New Zen microarchitecture details

[swilli89]

Prescott was a piece of crap. But it's interesting that you basically have to go back 12 years to find an instance of Intel truly falling flat on its face in the PC MPU market.

I'm merely recalling other instances of pre-release hype in good humor, and you immediately assume I'm trying to "level the playing field". And yeah it was a decade ago but if I remember correctly it was one of Intel's smaller design teams that derived the excellent Conroe architecture out of the Pentium-M.

I should know, because that was my third computer I built, with an E6600. I've had an XP-2500M, Athlon FX, E6600, I5-2500K.. the only trend you'll notice is that I get what gives me the most for my money. I have zero brand loyalty, but I do get excited about new stuff, wherever its coming from.

 

[Abwx]

Anyone who thinks that must be mad...there's just no way AMD has that much efficiency...because not even Intel is anywhere near that...and they got the budget to throw at such research.

And why not..?..

The datas released by GF did imply that the perf/watt improvement in respect of 28nm was roughly 2x (that s what i computed at the time) and the recently published slides confirm this with a 2.2-2.3x actual improvement..

 

[Shehriazad]

And why not..?..

The datas released by GF did imply that the perf/watt improvement in respect of 28nm was roughly 2x (that s what i computed at the time) and the recently published slides confirm this with a 2.2-2.3x actual improvement..

Which means that they could easily run their OLD arch at really low wattages...buuuuut....and here comes the but:

They promise:

- +40% IPC vs Excavator
- Entirely new Arch that will probably need revisions to be optimized fully
- Bunch of new tech that AMD hasn't really handled before
- incredibly small R&D funds
- insane core/thread amount for the wattage for a consumer CPU.



And I know it's apples and oranges...but look at Intel.

5690X:
8 cores 16 threads
3 Ghz BASE, 3.5 Ghz BOOST
22nm
140w


So let's say Intel ports this to 14nm....I do not think that they will manage to go below 95w.

Heck...even their 6700K 4core/8threads sits at 91W...and that chip is actually 14nm.

And I don't think AMD will have quite mastered the tech just yet.

So even if IPC was the same as Skylake...again dreaming here...you'd probably look at 4 Ghz with 95w tdp.



And yes I'm well aware that their CPUs function a bit different....but it seems pretty clear that AMD as well won't be pushing it beyond 3.5 Ghz boost for their 16thread monster...they might even end up boosting less...just because they are limiting their TDP this hard.



Again...I'm really not hating on AMD here, in fact the majority of CPUs I owned in my life were AMD CPUs...but expecting them to pull off a magic trick like a 4ghz 16thread machine with comparable performance to current gen Intel stuff with just 95W is pretty much asking for a miracle...and I'm speaking Moses parting the ocean level here.

If I end up being wrong, feel free to return to me to try and shame me...but I'll be long gone...playing on that AMD CPU. xD

 

[nerp]

Look, I built AMD systems exclusively for years and years. I am running a Skylake chip now. As much as I want Zen to be amazing, if it only matches Haswell, we all should be disappointed. AMD needs something to match or beat Intel at the same or less price. The idea here is that AMD needs to fill a product niche and be profitable at the same time. The only way that will happen is for the chips to either have blockbuster performance or offer the same performance at less cost without much compromise in terms of power usage or heat. Yes, the industry needs competition and we all benefit if AMD succeeds. But they have to deliver something that will find market success. Dreams do not meet payroll.

 

[Abwx]

Which means that they could easily run their OLD arch at really low wattages...buuuuut....and here comes the but:

They promise:

- +40% IPC vs Excavator
- Entirely new Arch that will probably need revisions to be optimized fully
- Bunch of new tech that AMD hasn't really handled before
- incredibly small R&D funds
- insane core/thread amount for the wattage for a consumer CPU.

Now that they took this path so good but i m convinced that an evolution of EXV modules would yield a higher throughput despite the lower ST IPC.

And I know it's apples and oranges...but look at Intel.

5690X:
8 cores 16 threads
3 Ghz BASE, 3.5 Ghz BOOST
22nm
140w

So let's say Intel ports this to 14nm....I do not think that they will manage to go below 95w.

At 14nm they would get 115W, and by removing two memory channels they can eventualy grab 10W more to get to 105W, so that would be feasible with a 10% improvement of their current process and a slight ipc loss, wich is doable on the mid term.

