an interview, that they believe they can deliver more meaningful upgrades with the current cadence.

I'm not sure if I buy it. The Zen2, 3, and 4 teams were assembled at about the same time, with Zen2 bearing the earliest fruit (obviously), so Zen3 and 4 should be ready at least in simulation if not in QS irrespective of when previous products actually reach market. TSMC is holding up their end of the bargain.

Vermeer is going to be a great product. It's just going to be on the market in one form or another for awhile.

 

[exquisitechar]

I agree. Under normal circumstances 12-15% IPC would be decent performance uplift. But from "completely new architecture" I would expect a much more. It looks like a fake results to me because I expect +40% in FPU heavy load from Zen 3. We have indirectly confirmed AVX512 for Zen 3 so there must be some nice FPU uplift.

I hope official Zen 3 benchmarks will show higher performance. If not then Golden Cove and Sapphire Rapids with new AMX SIMD and DDR5 will be pretty tough competition.

And SMT4 of course.

Those benchmarks are fake, but I have to say...Willow Cove just came out. Zen 3 will have to compete with Golden Cove a bit, but Golden Cove is really much closer to Zen 4 (which will break it in half).

 

[Ajay]

I want to see more data on the MT clocks for the 5900X in particular. Those CPU-z MT numbers are fishy. We need more data, though. In any case, Matisse gained some considerable MT clocks on some systems with 1.0.0.3ABBA. I would not be surprised if Vermeer gets similar AGESA updates.

Definitely. The results look borked. Could be anything from an earlier stepping to the BIOS or, AMD has done a better job clock boosting ST workloads somehow. Higher ST almost always leads to higher MT. CPU-Z is such a limited test, I'm sure it fits in L3$, probably even L2$.

If the 5700X has a 25% higher performance than the 3700X, I may sell my 3900X to buy one. I very rarely use all 12 cores - so I rather have the increased performance for my typical loads. Guess my eyes were bigger than my wallet

 

[moinmoin]

Sorry, but I'm just sick of people thinking that 15-18 months is an acceptable standard product cycle.

Just seems to be an odd thing to take such a strong stance about.

Personally I think the quality of a given product is much more important than whether it makes a date a couple months give or take (it's a different story if those months turn into years and the resulting products don't match what was initially promised, hi Intel).

They could have broken Intel's back by next year, and now it's basically impossible, so oh well.

That's entirely impossible in this dimension. The quantity Intel delivers to the markets is not something AMD can break, even if it had all the capacity at TSMC Apple has instead. And we all know AMD comes from near bankruptcy just a couple years ago, so it's more conservative at future allocation of capacity ordered than it should haven been in hindsight. Nobody can expect AMD to go all out full risk financially.

I'm sure AMD started stockpiling as far back as May and probably had significant supplies available as far back as August.

Honestly highly doubt that. Capacity priorities are the semi custom business customers (so Sony and Microsoft), existing datacenter order backlogs (so likely mostly Epyc 2 currently), and only after those all its own consumer products. Zen 3 likely sees sampling of Epyc 3 for several months now, but I don't think to the degree that stockpiling can have been happening since May.

 

[JujuFish]

Sorry, but I'm just sick of people thinking that 15-18 months is an acceptable standard product cycle. The cadence is supposed to be "12-18" months which really means 18 months, or maybe a little earlier if you're lucky, kthx have a nice day! If that's okay for you great, but I'm frankly getting tired of it.

How often do you upgrade your CPU? I don't know if I'm an outlier, but I've done it once in the last decade (2016). Might upgrade with Zen 3, might wait another couple years for other tech advancements to hit (DDR, USB, PCI-e, whatever).

 

[DrMrLordX]

Just seems to be an odd thing to take such a strong stance about.

Timing is everything.

Intel has been on what is essentially the same server CPU (read: not platform) since 2017. 2017. Cascade Lake-SP is nothing more than a bugfixed Skylake-SP with some frequency improvements here and there. Cascade Lake-AP was a failure that barely saw anyone even sell it. Cooper Lake was pulled from most of the market. Ice Lake-SP still hasn't shown up, and it's not clear that it ever will in significant volume. That leaves Intel with Sapphire Rapids as their next server product, one which may not show up until late 2021, and that's assuming they can execute well enough on 10SF/SFE (or whatever process they will actually use for it). It is entirely possible that Sapphire Rapids won't show up until 2022. We'll see.

