Details of AMD Zen 16-core x86 APU emerge

[ZaO]

Phenom can bridge from some AM2 boards (you are supposed to use AM2+) to AM3+, the FX range require an AM3+ board to work. Still, the AMD socket flexibility has proved very useful to me over the years.

I thought that a while back I'd read a forum post project about someone running an FX cpu on an AM2 board. I may be mistaken though. I'm not sure if the AM2 socket has enough pins for an FX chip actually. I might look that up later.. But I do know that my friend is running his Phenom II on an AM2 board (not AM2+). He's had that board for so long now..

I remember running my 955 on my AM2+ board. But even after a BIOS update, it still didn't run at full potential. I saw quite a difference when I upgraded to an AM3+ board. Still, it was awesome that I could buy a cpu and just run it straight away, then get a motherboard at a later date when I'd saved a bit more money Stuff like that is real cool for people on a budget. These days however, getting a new cpu nearly always means upgrading the mobo for me if the existing board is gonna hold it back.

[DanceswithUnix]

This looks like the kind of thing that's going to get targeted at Microservers, probably for dense HPC work rather than general servers. There isn't much demand to use the graphics parts for acceleration in general workloads. Maybe we're looking at 16 small cores along the lines of those in Xbox One, I'm fairly sure this isn't some super high end play, with 18 core HT Xeons out there a expensive 16 big core chip would be struggling straight away.

This is bound to be targeting 14nm, so the figures aren't actually that outrageous.

I have argued in the past that AMD should go onto a single socket, FM3 would make sense. If you double the resources of the 32nm FX8370 then that would get you 16 cores and a shedload of cache. Nice, and a doubling should take a process shrink which from 32nm gets you down to 22nm. So another shrink down to 16nm/14nm gets you some more transistors to play with (not quite another doubling). You can either add a big graphics subsystem, or my favourite would be to add a small graphics subsystem to make it APU pin compatible and shrink the die size a bit to make it cheaper to make. A consumer APU would be better served with 4 cores/8 threads for the near future and a bigger graphics part.

[azrael-]

I think I'm missing something here - when did a server or "compute engine" need fancy graphics.

Since people started using GPUs for HPC, so about a decade? Huge amounts of parallel compute has been in high demand in a variety of fields for a long time - this APU is aimed at that market.

Even worse, we've been saying for months/years that a high-end desktop competitor is needed and this seems to be ideal - so why the heck isn't it being "desktop-ized"?

It will be - Zen is an entirely new x64 compatible microarch that is likely to be propagated through AMDs entire product stack. This particular implementation is aimed at HPC, and was leaked from an HPC developer event from a few months back (check the Piledriver thread in hardware and there's a bit more discussion about it).

Wasn't this (Zen, that is) what AMD poached Jim Keller back from Apple for? The main architect behind AMD's runaway success, the K8.

[scaryjim]

I suspect it was more to get someone with experience of both ARM and x86 for the ambidextrous server/embedded approach they've taken. He'll certainly have input into Zen, though given he rejoined AMD in 2012, I don't think he'll have been there long enough to have been responsible for the initial Zen design: more likely he's finessed and enhanced whatever the x86 team were working on when he joined.

[crossy]

Indeed, if you want to be able to test a range of recent AMD motherboards a £14 Sempron 145 is your friend - AFAIK it'll run in any socket from AM2 to AM3+ (I've booted an old MSI K9N4 SLI AM2 board with mine). Still purchasable new, as well. That's platform longevity

Quite true - I am still running a 2010 vintage PhenomII 1090T and not really missing the performance gains (?) from having something newer - still able to play CoD, BF4, Bioshock at "acceptable" frame rates and everything else runs good enough. Actually my old 1090T runs rings around a laptop equipped with a decent i5 for video transcoding - which did unfortunately screw up my "justification" for getting a new rig. Oh, my original board was a decent Asus one from the 2010 range, but last year - on Cat's recommendation - I upgraded to a Sabretooth FX and it's made a big difference. And there's always a relatively cheap FX8370 if I need a boost.

