Speculation: Ryzen 4000 series/Zen 3

[dnavas]

By those, todays prices are cheap, you all are spoiled (that are whining about price)

One would hope so! I remember very well the price of my wife's original PC, and my A3000. I think I paid close to four figures(*) for an expansion box with 2MB of RAM in it. Actually, you have a point about the step-wise nature of perf/price advances, but agreeing is no fun, and I've been dragged into this, so in the finest tradition of sunk-cost fallacy....
https://aiimpacts.org/trends-in-the-cost-of-computing/ would have you believe that gflops/$ is supposed to be increasing 10x over five years. This is likely driven by GPU advances over the timespan measured, because anyone in their right mind ... but, what the heck, running the numbers:
x^60 = 10; x^15 = 1.8. If I was expecting an 80% improvement over 15 months, I think you'd rightly call it whining :>

Personally I'm more hoping for ~2x over 5 years. This past decade has been a disaster, and no one needs to take me to task over Intel -- I've been all over other forums kvetching about it. I'm personally in a bad spot because I'd like to see them take a long walk off a short cliff, but competition would actually be better.... At any rate (and I *swear* I did not do the math beforehand):
x^60 = 2; x^15 = 1.196. So, yeah, I'm hoping for about 19% (?!) performance improvement over 15 months with inflation-adjusted pricing. For anyone looking at the low-end price adjustments, you're getting just a bit more than inflation-adjusting increases

So, now I'm in the unenviable position of relying on Intel to create competitive products.

(*) [ed: http://amiga.resource.cx/adcoll/preview/ExpansionTechnologies_1987-09.jpg --I did not misremember!]

 

[eek2121]

Off topic, but, I wrote code on machines with turbo buttons on the front panel. That was long before 100Mhz was even a thing! Now give me back my cane and get off my lawn!

Same!

Regarding Rocket Lake, word on the street is Intel is going to ship a portable nuclear reactor/1000W chiller combo with each purchase.

 

[Markfw]

One would hope so! I remember very well the price of my wife's original PC, and my A3000. I think I paid close to four figures for an expansion box with 2MB of RAM in it. Actually, you have a point about the step-wise nature of perf/price advances, but agreeing is no fun, and I've been dragged into this, so in the finest tradition of sunk-cost fallacy....
https://aiimpacts.org/trends-in-the-cost-of-computing/ would have you believe that gflops/$ is supposed to be increasing 10x over five years. This is likely driven by GPU advances over the timespan measured, because anyone in their right mind ... but, what the heck, running the numbers:
x^60 = 10; x^15 = 1.8. If I was expecting an 80% improvement over 15 months, I think you'd rightly call it whining :>

Personally I'm more hoping for ~2x over 5 years. This past decade has been a disaster, and no one needs to take me to task over Intel -- I've been all over other forums kvetching about it. I'm personally in a bad spot because I'd like to see them take a long walk off a short cliff, but competition would actually be better.... At any rate (and I *swear* I did not do the math beforehand):
x^60 = 2; x^15 = 1.196. So, yeah, I'm hoping for about 19% (?!) performance improvement per year with inflation-adjusted pricing. For anyone looking at the low-end price adjustments, you're getting just a bit more than inflation-adjusting increases

So, now I'm in the unenviable position of relying on Intel to create competitive products.

And its not just CPU's. I will never forget, my first ONE GIG hard drive cost almost $1000 . It was SCSI, but at that time, no IDE existed like that. And it was 5 1/4 full height and weighed a ton.

 

[jamescox]

Really, "All Up in Arms" -- uppercase, even? I think you're over-reaching. Did you not read the "I'm far more interested" part of that post? There's a lot more there there. If I'm up in arms about something, it's about the utter lack of technical detail. Regardless I'm not "All Up in Arms". Goodness -- sensitive much?