Heck...even their 6700K 4core/8threads sits at 91W...and that chip is actually 14nm.

And I don't think AMD will have quite mastered the tech just yet.

There s nothing more to master, a transistor is a transistor and can be straightfowardly replaced by a better one without changing anything to the electrical schematic, what need to be changed is the layout.

And yes I'm well aware that their CPUs function a bit different....but it seems pretty clear that AMD as well won't be pushing it beyond 3.5 Ghz boost for their 16thread monster...they might even end up boosting less...just because they are limiting their TDP this hard.

Highest frequency is impossible to predict accurately with the available infos about the process, moreover given that it s a new uarch although it looks like a heavily beefed up EXV with doubled Integer exe ressource and an enhanced FPU since the one in EXV is already up to the task.

 

[deasd]

The L1 latency could be a hint that they are targeting high clocks. That's nothing solid, of course.

Haswell also has Latency of 4. It's understandable that reaching high clock is important at current situation which ILP and IPC hit a wall.

OTOH we don't know how many FO4 delay does Zen has, so it's hard to judge Zen to have high frequency design as bulldozer, I bet it to be little bit higher delay than K8 but lower than bulldozer.

BTW I have never seen AMD design such a wide core, this is already far better than bulldozer and K10 at this point. The only problem that prevent AMD is the yield which give bulldozer so much trouble.

(I feel disappoint to see so much 0 value post in this thread, and of course it's foreseeable.)

 

[YBS1]

Look, I built AMD systems exclusively for years and years. I am running a Skylake chip now. As much as I want Zen to be amazing, if it only matches Haswell, we all should be disappointed. AMD needs something to match or beat Intel at the same or less price. The idea here is that AMD needs to fill a product niche and be profitable at the same time. The only way that will happen is for the chips to either have blockbuster performance or offer the same performance at less cost without much compromise in terms of power usage or heat. Yes, the industry needs competition and we all benefit if AMD succeeds. But they have to deliver something that will find market success. Dreams do not meet payroll.

I don't think they have to actually beat or match Intel for this to be a success. I'm of the opinion if they could offer 6C/12T and 8C/16T CPUs that are within reasonable striking range of Haswell's single threaded performance and offer those at a price slotted around the top of Intel's quad lineup they might rack up some serious volume. How Intel would respond to this is the question.

 

[Dave2150]

And they originally wanted 850$ for Fury X if we are to believe rumours. But the 800$ 220W spaceheater was enough evidence.

But I like to see the product first. It may be a 200$ product because it performs like a 200$ product.

Performance/watt should hopefully improve drastically.

Spaceheater? 980ti's TDP is very similar to the FuryX's, especially when overclocked.

In fact an overclocked 980ti's TDP is higher than an overclocked FuryX.

 

[Madpacket]

Seems everyone's forgotten who's mainly responsible for the design this CPU. Bulldozer wasn't handled by the same core designer who's name I've gladly forgotten and given Jim Keller's track record I doubt we're going to see a repeat of that mess.

This chip needs to perform, no it will perform or AMD as we know will be sold off in parts with RTG being the first to go.

Even the Intel fanbois here better check their ego's and hope Zen pans out or we're all going back to paying for even more overpriced quad core X86 CPU's with 5% IPC every tick or tock. I mean look at Intel's margins, besides them only Apple can get away with this kind of highway robbery.

 

[sandorski]

Looking forward to its' release and reviews. Even if it falls just shy of Intel's best at this time, it will improve AMD's market position and start putting the pressure on Intel again.

 

[Madpacket]

Spaceheater? 980ti's TDP is very similar to the FuryX's, especially when overclocked.

In fact an overclocked 980ti's TDP is higher than an overclocked FuryX.

And at stock the Fury X is overall a slightly faster card, comes with a quiet cooler that allows for better multi card setups with better scaling. Oh and it's cheaper and will likely age better.

 

[videogames101]

Hmm, looks like every other CPU microarch in the last 4 years. Which isn't so strange being how every company reads the same whitepapers.

It's all about execution, we'll have to wait and see if AMD has it.

 

[Cookie Monster]

Nice one OP. Really good analysis and from the looks of it, does sound promising. This just may be what AMD needs (a large step in the right direction).

Tbh, it will all come down to price. If they can atleast have competitive products (maybe not as fast but somewhat competitive against Intel) it could shake up the market given how they price these.

I just realised the other day that they've been sitting on the same MB chipset/CPU uarch for YEARS.