Regardless, AMD has finally started to fight back. While Skylake-SP was the fastest server CPU in 2017 (Naples challenged it seriously, but arguably could not unseat it completely), Rome changed that in late 2018/early 2019. Against Milan, Cascade Lake-SP has no chance. Against Genoa, Cascade Lake-SP is a complete joke. Pitting Genoa, even for a few months, against the ancient-by-industry-standards Cascade Lake-SP would be a complete bloodbath. AMD can't beat Intel by interfering with their vendor contracts and their deep sales relationships, but they can beat them through sheer technical dominance. Eventually, those with an entrenched interest in continuing to choose Intel products would have to turn away from them just on the basis that they might no longer be competitive in their respective industries while contiuing to use Intel hardware. It is (or rather, was) AMD's one shot to gain a near-permanent foothold in the server space, which was the entire reason for Zen existing in the first place. Now Intel maybe has a chance to fight back with Sapphire Rapids, assuming they execute well enough to bring it to market quickly-enough to challenge Milan. Which, I'll grant, isn't an absolute certainty.

Personally I think the quality of a given product is much more important than whether it makes a date a couple months give or take (it's a different story if those months turn into years and the resulting products don't match what was initially promised, hi Intel).

Months add up. AMD's relaxed cadence puts Genoa maybe after Sapphire Rapids. Sticking to a 12-14 month cadence puts Genoa before Sapphire Rapids. Pretty simple really.

The quantity Intel delivers to the markets is not something AMD can break, even if it had all the capacity at TSMC Apple has instead.

Last I heard, TSMC can deliver nearly 3x the total wafer count that Intel currently can; granted, TSMC has a lot of customers outside of AMD and Apple.

Nobody can expect AMD to go all out full risk financially.

They did in 2017. If Zen went bust, the company probably would have folded. And I don't think I'm being overly-dramatic.

Honestly highly doubt that. Capacity priorities are the semi custom business customers (so Sony and Microsoft), existing datacenter order backlogs (so likely mostly Epyc 2 currently), and only after those all its own consumer products. Zen 3 likely sees sampling of Epyc 3 for several months now, but I don't think to the degree that stockpiling can have been happening since May.

Milan has actually been sampling for "select customers" since last year. Yeah they needed to spend a lot of wafer allocation on other projects - notably Xbox Series X and PS5. But they still have (apparently) been taking N7+ wafers for Milan dice for over ten months, and the ones that don't bin well enough for Milan SKUs can go to desktop Ryzen instead.

How often do you upgrade your CPU? I don't know if I'm an outlier, but I've done it once in the last decade (2016). Might upgrade with Zen 3, might wait another couple years for other tech advancements to hit (DDR, USB, PCI-e, whatever).

Depends. I had a big gap between 2009 and 2014, but I usually buy for platform reasons more than anything else. I went for FM2+ because I was interested in HSA (got an A10-7700k; later updated to a discount 7870k) and then AM4 because, Zen, why not? Skipped Pinnacle Ridge and grabbed a new AM4 rig for Zen2. My next plan is to jump on the DDR5 bandwagon with Zen4. Those of us who are waiting for that leap will be waiting for awhile. Zen3 will be a pretty big increase in performance, but it's not compelling enough for me to want to move on it, even if AMD rehashes it in the form of Warhol (which, by the way, is still just a rumour).

 

[Carfax83]

I agree. Under normal circumstances 12-15% IPC would be decent performance uplift. But from "completely new architecture" I would expect a much more. It looks like a fake results to me because I expect +40% in FPU heavy load from Zen 3. We have indirectly confirmed AVX512 for Zen 3 so there must be some nice FPU uplift.

I'm sure Zen 3 won't support AVX-512, but there was some speculation on these forums that it might have a 3rd AVX2 unit which could increase FPU performance substantially.

But that was just speculation as I recall. I don't think there were any leaks at all that supported it.

 

[inf64]

I'm sure Zen 3 won't support AVX-512, but there was some speculation on these forums that it might have a 3rd AVX2 unit which could increase FPU performance substantially.

But that was just speculation as I recall. I don't think there were any leaks at all that supported it.

There are no leaks nor benchmarks supporting AVX512 or better FP throughput. SMT4 is also a pipe dream, although it may be in store for some other Zen generation (Zen5 as the next tock?).

