New Zen microarchitecture details

[Doom2pro]

It's becoming apparent that AMD hit a walkoff grand slam in game 7 of the world series... I can't even begin to imagine how employees at Intel feel right now, knowing they lost the big game to a team with 1/5th the payroll... Amazing.

The Desktop market is back, can't wait to delve into it

 

[HurleyBird]

Stupid and childish chipset naming, having to have a number that's one bigger than Intel.

In a way I agree with you, especially if AMD's CPUs were still performing poorly it would come off as pretty desperate. But with the way things are shaping up, it's like they're just rubbing salt in the wounds and trolling Intel. It still comes off as somewhat dickish, but I don't think it makes them look like a "wannabe copy cat" that it otherwise would have.

 

[Atari2600]

It still comes off as somewhat dickish

With the way Intel acted the last time AMD had the performance lead, IMO, f**k 'em.

 

[Doom2pro]

Stupid and childish chipset naming, having to have a number that's one bigger than Intel.

Considering what Intel did to them during the Athlon 64 days, who cares? Paybacks a pregnant dog.

 

[laamanaator]

Except, manufacturers could make no frills boards with good vrms and only the necessary ports. This way, a320 would come in handy since it would drive the price down. Or even buying a cheap board no so good vrms for overclocking lower power (50W ?) QUAD core parts. It's stupid to lock down such features via chipset.

Tell me, why would companies like to bring down prices? What's the benefit for them in that kind of practice? Hint: none. AMD was wise to lock overclocking off from the cheapest of boards. Basically all AM3+ boards had very cr*ppy VRMs up to and beyond the 100€ mark. Asus was one of the few (In reasonable prices, the only) who made quality boards with good VRM for about a 100€. All others would barely be able to support 125W parts without VRM throttling. I think that AMD expected that that'll probably happen too for AM4, so it's better to lock OC off from the cheapest boards than have bad PR when a bad quality board has burnt someones house down when VRMs exploded.

 

[sm625]

Stupid and childish chipset naming, having to have a number that's one bigger than Intel.

Yeah but 18 months from now, Intel and AMD will both have x370 boards on the market. And Intel's will be brand new, so buyers will be tempted to assume that AMD's are also brand new. Thus, more sales. Seems rather smart to me, if not obviously dirty and low.

 

[Sven_eng]

Stupid and childish chipset naming, having to have a number that's one bigger than Intel.

The 7700K is worse, having one thousand bigger number for the same CPU.

 

[krumme]

Canard at the 3.6 4.0 stepping
"
Canard PC Hardware ‏@CPCHardware
@Dresdenboy Yep. Seems Quad-Core parts & APU are now expected much later (sometimes from BTS to 2018). But I'm not sure about the timeline.
"
Makes sense imo. They need capacity for desktop and servers and zen looks like a major success. Low margin apu will have to wait then. They can always use bristol for the <80 usd segment.

 

[ecogen]

Canard at the 3.6 4.0 stepping
"
Canard PC Hardware ‏@CPCHardware
@Dresdenboy Yep. Seems Quad-Core parts & APU are now expected much later (sometimes from BTS to 2018). But I'm not sure about the timeline.
"
Makes sense imo. They need capacity for desktop and servers and zen looks like a major success. Low margin apu will have to wait then. They can always use bristol for the <80 usd segment.

Wait, what? Aren't the 4 cores gonna be harvested from the same die?

 

[Atari2600]

Wait, what? Aren't the 4 cores gonna be harvested from the same die?

Could mean a few things:
- Not physically possible.
- They are very bullish on yields.
- They expect most defects to occur in areas that are required for any core to work.

 

[krumme]

How many quads is there besides from strict segmenting purposes?.
The cpu is 2 ccx of 4 dies. I mean even if yield is bad 6 cores will be plenty to iron it out. Or do i mistake that?

 

[HurleyBird]

Could mean a few things:
- Not physically possible.
- They are very bullish on yields.
- They expect most defects to occur in areas that are required for any core to work.

I mean, quad core is harvesting 50% of the dies. Even much larger GPUs don't go that far. They still can go down to hex core.