 

[richierich1212]

Thanks Dresdenboy. Hopefully AMD can turn things around with Zen.

 

[beginner99]

And I know it's apples and oranges...but look at Intel.

5690X:
8 cores 16 threads
3 Ghz BASE, 3.5 Ghz BOOST
22nm
140w


So let's say Intel ports this to 14nm....I do not think that they will manage to go below 95w.

TDP isn't power usage and there are multiple ways to calculate it. This is best seen in GPU where Nvidia since years basically has been lying about their TDP, eg. actually power use isn't that much smaller than AMD but the TDP is usually 50% less than for AMD comparable card. AMD could do the same thing here in CPU space.

Plus Skylake TDP is with the iGPU. If you disable it the CPU itself won't use that much.

 

[ClockHound]

Thanks, Dresdenboy, for posting your Zen overview in these sharktel infested waters. It's refreshing to read informative posts about highly technical stuff. Engineering stuff that seems to trigger schadenfreude emotional responses in some.

Someday I may understand the fanboy/tribalism that exists around impersonal commercial brands. Especially in the GPU and CPU areas. Probably more in the GPU market, where the target demographic appear to be more easily influenced by social engineering than the actual hardware engineering.

Has modern marketing become so successful that consumers must belong to a brand in order to belong?

 

[Carfax83]

Spaceheater? 980ti's TDP is very similar to the FuryX's, especially when overclocked.

In fact an overclocked 980ti's TDP is higher than an overclocked FuryX.

Overclocked 980 Ti is also at least 20% and more faster than overclocked Fury X, which justifies the higher TDP. Also, 980 Ti isn't using a liquid based cooler, which lowers the TDP..

Anyway, back on topic. I really hope Zen performs well. An AMD slot A Athlon CPU was my first self built PC, so I have a soft spot for AMD and I want them to succeed.

 

[IllogicalGlory]

Spaceheater? 980ti's TDP is very similar to the FuryX's, especially when overclocked.

In fact an overclocked 980ti's TDP is higher than an overclocked FuryX.

He means the FX-9590 that was originally sold for $800 and OEM-only, which is admittedly incredibly stupid.

On the other hand, I agree with your post. People love to apply the "spaceheater" moniker to AMD anything, especially video cards, even if the NV competitor is a mere 5-8% more efficient. The R9 290(X) series are the biggest targets for this, obviously. I reckon it's because of their reputation for running hot, never mind that GPU temperature has literally nothing to do with how much heat it emits.

 

[ThatBuzzkiller]

You need more than atomics to support HW transactional memory or do you even need them at all ?

Both are related to the concept of synchronization but their somewhat orthogonal. With atomics you need locks for synchronizing access to a region of memory. With HW transactional memory we now use "transactions" instead to define the critical sections of the code that must have restricted access by some threads. One special attribute unique to transactions is that they have the capability to restore the previous state just before beginning the transaction when there is an access conflict between threads by aborting the current transaction ...

 

[looncraz]

It's not about whether or not you could have bought HSW-E performance for the last 18 months anyway. It's that AMD is going to be selling 8 core Haswell-E levels of performance in the same $100-$200 range as the current octo-core FX processors. I don't recall the exact quote, but I think Lisa Su herself said something along the lines of "AMD cherishes its history as a value leader, and seeks to continue being a budget brand going forward." It was something like that, at least.

They most certainly will not be selling CPUs like that for that price range, that would be suicidal.

Zen is exciting because it represents AMD returning to being a viable alternative for the average, gaming, and even high performance and server systems.

If a 6C Zen costs $200, it will be performing like an i5 and no better. AMD will price similarly to Intel, except - perhaps - at the high end.

I *want* a 6C/12T Zen to cost exactly as much as the same from Intel. AMD's platform will be cheaper - enough to bring an easy $50+ price advantage by itself.

In the spring of 2017, I want to spend $150~$175 for a motherboard, $350~400 for a CPU, and $150 for DDR4 RAM. Which is to say, I want to spend $700 on a full-on AM4 upgrade to my i7-2600k@4.5GHz. I don't need any more IPC than I have, in fact, but I can absolutely use more cores.

At the same time, Intel should have already updated their HEDT platform, so Zen's prices will probably be a bit lower, but that just depends on Intel's pricing and market availability.

