New Zen microarchitecture details

[hojnikb]

I dont expect OEMs to really care about ZEN, most of the ZEN sales will be in retail for gamers.
ZEN APUs on the other hand will be the OEM dreams come true and im expecting those to be heavily OEM used in Laptops and Desktops.

Like they are now ?

 

[AtenRa]

Like they are now ?

There are ZEN APUs available today ??

 

[bakyt115]

good analyse of amds hot chips zen presentation

https://translate.google.com/translate?sl=auto&tl=en&js=y&prev=_t&hl=ru&ie=UTF-8&u=http://pc.watch.impress.co.jp/docs/column/kaigai/1037983.html&edit-text=

 

[CatMerc]

"And it is linked with the instruction fetish line from the L1 instruction cache and it..."

Considering how relatively good the translation has been before that moment, I was slightly caught off guard by that line. Was entertained.

 

[bjt2]

Ok, as "documented by Intel", let's check Intel Intel ia64 and ia32 architectures Optimization Reference Manual: http://www.intel.com/content/www/us...-ia-32-architectures-optimization-manual.html
If you don't want to follow that link and see yourself, there are two relevant screenshots:



As you can see, Port0 and Port1 support VEC FMA, VEC MUL and VEC Add. But to be even more specific, Intel even includes a small table



where they specified how many units could execute some selected instructions.
Also, InstLatX64 have a small table to compare Haswell, Broadwell and Skylake:
http://users.atw.hu/instlatx64/HSWvsBDWvsSKL.txt
I quote the relevant section:


And last, as you prefer Hardware.fr, here are their measurements of Skylake:
http://www.hardware.fr/marc/skl.txt

Yep, it's a bit tedious to check EVERY supported instruction, so let's go to the four relevant ones:



To explain it a bit:
VADDPS = vector ADD packed Single-precision
VMULPD = vector MUL packed Double-precision
YMM = ful 256bit AVX register
L: 4c = Latency: 4 clocks
T: 0.5c = Throughput 0.5 clocks per instruction, or 2 instructions per clock.

Byes

Yes, this is the same answer the other user gave me (without digging so deep and the screenshots).

 

[bjt2]

They might not be. But the chance of an entire CCX being bad may be lower than the odds of having a flawed iGPU. So AMD can just fuse off the iGPU and sell it as a quad, instead of having to take a working 8c and fuse off half the cores just to fill out a 4c sku.

Like the Athlon 845K...

 

[Dresdenboy]

AMD can't market if they're life depended on it. Zen may be the next coming of Christ but word of mouth will only carry it so far. They would need a multi million dollar worldwide marketing campaign which i believe they're incapable of pulling off.
Chip will be good but sales will be bad. Unlike their radeon graphics division which is making decent profit, Zen will tank because nobody outside enthusiast forums knows what Zen is.

They just need to deliver and hold some occasional events. Tech press, social media, OEMs, electronics stores will help with the rest. They are addressing HEDT, not the $299 sys buyer.

 

[SpaceBeer]

IT department in my former company decided to switch from i5 ultrabooks and desktop PCs to Kaveri and Carrizo based laptops and SFFs (from the same OEM) for office use. That's few thousand PCs, and most of people using them have no idea of differences between AMD, Intel, nVidia, etc. On the other hand, some of us needed more powerful workstations, so i7 or Xeon were only options.

People who make decisions know what they need for the job, and they'll buy the best thing they can for the money they got

 

[Insert_Nickname]

OEM's and large coporations don't need a massive marketing effort to know about Ryzen. I doubt AMD could even keep up with demand initially just from those two sources.

They just need to deliver and hold some occasional events. Tech press, social media, OEMs, electronics stores will help with the rest. They are addressing HEDT, not the $299 sys buyer.

Bingo. AMD doesn't need a huge marketing campaign for Ryzen initially, since they're targeting... well us... and the professional market. Those who need to know about Ryzen, already know. I my opinion they're much better of waiting until the mass market Raven Ridge APUs before running a full blown advertising campaign.

 

[guskline]

I'm only one person but I'm in for a RyZen 8c/16t cpu and 370 mb.

 

[Nothingness]

I'm only one person but I'm in for a RyZen 8c/16t cpu and 370 mb.

I might be in if:

1. performance is close to Broadwell (including some selected AVX(2) programs I use)
   
2. there's no TLB- or SMT/uop-like errata; this means I won't be an early adopter
   
3. price is really lower than similarly performing Intel.

IMHO there are too many unknowns at this point to make a decision.

 

[Dresdenboy]

good analyse of amds hot chips zen presentation

https://translate.google.com/translate?sl=auto&tl=en&js=y&prev=_t&hl=ru&ie=UTF-8&u=http://pc.watch.impress.co.jp/docs/column/kaigai/1037983.html&edit-text=

Nice! This is especially about the decoding, uop cache and further instruction processing.