 

[moinmoin]

Timing is everything.

Standing on its own and offering unique capabilities is more important. And there AMD already does well, with I/O capability, core/thread count, energy efficiency etc. Especially in the datacenter this is not a sprint but a marathon run.

While Skylake-SP was the fastest server CPU in 2017 (Naples challenged it seriously, but arguably could not unseat it completely), Rome changed that in late 2018/early 2019. Against Milan, Cascade Lake-SP has no chance. Against Genoa, Cascade Lake-SP is a complete joke. Pitting Genoa, even for a few months, against the ancient-by-industry-standards Cascade Lake-SP would be a complete bloodbath. AMD can't beat Intel by interfering with their vendor contracts and their deep sales relationships, but they can beat them through sheer technical dominance. Eventually, those with an entrenched interest in continuing to choose Intel products would have to turn away from them just on the basis that they might no longer be competitive in their respective industries while contiuing to use Intel hardware. It is (or rather, was) AMD's one shot to gain a near-permanent foothold in the server space, which was the entire reason for Zen existing in the first place. Now Intel maybe has a chance to fight back with Sapphire Rapids, assuming they execute well enough to bring it to market quickly-enough to challenge Milan.

Let's first see how Intel is doing with Ice Lake-SP, Sapphire Rapids' direct precursor, which should go head to head with Milan.

So far on 10nm Intel has not shown yet being capable of producing chips with more than 4 cores. Efficiency and yield still are questionable on this node. For Sapphire Rapids to take away all of Epyc's (current and future) selling points it has to match both its I/O capability as well as it core/thread count. Consider me skeptical that Intel manages to create such chips anytime soon in significant quantities at the necessary efficiency target, if at all. Nevertheless I fully expect AMD to continue considering best case scenarios for Intel to be the targets to prepare to compete against.

Last I heard, TSMC can deliver nearly 3x the total wafer count that Intel currently can; granted, TSMC has a lot of customers outside of AMD and Apple.

That's across all different process nodes. Intel moves its fabs between different process nodes (leading to the infamous 14nm scarcity). TSMC over years builds up capacity for new nodes in phases (so new nodes don't have the final possible throughput at the start).

They did in 2017. If Zen went bust, the company probably would have folded. And I don't think I'm being overly-dramatic.

Zen certainly was AMD's last straw, but it wasn't an all out financially risky move. In contrary they started with a slow roll out of different products in different market segments, which incidentally they still do today. I'm honestly glad AMD isn't as inpatient as you are.

Milan has actually been sampling for "select customers" since last year. Yeah they needed to spend a lot of wafer allocation on other projects - notably Xbox Series X and PS5. But they still have (apparently) been taking N7+ wafers for Milan dice for over ten months, and the ones that don't bin well enough for Milan SKUs can go to desktop Ryzen instead.

I highly doubt dies from early on of that sampling phase are usable for the final consumer product being launched soon. We will see, the week of manufacture should still be written on the IHS, isn't it?

 

[gruffi]

I agree. Under normal circumstances 12-15% IPC would be decent performance uplift. But from "completely new architecture" I would expect a much more. It looks like a fake results to me because I expect +40% in FPU heavy load from Zen 3. We have indirectly confirmed AVX512 for Zen 3 so there must be some nice FPU uplift.

"New architecture" and architecture from scratch are still two different things. Zen 3 is not the latter. It's still Zen, just fundamentally revamped. I expect ~20% more IPC than Zen 2 which would be a very good uplift. Don't expect something like the original Zen compared to Bulldozer. The rumors about +50% fp performance one year ago or so were likely based on Zen 4, not Zen 3. I think some leakers mixed that up. Zen 4 is expected to support AVX512, not Zen 3. Latest rumors suggest Zen 3 was more optimized for integer performance. Which is good for most workloads.

 

[inf64]

"New architecture" and architecture from scratch are still two different things. Zen 3 is not the latter. It's still Zen, just fundamentally revamped. I expect ~20% more IPC than Zen 2 which would be a very good uplift. Don't expect something like the original Zen compared to Bulldozer. The rumors about +50% fp performance one year ago or so were likely based on Zen 4, not Zen 3. I think some leakers mixed that up. Zen 4 is expected to support AVX512, not Zen 3. Latest rumors suggest Zen 3 was more optimized for integer performance. Which is good for most workloads.

