Discussion Intel current and future Lakes & Rapids thread

[JoeRambo]

I don't think AMD are increasing the numbers of cores per chiplet instead of just increasing the number of chiplets.

They are getting new IOD for sure, as current one is dragging them down with power use.
12Cx8 or 8Cx12 is impossible to know yet? We had real good leak for ZEN3 CCX size very early on, do we have anything on ZEN4 ?

I think it is in AMD's best interest to continue to rise CCX size, as their main competition is going to be not Intel, but ARM monolith monsters on same 5nm process in 2022.
For example SiPearl is 72 core monolith, supposedly with (ARM claims for V1) 50% ST performance uplift versus N1 on 7nm? Can someone fit 96C on 5nm?

 

[eek2121]

Your numbers/math are wrong. SPR is 60/56 cores, and will not come anywhere close to matching Genoa.

Note that it is rumored that up to 8 CPUs can be used in one system. I have to wonder how you can cool 8x 370w CPUs, so there is that...🤣

 

[eek2121]

They are getting new IOD for sure, as current one is dragging them down with power use.
12Cx8 or 8Cx12 is impossible to know yet? We had real good leak for ZEN3 CCX size very early on, do we have anything on ZEN4 ?

I think it is in AMD's best interest to continue to rise CCX size, as their main competition is going to be not Intel, but ARM monolith monsters on same 5nm process in 2022.
For example SiPearl is 72 core monolith, supposedly with (ARM claims for V1) 50% ST performance uplift versus N1 on 7nm? Can someone fit 96C on 5nm?

wrong thread.

 

[JoeRambo]

wrong thread.

Why so? Intel is moving to chiplets with Saphire Rapids, we are discussing the size of its chiplets, and how it compares to potential competitors chiplet based and monolithic chips. Does every post in this thread has to mention some Lake or Rapid?
8 vs 40 is different game from 8 vs 14, while 12 vs 14C is "same" . Question is what Milan will have versus Intel and ARM ( who seem to be staying monolithic with V1 gen products ).

 

[ashFTW]

The I/O takes up a greater portion on the smaller core dies.

For example ICX-SP LCC with 16 cores is at 370mm2 as well. The HCC with 28 cores is at 505mm2. The 40 core version is little over 600mm2.

You understand that each 370mm2 only has to have 1/4th the I/O.

 

[Gideon]

The lastest SR leak shows HBM2 is used, so it must be on the package, yet it's not visible on leaked pictures (AFAIK). So what gives?

Anyway, if it has HBM2 and CXL on top of DDR5 it has such an absurd memory bandwidth (and "L4" cache) advantage over Milan that the two are not really comparable.

 

[ashFTW]

Note that it is rumored that up to 8 CPUs can be used in one system. I have to wonder how you can cool 8x 370w CPUs, so there is that...🤣

SPR, at least for the Aurora, is 2P only. Cooper or it’s successor may continue the 4P and 8P for Intel.

Both Genoa and SPR for the upcoming supercomputer are Being liquid cooled. The Cray/HPE cabinets are designed to handle upto 500W CPU/GPUs.

https://www.hpe.com/us/en/compute/hpc/cray/argonne-national-laboratory.html

 

[ashFTW]

The lastest SR leak shows HBM2 is used, so it must be on the package, yet it's not visible on leaked pictures (AFAIK). So what gives?

Anyway, if it has HBM2 and CXL on top of DDR5 it has such an absurd memory bandwidth (and "L4" cache) advantage over Milan that the two are not really comparable.

I assume that the HBM2e version will come a bit later, maybe as late as Granite Rapids. It may also have Foveros based Rambo cache just like the 41 tile HPC Xe that has recently been showcased by Intel. The memory hierarchy gets much more complicated to adress the latency and bandwidth requirements with ever increasing cores..

 

[jpiniero]

SPR, at least for the Aurora, is 2P only. Cooper or it’s successor may continue the 4P and 8P for Intel.

Intel says Sapphire is up to 8S.

The lastest SR leak shows HBM2 is used, so it must be on the package, yet it's not visible on leaked pictures (AFAIK). So what gives?

It's optional.

 

[moinmoin]

Intel Sapphire Rapids to feature up to 56 cores, 350W TDP and 64GB of HBM2 memory - VideoCardz.com

Intel 4th Gen Xeon Scalable (Sapphire Rapids) More cores, more PCIe lanes, and support for next-gen interfaces. We received an anonymous tip with a slide featuring Intel’s future Xeon platform specifications. The slide is dated, as it still lists the Ice Lake server as a 2020 product, however it...

videocardz.com

There you go, 56 cores. Any other questions?