As shortly answered by Mike Clark already in the Q&A at the Hot Chips press briefing, a 256b fastpath double AVX(2) 256b instruction would come as 1 instruction out of the decoder and also just populate 1 slot in the uop$. Just when being dispatched to the FPU, it would become 2 uops. This should remove some potential bottlenecks in the front end, as part of BD-CZ AVX256 processing bottlenecks stem from that creation and handling of separate uops. They also limitied the decode throughput of accompanying instructions. IIRC BD could only decode one 256b instruction per cycle.

So aside from the known throughput limits, this might be interesting to watch in reality regarding AVX256 performance, as not all softwares using AVX 256b need to sustain a 2+ op/cycle throughput over long timespans.

 

[majord]

Love Hirofumi's work, but I notice on the Zen vs Skylake comparison, Skylake erroneously shows 5 x86 decoders

 

[Asterox]

OEM's and large coporations don't need a massive marketing effort to know about Ryzen. I doubt AMD could even keep up with demand initially just from those two sources.

Red Teams, they get much earlier a CPU performance NDA/information firsthand form AMD=test motherboard+CPU.This blue is common practice, marketing is for kindergarten and average PC users or gamers.

Chinese contract or job for AMD, this is based on very early firsthand information on future Zen CPU performanse.

http://www.anandtech.com/show/10268...ms-joint-venture-for-x86-server-socs-in-china

 

[zinfamous]

"And it is linked with the instruction fetish line from the L1 instruction cache and it..."

Considering how relatively good the translation has been before that moment, I was slightly caught off guard by that line. Was entertained.

Zen is really into feet, eh?

 

[mmaenpaa]

Bingo. AMD doesn't need a huge marketing campaign for Ryzen initially, since they're targeting... well us... and the professional market. Those who need to know about Ryzen, already know. I my opinion they're much better of waiting until the mass market Raven Ridge APUs before running a full blown advertising campaign.

I am very excited for AM4 & Ryzen. Finally after all these years I can build whole line of PCs (for SMB sector & gaming) with same MB, same heat sink, same memory. Of course the components must be able to handle top Ryzen sku & powerful GPU. With socket AM2/AM2+ we were able to offer more cores (3 Amd vs. 2 Intel, 4 Amd vs. 2+ht Intel) with the same money.

For young gamers start with APU (parents pay the bill after all), ugrade to dGPU & add memory if gaming interest continues, upgrade CPU if gaming get's serious

 

[zinfamous]

Bingo. AMD doesn't need a huge marketing campaign for Ryzen initially, since they're targeting... well us... and the professional market. Those who need to know about Ryzen, already know. I my opinion they're much better of waiting until the mass market Raven Ridge APUs before running a full blown advertising campaign.

Yep. large marketing campaign for the general public that is going to be standing in Best Buy or Microcenter and comparing OEM towers with Intel vs AMD is really only going to work fi Zen has already been out in the wild for some time, with established reputation, and sales can confidently inform customers that the AMD system is legit--only if it is legit.

AMD just doesn't have that reputation and they aren't going to get there with Joe Consumer by blasting advertising before Zen is actually released. It could be an expensive disaster if Zen underwhelms, but more importantly, it just wouldn't be effective without having any wind behind those sails if such an ad campaign is going on without any product on the shelves for x amount of time.

That being said, AMD really should start putting out some single-page teaser ads in various relevant magazines like, I dunno--do dudes still read Maxim? lol, and of course some webpage ad space/youtube ads. But they shouldn't go ball-to-the-wall with that stuff until there is real product to sell. Just inject some mindworms here and there through limited, targeted advertising.

I don't think they are ever going to be at a position to compete with "Intel Inside," which is almost on the level of "Kleenex" branding and market command, so it would be a colossal waste for AMD to try and devote capital into a strategy that goes after that level of branding.

 

[iBoMbY]

Well, it looks like you should tell that to AMD. Currently it seems very much like they are going to try to launch Ryzen as OEM product first ... without any customer base.

 

[Tup3x]

Well, it looks like you should tell that to AMD. Currently it seems very much like they are going to try to launch Ryzen as OEM product first ... without any customer base.

How did you reach in that conclusion? Judging by the events they have held and the things they have said, they are targeting gamers and heavy users who build their own PCs. At least that's the impression I got.

 

[podspi]

Anyone who builds their own PCs (ok, most people) are going to be aware of Zen and what it offers (or doesn't offer).


If OEMs can offer comparable performance/efficiency from AMD, they will. I honestly believe OEMs want diversity in suppliers, even better if AMD can offer some point of differentiation that people care about (GPU perf, etc).


I saw lots of Bobcat, Jaguar products available from OEMs back in the day. If the Zen core brings the goods, I imagine we'll see lots of AMD-powered laptops. Now, whether THOSE get marketing $$$ is another matter...

 

[KTE]

IT department in my former company decided to switch from i5 ultrabooks and desktop PCs to Kaveri and Carrizo based laptops and SFFs (from the same OEM) for office use. That's few thousand PCs, and most of people using them have no idea of differences between AMD, Intel, nVidia, etc. On the other hand, some of us needed more powerful workstations, so i7 or Xeon were only options.