Zen3 will have massive IPC jump when compared to previous new uarchitecture Zen(1).
Zen1 target was 40% versus SR(last gen of Bulldozer). AMD did overshoot the target and achieved ~52%.
Zen3 might have had a similar of 40% target versus Zen1. Zen+ brought 3.5%, Zen2 brought another 15%. In order to be 40% faster than Zen1, Zen3 needs to have 17.5% IPC uplift versus Zen2: 1.035 x 1.15 x 1.175~=1.4 or 40%

 

[Kenmitch]

Not much longer to wait and see how Zen 3 turns out. Maybe it'll be a 2 for 1 announcement with some RDNA2 teasing?

 

[dr1337]

Zen3 might have had a similar of 40% target versus Zen1

Sure, but zen 3 also has a LOT more in common with zen 1 than zen 1 did with even excavator(the actual last gen of f 15h), its an evolution. The language of "new architecture" makes it sound like it could be a brand new design but really I think he meant in the sense of a true generation. Tbh if zen 3 is some crazy all new radical core design, 20% ipc is good, but not really the most impressive improvement.

Im still happy with 15-20% all things concerned. This consistent improvement is a breath of fresh air compared to the last 7 years before zen. I think if they can keep up this pace they will have no problem dealing with intel and ARM.

 

[Saylick]

Hahaha, it wasn't that long ago where we'd only get a 10% IPC increase every 2 years and now we're getting 15% with every major architectural overhaul with Zen. We should be happy if we get only 15% IPC gains and really shouldn't fault AMD for not getting more to be honest. 15% per generation is a SOLID improvement for x86 processors. Anything on top of that is icing on the cake. Hell, we should be glad that we don't see a regression in clocks, because we all know how happy we were with Intel hitting 18% IPC gains but then not gaining ST performance because of the 20% clock reduction...

 

[eek2121]

Those benchmarks are fake, but I have to say...Willow Cove just came out. Zen 3 will have to compete with Golden Cove a bit, but Golden Cove is really much closer to Zen 4 (which will break it in half).

Looks believable to me, especially after that CPU-Z leak. If anything the ST numbers should be a bit higher.

 

[Richie Rich]

I'm sure Zen 3 won't support AVX-512, but there was some speculation on these forums that it might have a 3rd AVX2 unit which could increase FPU performance substantially.

But that was just speculation as I recall. I don't think there were any leaks at all that supported it.

Norrod said that with Zen2 they were not able target all servers (meaning HPC segment). IIRC he said that Zen3 will change that and target 100% of servers. When I read between those lines I understand lack of AVX512 is the biggest problem in HPC. AMD can solve that by supporting AVX512 (maybe not full speed but rather half-speed like Zen1 did handle AVX) or by introducing their own SIMD instruction set (modern version of AVX512 with support of matrix multiplication as a AMX replacement).

The later is super bold move however not completely sci-fi. Remember K12 development was canceled back in 2015 exactly when SVE instruction set was finalized (Fujitsu A64FX development started) so AMD had SVE spec at the begining of Zen3 development. There is a pretty good chance that AMD decided to develop something better than AVX512 inspired by SVE. Maybe better than SVE itself.

"New architecture" and architecture from scratch are still two different things. Zen 3 is not the latter. It's still Zen, just fundamentally revamped. I expect ~20% more IPC than Zen 2 which would be a very good uplift. Don't expect something like the original Zen compared to Bulldozer. The rumors about +50% fp performance one year ago or so were likely based on Zen 4, not Zen 3. I think some leakers mixed that up. Zen 4 is expected to support AVX512, not Zen 3. Latest rumors suggest Zen 3 was more optimized for integer performance. Which is good for most workloads.

+50% FPU performance can come from somebody who saw uarch slides under NDA with 3xFPU configuration insted 2xFPU in Zen1/2. It definitely makes sense to me as more pipes are a great way how to increase IPC without need to increase SIMD width. As good example is new Cortex X1 core doubling FP units to 4x128-bit in compare to sister core Cortex A78 having only 2x128-bit. If ARM can easily double FPU units then I see no problem for Zen3 to have only +50% FPU pipes.