But, if you want to talk pricing ... I don't know how old you are, but I'm from a time when processors got 100% faster every year and a half, and prices were as likely to slip as rise. So, yeah, 20% is not impressive, neither is paying twice as much for twice the cores. The whole point of the industry is perf/$ -- 20% perf gain minus 10% price rise == 10% value increase. As I posted somewhat long ago, my only real hope is that 39x0 TR pricing will see a reset. We'll see :shrug:

Complaints about the tiny price increase are ridiculous. I still use an i7-920 for a Linux machine (24 GB of ram and an SSD). It is a 920 (2.67 GHz) because the 975 (3.33 GHz) was $999 and the 920 was only $284. That is $715 for an extra 666 MHz, and that is it. I guess you guys don’t want AMD to make enough money to continue to break up the Intel monopoly? Do you really want to go back to Intel monopoly pricing? If it AMD had not come along with Zen, Intel would have held onto 4 core as long as they possibly could to prevent cannibalizing much more profitable Xeon sales. we would have been paying really high prices for less cores than are in my phone if it hadn’t been for Zen. AMD selling 16 core / 32 thread parts not only hit Intel desktop parts, it also hit their much higher priced Xeons. I am still stuck with a 4 core skylake Xeon desktop at work that probably was more expensive that 12 or 16 core Ryzen of the time. I would love to have one of the 16 core processors for compiling. I often get on a bigger machine in the server room, so the 4 core Xeon has turned into an expensive vnc viewer box.

If anything, the price increase is very conservative. This gives me some hope that they have very good supply. Also, comparing the value increase over Ryzen 3000 is also ridiculous. Most people that have a ryzen 3000 cpu do not need to upgrade to a Ryzen 5000. Most games will run perfectly fine on really old cpus. LTT did a video where they tested old cpus with a newer upper midrange (2060 super maybe), and the really old cpus still did fine. I don’t remember which ones they tested; back to 2012 (edit: 2012, not 2023; I am not from the future) maybe.

Moores law actually died around the 2006-2007 timeframe when they did a process shrink and power consumption went up rather than down due to leakage. That is why old core 2 duo and core 2 quad machines still do very well. I still use a 17 inch core 2 duo MacBook Pro. With a fast samsung ssd in it, it still does everything I want it to fine (mostly web browsing and YouTube). When it dies (has had some apparent gpu related crashes; it has 2 Nvidia GPUs, one for internal display and one for external), I am going to want a higher end AMD laptop, so hopefully the OEMs deliver that. I might buy a desktop Ryzen 5000 also, but I have exceptionally old hardware.

Going from a Ryzen 3000 to 5000 may be worthwhile for some people. If a lot of them upgrade then there may be a lot of used 3000 cpus for cheap. It wouldn’t be hard, jut upgrade the bios and swap the cpu. I wouldn’t recommend that to my friends though, unless they really need a new system. Some of my friends are using very old systems, but it doesn’t take much to run WoW. They probably don’t need the performance and while Ryzen 5000 looks great, it is in a strange spot. If I buy a motherboard and memory now, this is probably the last AM4 DDR4 Zen cpu. I am unlikely to upgrade it soon; I keep hardware a long time, so it really isn’t an issue for me. For more frequent upgrader, it is a consideration.

Zen 4 is probably going to be a bigger change in some respects. I suspect the integer units will be similar to Zen 3, but probably radically new floating point. Zen 4 will probably be s stacked device of some kind so they will probably widen the data paths significantly. The 32-bit serdes based, on package, infinity fabric will probably be replaced with something like 1024-bit interposer type links. The data paths will need to widen to accommodate that. It has been (I think) 256-bit internal links (in each direction) with Zen 2. Zen 3 may just be higher clock IO die. Zen 4 they probably go to 1024-bit or more with active interposer or other stacked die. They will be able to use HBM and GPU chiplets also. I don’t know if we would see HBM on a desktop part though. With the way the Zen 3 die looks, I am wondering if Zen 4 chiplets will be 2 CCX with 16 cores total at 5 nm.

 

[dnavas]

And its not just CPU's. I will never forget, my first ONE GIG hard drive cost almost $1000 . It was SCSI

Yep, I bought one of those at that price as well.
...and I remember being giddy about it too! Hard drive prices have kind of flattened in recent years. SSD pricing is still being driven down, which is great, but it's hard to get the same kind of excitement when it's basically treading over the same area that spinning rust was at a decade (?) ago. Network gear is finally starting to see some life again. And of course TR (almost on topic) has been a true breath of fresh air, if only they hadn't had to go mucking about with the socket.

Ryzen 1800X for $500, ThreadRipper 16C/32T for $1000.

