REGISTER | LOGIN
Breaking News
Blog

AMD’s Zen Takes on Intel

Competition returns to x86 in 2017
NO RATINGS
Page 1 / 2 Next >
View Comments: Newest First | Oldest First | Threaded View
GSMD
User Rank
Author
Re: ...
GSMD   8/22/2016 5:44:11 AM
NO RATINGS
1. On the contrary, it is easier to commoditize the server. Product cycles are larger, variations are lesser and IP blocks used are smaller.

2. x86 is not an exemplar for server processor design cycle timelines since it has very specific verification issues related to backward compatibility and teh complexities of the ISA. Thsi makes platform issue more complex too.

3. Simpler ISA like RISC-V do result in effort reductions in core verification. Where it gets more complex is the memory model and cache coherency brought about by a waeker memory model. Can make it way more complex. Add TM and yiu have the perfect verification storm ! But the right High level language and formal verification can reduce this effeort.

4. Use of modern HDLs/HGLs make processor designs a lot simpler. RISC-V designs are based on Chisel or Bluespec predominantly, which do make life a lot easier. Can result in a 4-6 fold reduction in effort in some areas. Also the more formal semstics of these languages make verification a lot more tractable. For some reason most extant processor design still uses older languages (Intel does have higher level languages in verification). I can only atribute it to sheer cussedness on the part of  the design community ! Even in my own company I could not get my engineers to shift languages.

5. The platform issue is way overblown. Open source plus standardization makes platform development and verification a lot more tractable. The focus should be on SKUs. It is not that the platform is not important but that the Linux/GCC/LLVM community is more than capable sorting out that issue if early silicon/FPGAs are released. Of course if the cores are open source, the platform effort becomes  simpler. If my efforts at sharing platform components between RISC-V and OpenPower succeed, then platform development can be shared across multiple processor families.. I am aiming at sharing socket/pinout definitions. Let us see how that works out. Cache coherent interconnects, I/O channels etc should be orthogonal to core families and ideally should be standardized independently. In other words platforms and cores/ISA should be largely decoupled.

6. Backend is still an art form but P&R is mostly automated these days and while special libraries are used they are primarily in memory and fabric areas.Till fabs reveal the front end of their libraries (the full library need not be disclosed) , physically aware synthesis cannot be done in the open. Hopefully fabs will see the light of the day. If that happens, then open sources cores can be synthesised by anybody and optimal digital designs can be achieved.

Gondalf
User Rank
Author
Re: ...
Gondalf   8/21/2016 5:19:12 AM
NO RATINGS
All it's possible at this point :). AMD at Intel FABs?? why not??

Still i think Server space will remain safe from this commoditization we are looking in phone space. Yes there is a lot of new IP around but this doesn't mean it will be cheap neither that the "core" is the most important thing; in server space is all leading edge, optimized at an high level. The plataform and the drivers that are the real discrimine of a server solution.

Look at recent AMD presentation.....they are at an FPGA level in some parts of the plataform, trying to debug it. In my knowledge the plataform is the "thing" that gives the larger numer of delays to a new product. After all, once debugged, a cpu can ship a little slower than expected with some clock speed enhancements few month later.

Look at Intel, announcing a new Server product, it NEVER emphasize the SKU but the plataform instead and its features and new drivers. Absurd thing some ARM server producer are all on the SKU and they say nothing or so on the plataform......at the end of the day the most difficult thing to realize in an reliable manner good enough to impress IT managers.

 

GSMD
User Rank
Author
Re: ...
GSMD   8/21/2016 3:48:10 AM
NO RATINGS
Intel should offer to fab Zen's in its foundries ! They are going to lose business to AMD anyway, might as well pick up some foundry business. They are doing this with ARM . Seems to be a win win situation ! But seriously with open source cores gaining traction, margins in the IP and processor business may drop. Foundry margins may be low but cutting edge nodes have almost insurmountable entry barriers.

Gondalf
User Rank
Author
Re: ...
Gondalf   8/20/2016 6:07:45 PM
NO RATINGS
Unfortunately it is not :). Zen looks good and it will give good money to AMD helping Intel to beat ARM armada, but....

Pressed by the press AMD said: "comparable TDPs to Broadwell-E"

https://www.pcper.com/reviews/Processors/AMD-Zen-Architecture-and-Performance-Preview

So entusiast FX SKU will be around 140W like Broadwell E.

All this respect a simple law: at same cpu design expertise and at the same process node, equal performance means equal average power consumption.....expecially if the ISA is the same. Low hanging fruit were already taken in cpu science, new good ideas give only little incremental benefits.