Zen3 will have massive IPC jump when compared to previous new uarchitecture Zen(1).
Zen1 target was 40% versus SR(last gen of Bulldozer). AMD did overshoot the target and achieved ~52%.
Zen3 might have had a similar of 40% target versus Zen1. Zen+ brought 3.5%, Zen2 brought another 15%. In order to be 40% faster than Zen1, Zen3 needs to have 17.5% IPC uplift versus Zen2: 1.035 x 1.15 x 1.175~=1.4 or 40%

Exactly. And if Zen4 will be another +15% IPC uplift then it's 1.4 x 1.15 = 1.61 or 61% over Zen1. This is pretty big IPC uplift which suggest wider uarch. Just for comparison, Intel's extracted from 4xALU uarch (Haswell to Ice Lake) approximately 29% IPC (according to Geekbench5 in my TOP20 IPC table here.). Jim Keller confirmed something significantly wider (probably Golden Cove).

So Zen3/4 uarch with IPC uplift up to 60% really suggest something wider than 4xALU and 4xFPU pipes in current Zen2. K12 was 4xALU+1xBranchUnit and ARMs like Cortex A77/78/X1 are using 4xALU+2xBranch config (unblocking ALU units for useful calculation while two branch pipes handle branch instructions, there is usually 20% branch instructions in avg code). So I see K12's config 4xALU+1xBr as a minimum. Apple cores have 6xALU config since 2017 so even this config wouldn't surprise me much. So Zen3 might be pretty wide beast: 6-wide integer core and 6-wider FPU core.

 

[A///]

Well guys anyone Remember adored?
Let's wait for the reviews or reputable leaks

We do. They were running figures that were "leaked" to them or they estimated. They didn't have anything to back up their claims with.

 

[A///]

Oh. He's back.

I understand your frustration and I do hope they find ways to speed it up in the future to something more like a yearly cadence.

He's repeating himself ad nauseam. He's repeating the same stuff he replied to me a couple weeks back. Never thought I'd see someone advocate AMD to what Intel had done for years, even before their troubles began.

Don't bother debating him. It's a waste of your valuable time. You can lead a horse to water, you can't make them drink it.

Not much longer to wait and see how Zen 3 turns out. Maybe it'll be a 2 for 1 announcement with some RDNA2 teasing?

It was likely from the start to keep them in the news and boost stock futures. NVidia announced they were bumping up the 3070 release to the 29th of October this morning. Either they've fixed the deeply ingrained issues with their hardware or they know the sheep will buy it based on brand name, when the card isn't going to be terribly impressive.

 

[Mopetar]

Not much longer to wait and see how Zen 3 turns out. Maybe it'll be a 2 for 1 announcement with some RDNA2 teasing?

I thought someone posted a tweet from someone at AMD (maybe it was a tweet from Lisa Su) about an RDNA2 tease at the upcoming Zen 3 event. There's an RDNA2 event scheduled for the end of October. We could see a possibility of AMD announcing the Zen 3 release date to be late October / early November with the event being about a month ahead of time and then having RDNA2 launch shortly after it's event so that they overlap. That seems somewhat unlikely to me, but who knows for sure.

I wouldn't expect anything substantial as far as GPU information at the Zen 3 launch. If AMD has Intel beat on gaming then I expect them to show a Zen 3 CPU going against an Intel one using some NVidia GPU, probably a 2080 Ti. The GPU tease will probably be AMD running the same test, only this time with one of their own GPUs which gets better FPS. Obviously they could use something other than a 2080 Ti for the comparison, maybe a 3080 if they're being cheeky, but I don't think we get any substantial information about their new GPUs and even if they do a little hype tease, it's going to be in some game/benchmark that's most favorable to their products.

 

[A///]

I thought someone posted a tweet from someone at AMD (maybe it was a tweet from Lisa Su) about an RDNA2 tease at the upcoming Zen 3 event. There's an RDNA2 event scheduled for the end of October. We could see a possibility of AMD announcing the Zen 3 release date to be late October / early November with the event being about a month ahead of time and then having RDNA2 launch shortly after it's event so that they overlap. That seems somewhat unlikely to me, but who knows for sure.

I wouldn't expect anything substantial as far as GPU information at the Zen 3 launch. If AMD has Intel beat on gaming then I expect them to show a Zen 3 CPU going against an Intel one using some NVidia GPU, probably a 2080 Ti. The GPU tease will probably be AMD running the same test, only this time with one of their own GPUs which gets better FPS. Obviously they could use something other than a 2080 Ti for the comparison, maybe a 3080 if they're being cheeky, but I don't think we get any substantial information about their new GPUs and even if they do a little hype tease, it's going to be in some game/benchmark that's most favorable to their products.