For the record, I own both. Exactly because they were great values for the price. Looking forward to the next one!

 

[nicalandia]

For the record, I own both. Exactly because they were great values for the price. Looking forward to the next one!

Someone here was stating that they don't like paying double for twice the cores...… So I take it that at $1000 that 16C/32T was "unreasonable" priced... goes to show you that some people have been spoiled by AMD too much, just a few years ago you had to pay $500 more for two extra cores(Intel 6950X).

 

[Markfw]

Someone here was stating that they don't like paying double for twice the cores...… So I take it that at $1000 that 16C/32T was "unreasonable" priced... goes to show you that some people have been spoiled by AMD too much, just a few years ago you had to pay $500 more for two extra cores(Intel 6950X).

At the time, I got up to 5 1950x's and most were at $900-$1000, and at the time I thought they were great. Now I am down to one, and have changed over to EPYC for more cores, and a 3950x for speed and a few cores.

Today ? That would be a horrible value (the $1000 for 16 cores) And those are slow. My 3950x is twice as fast as my 1950x, and for $250 less.

Computers get worthless fast. I only got like $300 for my 1950x's....

 

[dnavas]

Moores law actually died around the 2006-2007 timeframe when they did a process shrink and power consumption went up rather than down due to leakage.

Dennard, not Moore. https://en.wikipedia.org/wiki/Dennard_scaling#Breakdown_of_Dennard_scaling_around_2006
Common mistake, and I can never remember the right name either :shrug:

Zen 4 is probably going to be a bigger change in some respects.

I'm quite looking forward to CCIX, pcie5, and heterogenous compute platforms. I've been saying for years now that replacing half of the chiplets with targeted accelerators would be *amazeballs*

Someone here was stating that they don't like paying double for twice the cores...… So I take it that at $1000 that 16C/32T was "unreasonable" priced... goes to show you that some people have been spoiled by AMD too much,

Strictly speaking, the 2950x I bought was less than twice the cost of my 1800X (1.8x) and it was rather faster than twice as fast. The teething problems with TR1 were enough to keep me away. The 3970x was more than twice the cost of the 2950x (2.2x) and required yet another motherboard. I drew a line. Anyone is, of course, free to disagree with that line, but that's where I drew it. If I were still using the 1800x box, the 3900x would be an amazing replacement. However, I'm far more interested in the TR line at this point, which is why I'm more interested in the wider FE than desktop price increases (I am worried about what they portend for TR4, though).

 

[Veradun]

Welp, looks good. I'm a bit disappointed there's no 5700x part, because I don't want to pay $450 for an 8 core CPU. I'm a bit surprised the 5600x will be $300. The $50 increase in most stacks makes sense, but feel like that's a bit much for a more mid ranged part. Thinking I'll probably upgrade my 1600 to a 5600x, then.

The Elephants in the room are the very very obvious 5600 and 5700x parts that will be released later. Premium expensive parts first, value parts last is an industry standard.

You'll probably be able to get them in Q1.

 

[eek2121]

The Elephants in the room are the very very obvious 5600 and 5700x parts that will be released later. Premium expensive parts first, value parts last is an industry standard.

You'll probably be able to get them in Q1.

I have read from a couple different sites that non-X parts are no longer going to be sold.

 

[moinmoin]

I have read from a couple different sites that non-X parts are no longer going to be sold.

Not at launch at least. And AMD of course wants plenty demand nevertheless.

I think Ian Cutress discussed the whole issue well in his article:

The success of AMD’s Ryzen 3000 family is easy to see. It offers a very attractive performance per dollar proposition, and nine of the top ten processors on Amazon’s best sellers list are from AMD. Here’s that list, with the pricing as given in our buyer’s guide last week:

Price Options [#] is Amazon Best-Seller Position
#	AMD	Price
[1]	Ryzen 5 3600	$205
[2]	Ryzen 7 3700XT	$295
[3]	Ryzen 5 3600X	$209
[4]	Ryzen 5 2600	$149
[5]	Ryzen 9 3900X	$430
[6]	Ryzen 7 3800XT	$340
[7]	Ryzen 5 3600XT	$230
[8]	Ryzen 7 2700X	$218
[9]	Ryzen 3 3200G	$100

This is an impressive list, however there a theme I have noticed. Starting from the top, the first four best-selling processors are below $299. The lowest Ryzen 5000-series processor in this launch comes in at the $299 price bracket. Out of that top nine, only #5 and #6 are above $299. This is a clear indication of where the bulk of the market is, especially as AMD is likely in increasing both market share and revenue.