Funny enough the quality of a manufacturing node is becoming even more relevant lately.

 

 

realjjj
User Rank
CEO
Re: ...
realjjj   8/19/2016 6:50:32 AM
NO RATINGS
There are some hints that make 95W almost certain. I am assuming that they have only dual channel memory and they do have less L3 cache than Broadwell-E. AMD stated that some SKUs will ship at higher clocks than 3Ghz and 3GHz is not a lot lower than what Brodawell-E does when fully loaded.  Assuming perf is competitive with Broadwell-E , the power metric seems interesting. Too little info to be certain of anything ofc but it is encouraging.

The MCM theory is just a theory as far as i can tell, it doesn't make a lot of sense to go that route especially when the 8 cores die shouldn't be big. Could be 2x16 but remains to be seen how area efficient AMD is. Could they do 8 cores in 120-150mm2 ? Broadwell-E is about 246.2mm2 and Skylake GT2 4C half that with the GPU being maybe some 40% of the area. Intel is using taller cells in some areas while AMD might have focused on density. Obviously Intel has the tighter process but Zen 8C could be surprisingly small. AMD is supposed to have modular designs and be able to put something together without the high costs. If they have that, why go MCM? The module should actually be 4 cores with 8MB L3 and AMD would just easily scale up the module count. They got to at least have a 16 cores die. If AMD went with MCM remains to be seen what kind of packaging they use, they could use an interposer to mitigate the downsides. The downsides are not something they wouldn't be aware of, unfair to assume that they don't do anything about it. Nvidia went over 600mm2 on 16FF already and in server AMD could afford to do that too next year, if needed. I'm sure they did the math and decided on what they thought is the best solution. Expecting a 8 cores die and nothing more seems overly pessimistic.

As for stopping ARM, why would anyone want that? Competition is good for the entire world, except Intel and Intel is just a corporation, no point in loving or hating it.

If AMD wasn't deceiving with this benchmark and they really got the goods, it is very encouraging. Hopefully we find out a bit more next week.

Gondalf
User Rank
Author
Re: ...
Gondalf   8/19/2016 5:17:54 AM
NO RATINGS
Obviously a competitive 8 core Zen does not fit in 95 W. This is feasible only if low clocked, so unlikely. IMO they will stay on actual 125W TDP plataform figure.

About the 32 cores monster, it is a monster for real?? After all they are only four 8 core modules on an MCM package, this will not help the performance and latency in general. The communication between dies waste a lot of power and latency unfortunately.

About bandwidth, Purley will have six memory channels and will be single die, with a great advantage on performance and power consumption. 

Anyway we have, more or less, the second source that will stop ARM to fight to enter in server space.

 

realjjj
User Rank
CEO
Re: ...
realjjj   8/18/2016 4:22:13 PM
NO RATINGS
Forgot something ...

AMD should consider selling a 8C/16T Zen in notebook. Preferable at 45W but even 65W could work. At that kind of perf a discrete GPU would be used anyway and there is absolutely no reason to be stuck on 4 cores in notebook just because Intel had no competition for a few years. Might not be very high volumes but it's high ASP and a hole in Intel's armor.

 

realjjj
User Rank
CEO
...
realjjj   8/18/2016 12:44:09 PM
NO RATINGS
L3$ might be 8MB per quad cluster so 8+8 in total.

For Zen desktop if they have only two memory channels it would be interesting. In consumer there would be a minor perf loss but substantial savings in area, power and system BOM.

For the 32 cores monster, again the number of memory channels could make a difference. Intel goes up to 24 cores but with only 4 memory channels max bandwidth is only 102GB/s.

Power could be interesting too. Zen 8 cores is rumored at 95W while Broadwell-E is rated at 140W. Shipping clocks are unclear but this is very encouraging, assuming the perf in Blender is not a one hit wonder.

 

@Rick

PCIe 4.0 might be more useful for XPoint SSDs than GPUs next year.  Costs also matter for first gen Zen since Zen needs to compete with the 150-300$ Intel offerings where Intel has the crippling "advantage" of having an integrated GPU that enthusiasts don't need and only 4 cores.  If they really got the perf and maybe even win in power, this gets really interesting.

rick merritt
User Rank
Author
Welcome back, AMD!
rick merritt   8/18/2016 11:51:38 AM
NO RATINGS
It's already been quite a year for the company so far with Polaris and two China JVs announced.

One nit is AMD's Zen chip sets will iniitally support PCIe Gen 3 and I'd be surprised if we didnt see some PCIe Gen4 graphics cards emerge next year.

Like Us on Facebook
EE Times on Twitter
EE Times Twitter Feed