I'm most excited, though, about being able to recommend AMD systems again. AMD really only fits for HTPC applications these days, as even a lowly Pentium offers better performance per dollar every way you look at it for every other possible application. APUs are capable enough for people who like to play on some of their old games, for example, but if you need to use a dGPU, Intel is the only logical choice.

 

[turtile]

Why would Zen ship with a 6 core product? They are in 4 core clusters instead of the current 2 in a module. As for frequency, I remember AMD's target was 3Ghz way back but that could have changed by now.

 

[JDG1980]

And I know it's apples and oranges...but look at Intel.

5690X:
8 cores 16 threads
3 Ghz BASE, 3.5 Ghz BOOST
22nm
140w

That official TDP is substantially overstated. Tom's Hardware ran some tests and determined that, at a full stress load, the i7-5960X uses an average of just 106W and peaks at 122W. (This is measured at the 12V input, so it includes the FIVR.) Intel could have claimed a TDP of 125W or even 110W without stretching the truth much.

When overclocked to 3.5 GHz, the same i7-5960X exhibited 121W average, 141W maximum. Pushing it to 4.0 GHz spiked power usage to 146W average, 165W maximum. They also ran a test at 4.0 GHz with two cores disabled; in hex-core mode, it averaged just 115W and peaked at 137W.

Now consider that this was all at 22nm. I know Intel's 14nm process is slightly denser than Samsung/GloFo/TSMC FinFET, but the foundry FinFET processes still beat Intel 22nm by a decent margin. It seems quite possible that AMD could do an 8C/16T chip with 95W TDP at 3.0 - 3.5 GHz base clock (and higher turbo on lightly-threaded workloads), especially if this TDP corresponds more to the "average" figure under full load rather than "maximum" momentary spikes. For the cut-down 6C/12T chips, 4.0 GHz might well fit within the 95W TDP budget, if the architecture and process node can handle it. Assuming IPC is in the Sandy Bridge to Haswell range, I could see the 6C/12T version of Zen becoming a very popular chip if it's sold at a competitive price ($249-$299, roughly in line with what Thuban cost when it was first released).

Was it ever confirmed if Summit Ridge would have dual-channel or quad-channel RAM support? Haswell-E has quad-channel RAM, and if Summit Ridge can get away with a smaller memory controller, that could be a considerable savings in both die space and power usage. Dual-channel would probably be enough for HEDT, while the big server applications would wait for Zeppelin (assuming it's not just a rumor).

 

[NTMBK]

That official TDP is substantially overstated. Tom's Hardware ran some tests and determined that, at a full stress load, the i7-5960X uses an average of just 106W and peaks at 122W. (This is measured at the 12V input, so it includes the FIVR.) Intel could have claimed a TDP of 125W or even 110W without stretching the truth much.

What did this "torture test" consist of? Did they put a large continuous load over the PCIe links, too? Driving 40 PCIe 3.0 lanes is going to be awfully power hungry, and that controller is integrated into the CPU.

Was it ever confirmed if Summit Ridge would have dual-channel or quad-channel RAM support? Haswell-E has quad-channel RAM, and if Summit Ridge can get away with a smaller memory controller, that could be a considerable savings in both die space and power usage. Dual-channel would probably be enough for HEDT, while the big server applications would wait for Zeppelin (assuming it's not just a rumor).

The AM4 platform is only dual channel, so if Summit Ridge is going to share a socket with Bristol Ridge then it has to be dual channel too. Hopefully it supports some high memory speeds!

 

[ShintaiDK]

Spaceheater? 980ti's TDP is very similar to the FuryX's, especially when overclocked.

In fact an overclocked 980ti's TDP is higher than an overclocked FuryX.

FX 9xxx series.

The Fury X was used as price compare that AMD doesn't give anything away for cheap because fans dreams of it.

 

[jihe]

It is a fact -- before BD launched, these forums were buzzing with claims that AMD was going to kick SNB into the stone age with 8 cores, more IPC than Phenom II, etc. Your join date is 2008, so I am sure you saw all of the internet buzz around Bulldozer, too.





That hype led to a lot of disappointment/angst, particularly among those who held off on buying SNB because they were told to "wait for Bulldozer."





Zen might be different, but to borrow the words of swilli89, "knowing AMD" it'll under-perform comparable Intel parts. Past performance is the best indicator of future performance and AMD has a lot to prove in CPUs.

Knowing AMD? Past performance??AMD has a lot to prove in CPUs????? Were you not around in the a64 days?