They might shift that date up if they want to beat NVidia who are releasing the 3070 on the 29th. If the Navi2 cards are any good they'll be at least better or matching a 3070 for cheaper. IDK why AMD would shoot themselves in the foot by not launching earlier.

 

[DrMrLordX]

Standing on its own and offering unique capabilities is more important. And there AMD already does well, with I/O capability, core/thread count, energy efficiency etc. Especially in the datacenter this is not a sprint but a marathon run.

They did that with K8 Opteron and nearly got wiped out. They couldn't do to Intel what Intel eventually did to them.

Let's first see how Intel is doing with Ice Lake-SP, Sapphire Rapids' direct precursor, which should go head to head with Milan.

I'd love to know more about IceLake-SP. Still waiting.

For Sapphire Rapids to take away all of Epyc's (current and future) selling points it has to match both its I/O capability as well as it core/thread count.

Not exactly. Intel just needs a 10nm server product that can replace Cascade Lake-SP. That's it. They can bring along an enormous number of customers without even having to win any benchmarks. Assuming they can keep their customers distracted long enough for them to not notice that AMD has updated their server platform AND server CPUs three times prior to Sapphire Rapids' release. It's really not clear how many times AMD has to improve their products while Intel does essentially nothing to finally make loyal vendors say, "you know what? Forget your incentives Intel, we'd rather not be an Intel exclusive shop anymore".

IceLake-SP appears to be not capable of fulfilling that role due to yields. I guess the hope is that 10SFE and

That's across all different process nodes.

Okay, fair point, but I'm pretty sure Intel's 800k number is also across all their process nodes (or at least 14nm + 10nm).

Zen certainly was AMD's last straw, but it wasn't an all out financially risky move. In contrary they started with a slow roll out of different products in different market segments, which incidentally they still do today.

. . . sort of? The first Zen products out of the gate were Summit Ridge. Had Summit Ridge failed and EPYC failed behind it, that would have been it, kaput, finito. I doubt the second-wave Summit Ridge chips would have made it to market in significant quantity, nor would we have seen much of Raven Ridge either, much less Threadripper. They really only had two rollout phases to make or break their future. Fortunately things worked out for them.

I'm honestly glad AMD isn't as inpatient as you are.

Really? Why? Imagine this timeline:

Matisse: May 2019
Vermeer: August 2020
Raphael: Sept/Oct 2021

That, friends, would be beautiful, and good for everyone except maybe OEMs that got caught with their pants down by buying up too much stock of previous-gen CPUs (but really, I think they could manage). Imagine how embarassed Intel would have been had AMD managed to get Vermeer out before they could officially launch their awesome 4c 10SF mobile CPUs. And I'm not talking about some buggy not-as-performant Vermeer compared to the one we're seeing in a few days. I mean the same chip. Do you really think AMD wouldn't have had the silicon ready by then? The only thing I would worry about would be firmware, which admittedly was still a problem for Matisse, even though AMD held it until July 2019. ah heck that.

I highly doubt dies from early on of that sampling phase are usable for the final consumer product being launched soon.

I agree! We won't see package dates from January 2020 or anything silly like that. But May 2020? I could maybe see that. Pretty sure date-of-manufacture is going to be for the final packaging date rather than diffusion.

Oh. He's back.

Yeah, I gave up on you when you started accusing me of throwing toys, or something stupid like that.

Never thought I'd see someone advocate AMD to what Intel had done for years, even before their troubles began.

You just don't get it. AMD is becoming too much like Intel. They need to stop! AMD needs to behave like they have fierce competition even when they don't.

 

[jpiniero]

They might shift that date up if they want to beat NVidia who are releasing the 3070 on the 29th. If the Navi2 cards are any good they'll be at least better or matching a 3070 for cheaper.

I dunno, look at Zen 2's 'official' pricing. Brand for Brand it's not that much cheaper than Intel despite being slower in games. You'd have to think that the 3070/2080 Ti competitor would have more than 8 GB but I wouldn't expect it to be much less than $499.

Same deal with Zen 3. If they end up doing the 8 core and 12 core first I would expect the MSRP to be decently more than $399 and $499, maybe $449 and $549. The street price will of course vary.