In discussing the pricing with AMD, I noted that the new Ryzen 5000 processors are not only replacing hardware with a $50 higher MSRP, but also replacing hardware that routinely sells below MSRP. This makes the differences more akin to $90-$150. This of course changes some of the dynamic when we start discussing performance per dollar.

AMD’s response to this commentary was the one I would have given if I were in their position. The Ryzen 5000 series is a new product, and the claim of market leading performance means that the early adopters and AMD enthusiasts that want the best on day one will be able to get the hardware they desire. During the initial phase in almost all launches, users looking for the best bang-for-buck build will always look to purchasing the previous generation, which is almost always offered at a good discount as stock transfers to the latest product. AMD believes it has set the pricing of the new Zen 3 processors where it remains competitive, but still balances the message that AMD claims it has the best, most efficient processors.

Lower price point SKUs will come given time.

 

[amd6502]

Just few days to confirmation Zen3 has SMT4!

not in 5nm? i believed in you nosta! also no smt4? oops.

Well, almost certainly not for PC anyway, and definitely not enabled by default. For server most likely also not until Zen4.

IMHO there's somewhat of a chance that the FPU code might be running SMT4 on core pairs underneath the hood. It'd still be SMT2 to the user.That could explain an almost 20% IPC boost for single thread float loads.

I was hoping about more details from the Oct 8 presentation.

The only thing suggested was an improved front end via the "zero bubble branch prediction". But it would be surprising if they got this 19% ipc integer boost without widening the core from Zen2's 4+3 integer core.

 

[jamescox]

Dennard, not Moore. https://en.wikipedia.org/wiki/Dennard_scaling#Breakdown_of_Dennard_scaling_around_2006
Common mistake, and I can never remember the right name either :shrug:



I'm quite looking forward to CCIX, pcie5, and heterogenous compute platforms. I've been saying for years now that replacing half of the chiplets with targeted accelerators would be *amazeballs*



Strictly speaking, the 2950x I bought was less than twice the cost of my 1800X (1.8x) and it was rather faster than twice as fast. The teething problems with TR1 were enough to keep me away. The 3970x was more than twice the cost of the 2950x (2.2x) and required yet another motherboard. I drew a line. Anyone is, of course, free to disagree with that line, but that's where I drew it. If I were still using the 1800x box, the 3900x would be an amazing replacement. However, I'm far more interested in the TR line at this point, which is why I'm more interested in the wider FE than desktop price increases (I am worried about what they portend for TR4, though).

I am actually still not sure that AMD will ever support AVX512. It is still very niche and probably should be replaced with something better. If they can use an RDNA or CDNA chiplet that uses the same cache coherent memory space as the cpu, then you get a huge amount of FP power without the latency of a separate GPU.

 

[dnavas]

I am actually still not sure that AMD will ever support AVX512. It is still very niche and probably should be replaced with something better.

Probably so.

If they can use an RDNA or CDNA chiplet that uses the same cache coherent memory space as the cpu, then you get a huge amount of FP power without the latency of a separate GPU.

Yeah, but it's difficult to embed a call across processors in a serial stream of code. Cache coherency isn't the whole story, sadly, but the ability to minimize round-tripping through memory does give you some decent benefits. Regardless, you need the attention of programmers to use it, and CUDA and quicksync currently have that attention. This is going to be an uphill slog for AMD, but at the same time, in the best of worlds, it could be a huge opportunity.

 

[AmericanLocomotive]

I really don't understand people defending the price increases. It's +$50 across the board, which impacts the lower end parts a lot more. Plus with the removal of the stock HSF (which while yes, many people didn't end up using it, it will still cool the CPU adequately), the effective price is really more like +$80 across the board compared to last gen CPUs. A lot of the value is gone.