People who make decisions know what they need for the job, and they'll buy the best thing they can for the money they got

Is this a corporation or a small/medium business?

Enterprise (Fortune 500) rarely ever buy highend desktop CPUs but the CPU MFG that's 'winning' has the Halo Effect so it trickles down. Top 500 firms outsource the powerful compute servers, and keep notebooks, thin clients or low-end i3/i5s for local use. So it would bode well for AMD to deliver at the highend.

Most of all IMO if AMD can get into top-end MBPs, then there's a good chance mobile OEMs will spec their full new lineups on AMD, as Apple has that Halo Effect for OEMs/ODMs. Thats the market AMD needs. Mobile shipments per year are around 160mill units and Desktop 100mill.

Server/workstation right now is huge tho serving some critical breakthrough markets... Cloud is currently $90bill, Cybersecurity $80bill, Gaming $70bill and VR expected around $30bill.

OEM's and large coporations don't need a massive marketing effort to know about Ryzen. I doubt AMD could even keep up with demand initially just from those two sources.

Both do. Thats where the money is, not home enthusiast users

Without them, Zen won't go anywhere even if 200% better than Skylake. Don't forget what happened with K8 and LLano. But that's the Marketing Managers job to contact OEM/ODMs/Corps and make them aware of new product offerings. The sales pitch HAS to be done, as you are a business wanting awareness->uptake.

In this world, unfortunately, sales pitch and support is what sells, less of the product itself.

good analyse of amds hot chips zen presentation

https://translate.google.com/translate?sl=auto&tl=en&js=y&prev=_t&hl=ru&ie=UTF-8&u=http://pc.watch.impress.co.jp/docs/column/kaigai/1037983.html&edit-text=

Nice! This is especially about the decoding, uop cache and further instruction processing.

As shortly answered by Mike Clark already in the Q&A at the Hot Chips press briefing, a 256b fastpath double AVX(2) 256b instruction would come as 1 instruction out of the decoder and also just populate 1 slot in the uop$. Just when being dispatched to the FPU, it would become 2 uops. This should remove some potential bottlenecks in the front end, as part of BD-CZ AVX256 processing bottlenecks stem from that creation and handling of separate uops. They also limitied the decode throughput of accompanying instructions. IIRC BD could only decode one 256b instruction per cycle.

So aside from the known throughput limits, this might be interesting to watch in reality regarding AVX256 performance, as not all softwares using AVX 256b need to sustain a 2+ op/cycle throughput over long timespans.

I think the uop cache being not fully inclusive or exclusive is the key for me... Yet it's able to hold the same data as in the L1. Seems like an evolution from Intels inclusivity.

I haven't read the rest yet...

I might be in if:

1. performance is close to Broadwell (including some selected AVX(2) programs I use)
   
2. there's no TLB- or SMT/uop-like errata; this means I won't be an early adopter
   
3. price is really lower than similarly performing Intel.

IMHO there are too many unknowns at this point to make a decision.

I'm in this boat but I will buy some Summit/Bristol Ridge anyway to experiment the 'new beginning'. It really sounds fun, deciphering new tech

So far the clocks are looking awesome...

Sent from HTC 10
(Opinions are own)

 

[Doom2pro]

good analyse of amds hot chips zen presentation

https://translate.google.com/translate?sl=auto&tl=en&js=y&prev=_t&hl=ru&ie=UTF-8&u=http://pc.watch.impress.co.jp/docs/column/kaigai/1037983.html&edit-text=

AMD is also expected to load a considerable amount of cash.

In more ways than one

 

[iBoMbY]

How did you reach in that conclusion? Judging by the events they have held and the things they have said, they are targeting gamers and heavy users who build their own PCs. At least that's the impression I got.

Consumer mainboards are a long time from being ready (no hard dates, and no final specs), OEM boards are ready for a long time. They proudly showed a whole range of OEM PCs.

 

[AtenRa]

Consumer mainboards are a long time from being ready (no hard dates, and no final specs), OEM boards are ready for a long time. They proudly showed a whole range of OEM PCs.

The only AM4 OEM systems i have seen are with APUs BristolRidge not with ZEN. Also those high-end ZEN CPUs are not what OEMs are after for entry-mainstream systems, those ZEN CPUs that will launch now are only good for Gaming/Workstation systems.

Im not saying we may not see any ZEN + dGPU OEM systems but those will be low volume, most of the sales it will be from the retail/e-tail DIY market.

 

[Abwx]

Consumer mainboards are a long time from being ready (no hard dates, and no final specs), OEM boards are ready for a long time. They proudly showed a whole range of OEM PCs.

So all these people did was to stuck some random components on mock cards..?..

Specs not being published doesnt mean that there s no specs, otherwise those cards wouldnt even exist.

https://www.computerbase.de/2017-01/amd-ryzen-cpu-x370-chipsatz/