 

[lolfail9001]

I mean, quad core is harvesting 50% of the dies. Even much larger GPUs don't go that far. They still can go down to hex core.

AMD actually harvested half of Polaris 11 that is tiny on it's own.
Also, CCX structure sort of excludes hex cores from consideration.

 

[CatMerc]

AMD actually harvested half of Polaris 11 that is tiny on it's own.
Also, CCX structure sort of excludes hex cores from consideration.

Still haven't seen any reliable source claim so.

 

[Dresdenboy]

Naa thats cheating . Its like one week back or so. What matters is 6 month plus.

Zen will hit the market with 3.6GHz base/4.0GHz boost and exceed AMD's own cautious IPC estimations.

SCNR Naah, I'm happy I don't have to correct my 15 months old estimation, which was actually based on intuitive processing (a.k.a. Deep Learning ) of all collected information back then + some considerations.

Now let's wait for the prices. Special editions etc. might be interesting as well. With their "one die to rule them all" strategy they'll have plenty dies to pick from for some higher priced SKUs.

 

[The Stilt]

1, 2, 4, 6 and 8 cores are entirely possible configs on Zeppelin.
Some of the options are just are not necessarily feasible, since there are some restrictions.

Both CCXs need to have the same amount of cores and L3 usable.

i.e A die with one fully functional CCX + L3 complex and one CCX with two defective cores will be automatically a quad core part, despite there are six functional cores on the silicon. Six cores can only be achieved with 3+3 config.

 

[itsmydamnation]

You have to remember that Intel's max power consumption have to be measured on AVX/AVX2 workloads as they have 256 bit FP units. Most common workloads like gaming or productivity rarely (if at all) use the 256 bit FP units. AMD's Zen has only 128 bit and not 256 bit FP units. Intel will have higher performance in AVX/AVX2 workloads wrt Zen but at higher power cost. If we measure power consumption in an AVX/AVX2 workloads then we are likely to see Intel get closer to rated TDP.

arg this bugs me so much (it important to get the semantics right dammit!)
Its not AVX workloads, its 256bit ops, any 256bit ops, meaning both avx and avx 2 can target 128bit vectors if they want. For example the extra instructions in avx/avx2 vs see, or because their console code is tuned to 3 operand 128bit avx etc. ( first thing that bugs me out of they way)

Also its not really the width of the units that's the limiting factor for Zen because it has more FP units then skylake. Its the load store bandwidth in and out of the core's. Zen has 256bits load and 128bits store vs 512/256 of >Haswell.

The point being AMD AVX and AVX2 performance is fine, there isn't some magically thing making AMD crap at those instruction sets( BD and PD had real 256bit instruction issues). Its that intel have an advantage on anything thats at 256bit operation. At the same time AMD/ZEN have an advantage on 128bit operations because they have more units.

If i was amd i wouldn't go chasing 256 or 512bit avx performance or SMT 4*, i would be using the massive die and power budget those things cost to increase clocks and IPC. If you look at AMD GPU's ( or NV) they are becoming much better at being CPU like. The more GPU's compute capacity becomes flexible and general the more a 512bit CPU becomes a jack of all trade master of none. If you have a master at both you can just eat them from both sides.

The General server base doesn't care about really wide vectors, thax to intels own segmentation neither does the consumer market.

*those rumors from Fottemberg aren't worth the bits in the database they are stored on unless AMD plan on basically copying a power9 style methodology to core design which is really a more unified version of CMT.

 

[Dresdenboy]

arg this bugs me so much (it important to get the semantics right dammit!)
Its not AVX workloads, its 256bit ops, any 256bit ops, meaning both avx and avx 2 can target 128bit vectors if they want. For example the extra instructions in avx/avx2 vs see, or because their console code is tuned to 3 operand 128bit avx etc. ( first thing that bugs me out of they way)

Also its not really the width of the units that's the limiting factor for Zen because it has more FP units then skylake. Its the load store bandwidth in and out of the core's. Zen has 256bits load and 128bits store vs 512/256 of >Haswell.

The point being AMD AVX and AVX2 performance is fine, there isn't some magically thing making AMD crap at those instruction sets( BD and PD had real 256bit instruction issues). Its that intel have an advantage on anything thats at 256bit operation. At the same time AMD/ZEN have an advantage on 128bit operations because they have more units.