This story: http://hexus.net/tech/news/cooling/8...r-hits-market/ reminded me of another big plus for AMD - they've kept the same cooler mounting holes since AM2 as well (and the same stock bracket since 754/939), including on the FM sockets. That's a huge amount of longevity for your good quality CPU cooler, which lets be honest is a piece of kit that shouldn't need changing unless oyu manage to completely wreck it....

Socket longevity yes - but I'm not exactly convinced by the design. I've had to replace that plastic ring too many times because a heavyweight cooler has sheared off the lugs after a while. One of the big pluses of the Noctua NH-DH14 I've been using for a long while is that Noctua replace that with their own mounting which is more secure and stronger.

[DanceswithUnix]

I suspect it was more to get someone with experience of both ARM and x86 for the ambidextrous server/embedded approach they've taken. He'll certainly have input into Zen, though given he rejoined AMD in 2012, I don't think he'll have been there long enough to have been responsible for the initial Zen design: more likely he's finessed and enhanced whatever the x86 team were working on when he joined.

I think these days finesse and overall structure are what it comes down to. Detail stuff like branch predictors are going to be taken from the best that the existing cores have to offer, otherwise time to market is just too long and the existing components are pretty good.

[GuidoLS]

I wouldn't argue massively with that. But, being roughly a gen or two behind with tech, or at least in regards to hardware performance, isn't a big deal. There are plenty of people who keep their hardware for years (sometimes, not much changes in a couple years either). One of my friends who games on pc, playing all the latest, demanding games stills runs a Phenom II 965 (at stock clocks might I add). Gaming is really my strongest area of knowledge when it comes to hardware, so that's the main thing I can comment on. Amd have been competative in the mid range for years. It's just the high end where they're not doing so good. I'm more of a mid-high end sort of guy. So, something like a Phenom II would definitely not cut it for me (I already got rid of my 955 back at the start of 2012 to get a 2500K). An FX, yeh, sort of.. But I feel I would be better buying Intel at this point in time. Though if I already had a good, built computer with an AM3+ board, I'd almost definitely pickup an FX for a cheap upgrade! See, that's another place where Amd are great, and I don't hear enough praise for - they keep their chipsets going for a long time with really decent upgrade paths! I mean, you can run FX chips on some old AM2 boards as far as I'm aware (though I'm sure they'll be held back a bit on that chipset). Just think how long ago those came out.. Amd are more consumer friendly, imo. They just don't always have the latest and greatest top end stuff when it comes to cpu's. So, for mid range and budget gamers - Amd is awesome As for Gpu's - Amd are still doing great at the high end!

I'm not just a generation behind - I lost count, but I think it's at least 6 by this point (core2quad 9550). partnered with a 9800GT, which I think is 7 or 8 behind. And, interestingly (or not) the only DX9/10 game I haven't been able to run well has been the new Wolfenstein (and Crysis 3, but nobody ran that well that didn't have a 6 figure salary, or more money than sense). No DX11 games, but so far, I've only missed one I wanted to play, and am probably going to end up having to spill the $$ for a card before Witcher 3 comes out... so I know just what you mean when it comes to mid-level gamer. I'm not oblivious to the fact that the reason I'm still good on 90% of what's coming out is because my kit, while a bit long in the tooth, were top of the line (or close, with the Intel chip) when new.

And in the home PC market, companies can more or less get away with catering to people that don't mind mid-range. I think it's a huge mistake to try to compete, head to head, with Intel in the high end server market with an announced chip that doesn't even carry 1/4 of the processing power. And with a SoC to boot... makes no sense to me.

[DanceswithUnix]

And in the home PC market, companies can more or less get away with catering to people that don't mind mid-range. I think it's a huge mistake to try to compete, head to head, with Intel in the high end server market with an announced chip that doesn't even carry 1/4 of the processing power. And with a SoC to boot... makes no sense to me.

Remember AMD have SeaMicro as one of their available technologies. An SoC, specially with on package memory, would be ideal for dense server farms.