 

[turtile]

I really don't understand people defending the price increases. It's +$50 across the board, which impacts the lower end parts a lot more. Plus with the removal of the stock HSF (which while yes, many people didn't end up using it, it will still cool the CPU adequately), the effective price is really more like +$80 across the board compared to last gen CPUs. A lot of the value is gone.

I figured this would be the case after the XT series was released so close to the Zen 3 launch. At the same time, if they priced the new CPUs at the same price as Zen 2, they would only hurt their own margins. I think we will see a price drop once Rocket Lake comes out and the inventory of Zen 2 is lower.

 

[USER8000]

I really don't understand people defending the price increases. It's +$50 across the board, which impacts the lower end parts a lot more. Plus with the removal of the stock HSF (which while yes, many people didn't end up using it, it will still cool the CPU adequately), the effective price is really more like +$80 across the board compared to last gen CPUs. A lot of the value is gone.

AMD fans attacked Intel and Nvidia for price increases. Now its OK because AMD. Ryzen 5600x is 3600 successor(same cooling needs,and same wraith stealth cooler).

 

[SPBHM]

supply must be pretty limited to justify those prices...

 

[Antey]

now we can only hope for intel to release a competitive cpu lineup and make amd lower their prices.

 

[blckgrffn]

I am not going to defend price “increases” (is it an increase when it is the price of a brand new SKU 🤔) but did they say that the 3xxx line of Ryzen is discontinued as of immediately?

AMD did not just stop selling CPUs under $299 as of November 5th, they just might be one step back from latest and greatest. If you are on a budget, Zen 2 is still extremely viable. I just built a 3400G rig *today* and it’s only Zen+ but so capable for the work machine that it is. There is no Intel equivalent (GPU power) even today on the desktop.

Also, I believe the new 12c part in particular is still a crazy value compared to the CPUs of the not-so-distant past. So much cache!

 

[cellarnoise]

Competition is great! We all need more!

I am biased though as I wish Zen3 had even more performance for the $, but after buying and still on Zen1, I might just purchase.

We should all hope that performance per $ keeps improving as the current manufacturing tech keeps climbing fast. Let along the geo-political issues with tech advancement.

10 to 15 years ago were amazing years for cpu power. Maybe quantum type stuff will get us back to the recent old days?

 

[ThatBuzzkiller]

I am actually still not sure that AMD will ever support AVX512. It is still very niche and probably should be replaced with something better. If they can use an RDNA or CDNA chiplet that uses the same cache coherent memory space as the cpu, then you get a huge amount of FP power without the latency of a separate GPU.

What is this "something better" ? And don't say it's GPUs since they still can't even run standard C++ ...

It's bad enough that some or all parts of the silicon can't even run standard C++ code but it sucks even more for a programmer that they'll need to deal with 2 different compilers. 1 for the CPU/host code and 1 for the GPU/device code ...

We can barely convince AAA game developers to adopt stupid concepts like shading languages or shader IRs so can you imagine how well it'll go down with many other smaller developers needing to use another proprietary vendor compiler ?

 

[A///]

It's a safe bet to assume 5% better IPC than Tigerlake at this point. I guess AVX512 stuff will be last area where Icelake derivatives will be ahead. Unless Tigerlake outclocks Cezzane by more than 5% while running at lower power and adding 2x cores and Rocektlake (which will have lower IPC than Tigerlake) magically clocks to 5.5Ghz, AMD will have performance and perf./watt lead in desktop, server and laptop segments.

TGL uses an onboard accelerator. That alone helps it across the board. It's also why AMD is looking into buying Xilinx on stock.

 

[A///]

Although to be fair the presentation lacked a lot in technical detail, so I expect we'll continue to hear from the peanut gallery. "Wider FE" tell me more? Marketing took front seat, which explains the pricing as well, and is exactly as I feared -- the piper is being paid. The downside is exactly as yeshua put it -- "Customers first, my buttocks." AMD can't afford to lose customer enthusiasm, and raising prices in challenging times is tone deaf.

We'll see more details in the coming weeks. This isn't Hot Chips. I would say the high prices give AMD to breath. They did state prices are subject to change on their slides. If the think RKLS is a threat, they can slash prices and come out on top like sweet cream in milk.

 

[A///]

It's funny AMD has leaked yet to be announced 5900XT and 5950XT in their own presentation:

@DrMrLordX is quietly raging at that.