Not to forget the missing 3rd AGU, ignoring the stack memfile. After getting a RyZen CPU I'll do some microbenchmarking for mem ops throughput. ^^

 

[Abwx]

Zen has 256bits load and 128bits store vs 512/256 of >Haswell.

.

That s for AVX2 but if there s only 128bit loads , like AVX128, then it will be one 256b load and one 128b store, because it doesnt have the ressources to exe two 128 bit loads using two different instructions, we are talking of SIMD here, so we are back to the same number as Zen.

 

[KompuKare]

Makes sense imo. They need capacity for desktop and servers and zen looks like a major success. Low margin apu will have to wait then. They can always use bristol for the <80 usd segment.

Well, sense with what resources they have.
But if AMD had more manpower I'm sure they could release more and sooner. I mean with the looses over the last decade it's amazing they have (hopefully) made such a comeback.
However, if they had spare cash I'm sure that once they had accurate projections of Zen's performance they would have hired lots of people for validation etc. Risky and might have been unpopular with shareholders (hiring before revenue), but if it would have knocked of a few months (or more for the lower end parts) it would probably be money well spent.
I think currently AMD are rather stretched for everything which was also noticeable with the Polaris launch where I'd guess the PS4 Pro work held up the launch. Would not be surprised if the Project Scorpio work is delaying Vega.
Obviously some things like tapeouts take time and simply throwing more people at problems doesn't always work, but in this case it probably would have made sense.

 

[TemjinGold]

Something I've been wondering: Will 1 Ryzen Ghz = 1 KL Ghz? Not asking to cast doubt on Ryzen, I'm just curious if anyone would know because I remembered back in the old days (Thunderbird era or so) AMD used to market that one of their mhz was better than one of Intel's. We see leaks of 3.6 base release and whatnot but is that equal to 3.6 in Intel numbers too?

 

[lolfail9001]

Something I've been wondering: Will 1 Ryzen Ghz = 1 KL Ghz? Not asking to cast doubt on Ryzen, I'm just curious if anyone would know because I remembered back in the old days (Thunderbird era or so) AMD used to market that one of their mhz was better than one of Intel's. We see leaks of 3.6 base release and whatnot but is that equal to 3.6 in Intel numbers too?

Depends on workload, from everything i have seen and heard.

 

[krumme]

Well, sense with what resources they have.
But if AMD had more manpower I'm sure they could release more and sooner. I mean with the looses over the last decade it's amazing they have (hopefully) made such a comeback.
However, if they had spare cash I'm sure that once they had accurate projections of Zen's performance they would have hired lots of people for validation etc. Risky and might have been unpopular with shareholders (hiring before revenue), but if it would have knocked of a few months (or more for the lower end parts) it would probably be money well spent.
I think currently AMD are rather stretched for everything which was also noticeable with the Polaris launch where I'd guess the PS4 Pro work held up the launch. Would not be surprised if the Project Scorpio work is delaying Vega.
Obviously some things like tapeouts take time and simply throwing more people at problems doesn't always work, but in this case it probably would have made sense.

I kind of think the same and if we look at r&d budget it have increased slightly the last 2 years as i recall. So they have started to spend.
The wsa deal with an option for mubadala and the dilution of the sharevalue to pay debt was surely a result of positive zen tapeout results. You need some leverage to pull it off. Looking at share price it seems to me amd used the situation so to speak. Its tough driving a business like that but it was needed.
I simply think they need 14nm lines more than manpower for eg verification. And if Fottemberg rumors is worth going for its probably going to be samsung us 14nm where gf is situated too.

 

[Doom2pro]

Could mean a few things:
- Not physically possible.

If they can't harvest 4 out of 8 cores, they messed up big time... No way you create a design like this and them gimp your binning prospects in the process.

 

[cytg111]

If they can't harvest 4 out of 8 cores, they messed up big time... No way you create a design like this and them gimp your binning prospects in the process.

4 cores is shooting for the apu right? on the die shots we have seen so far where is there room for the igp? I dont recall there is any.. thus it is possible theyre going with another die alltogether for 4 core parts.