[ZaO]

I'm not just a generation behind - I lost count, but I think it's at least 6 by this point (core2quad 9550). partnered with a 9800GT, which I think is 7 or 8 behind. And, interestingly (or not) the only DX9/10 game I haven't been able to run well has been the new Wolfenstein (and Crysis 3, but nobody ran that well that didn't have a 6 figure salary, or more money than sense). No DX11 games, but so far, I've only missed one I wanted to play, and am probably going to end up having to spill the $$ for a card before Witcher 3 comes out... so I know just what you mean when it comes to mid-level gamer. I'm not oblivious to the fact that the reason I'm still good on 90% of what's coming out is because my kit, while a bit long in the tooth, were top of the line (or close, with the Intel chip) when new.

And in the home PC market, companies can more or less get away with catering to people that don't mind mid-range. I think it's a huge mistake to try to compete, head to head, with Intel in the high end server market with an announced chip that doesn't even carry 1/4 of the processing power. And with a SoC to boot... makes no sense to me.

I think you'll be happy when you get a DX11 card. I remember going from DX9 to DX10 was no big deal at all.. Infact, I usually ended up running games in DX9 mode anyway, because the difference in visuals was so hard to notice, and DX9 games ran so much faster. DX10 was crap to me. But when I went to DX11 - Huge difference! Tesselation is one of the most notable ones. Lighting as well. It seems way better optimised than DX10 too. Maybe just make the jump straight to DX12 when you get a new card. It's hopefully going to be much more efficient. But, new ones are always "supposed" to be way better eh..

[GuidoLS]

Remember AMD have SeaMicro as one of their available technologies. An SoC, specially with on package memory, would be ideal for dense server farms.

Not any more...

As part of its 1Q 2015 earnings release, AMD has announced that it is leaving the high density microserver market, effective immediately.

AMD bought SeaMicro in 2012 for $334 million to get a foothold into the microserver business. At the time, SeaMicro built systems containing dozens of Intel Atom and Xeon processors connected to a shared storage and network fabric.

Since the acquisition, AMD has only released a single new SeaMicro system, the SM15000. This could use either AMD Opteron systems (using the Piledriver core) or Intel Xeons (using the Ivy Bridge core). With today's announcement, it's clear that system will also be the last new SeaMicro system to be released.

Source and more

[DanceswithUnix]

Not any more...



Source and more

Ooh I missed that one, must have hurt to have to sell that. Their ARM and x86 kit still seems very much tuned to dense microservers though. Perhaps they think the likes of HP will sell enough.

[DanceswithUnix]

Fudzilla now claims there will be a non APU Opteron version with 32 cores.

http://www.fudzilla.com/news/process...-zen-x86-cores

Shall look forward to trying one if true

[Kanoe]

The graphic component of the APU is also impressive in its scalability and use of new tech. 'Greenland' stream processors are on board and capable of addressing up to 16GB of HBM (high bandwidth memory). This GPU offers half rate double precision compute capability enhanced ECC and RAS and HSA support.

And the CPU will be able to access the HBM memory to due to HSA which makes things very interesting.

The GMI tech will be interesting too as it will have to tie all the memory together so that CPU and GPU can both access the HBM and DDR4.

This chip will defo happen, it will be big, it will be expensive and it is not for consumers.

[DanceswithUnix]

This chip will defo happen, it will be big, it will be expensive and it is not for consumers.

On 14nm it doesn't sound that huge.

Consider if you re-did the FX8370 on 14nm, how many cores could you fit on it the same size silicon?

The non APU version seems to not come with HBM as well, but then with 64MB of L3 it probably doesn't need it.

[Xlucine]

Fudzilla now claims there will be a non APU Opteron version with 32 cores.

http://www.fudzilla.com/news/process...-zen-x86-cores

Shall look forward to trying one if true

Just like the 16 Zen core high performance market APU, each core has 512KB of L2 cache and four processors share 8MB L3 cache. The highest end part will come with eight clusters of 4 cores and if you do the math this server oriented CPU will come with 64GB of L2 cache and 16MB of L2 cache for its CPU cores.

Wow, what a chip!