To me the most interesting info is the part by momomo_us calling the 44c 270w SR chip the "high volume" one. Remember SR is a 4 x 14c CPU tile MCM. For 44c to be a high volume part each CPU tile has 3c (or a 5th part of the tile) disabled (assuming Intel disables cores in a symmetric way like AMD does). If true yield still has a ways to go.

 

[Exist50]

SPR w/ HBM is more than just "optional". It's different silicon entirely.

 

[firewolfsm]

Meanwhile genoa is expected to offer 96 cores, 12 channels of DDR5 (which should help to offset lack of HBM) with a 10% lower TDP. It seems 2022 will only increase the server gap between Intel and AMD.

 

[Hougy]

If Golden Cove has 20% better IPC, it should be around 44% bigger

 

[IntelUser2000]

You understand that each 370mm2 only has to have 1/4th the I/O.

You understand that the 370mm2 die is 16 core Icelake-SP right?

20% increase in perf/clock usually results in much greater die area increase, for one.

By having 4 dies, yes you might decrease some I/O, but you actually have to add more because you need extra to have the dies communicate with each other.

The fact that they can maintain 370mm2 die while using a much more advanced core plus all the advancement elsewhere suggests they are actually being more area efficient than Icelake.

 

[eek2121]

To me the most interesting info is the part by momomo_us calling the 44c 270w SR chip the "high volume" one. Remember SR is a 4 x 14c CPU tile MCM. For 44c to be a high volume part each CPU tile has 3c (or a 5th part of the tile) disabled (assuming Intel disables cores in a symmetric way like AMD does). If true yield still has a ways to go.

I don't understand why you guys are being so critical of Intel in an Intel thread, but the die sizes on this thing are larger than anything AMD has put out with the exception of Navi 21...and there are multiple dies.

For SR 44c to be the volume part, that means their margins are likely around 65%. It also means that at worst, Remember that when this chip is launched. Some of the smarter geeks here might be able to figure out a rough defects per mm2 based on leaked information thus far. We will see. Something to keep in mind is that if a single chiplet has a bad core, Intel will have to disable at least 1 core per chiplet to make the configuration work. This means their yield on chiplets could be as high as 92%.

I get the scepticism regarding Intel and encourage it. I use AMD CPUs personally and there isn't a single Intel CPU here, however this is an Intel thread that is discussing future Intel chips. Please focus on discussing the topic. If you want to discuss Intel vs AMD competitiveness, feel free to create a thread if you want and those that want to contribute will contribute. The rest of us will continue to contribute to both AMD and Intel threads. Some of us love tech, no matter the vendor.

Meanwhile genoa is expected to offer 96 cores, 12 channels of DDR5 (which should help to offset lack of HBM) with a 10% lower TDP. It seems 2022 will only increase the server gap between Intel and AMD.

Sapphire Rapids is 8S. EPYC is 2S. That means a theoretical SPR system can have as much as 448 cores, 896 threads. I'm not saying you'll see something like that, but don't count on AMD core density to save them unless Genoa supports more sockets. Note that if AMD continues down the chiplet path for Genoa, they will find it challenging to expand beyond 2S...

 

[Exist50]

Sapphire Rapids is 8S. EPYC is 2S. That means a theoretical SPR system can have as much as 448 cores, 896 threads. I'm not saying you'll see something like that, but don't count on AMD core density to save them unless Genoa supports more sockets. Note that if AMD continues down the chiplet path for Genoa, they will find it challenging to expand beyond 2S...

There's very little demand for >2S systems, especially >4S. All the more so as you pack more cores and IO per socket.

 

[LightningZ71]

Our best rumors have Genoa at 96 cores per socket. I don't really see AMD pushing up from 2P for at least a couple of generations. Even then, at 2P, assuming that the rumors that have it at 96 cores per socket max, that's 192 cores and 384 threads in a 2P x86-64 server. If Intel is pushing 44 core SPR in the market, they can't even meet that with a 4P board (44 cores X 4 sockets is 176 cores/362 threads)! That leaves Intel to having only one type of system with a core count advantage, 8P systems, and they make up a very small slice of the total server market.

However, what we don't know is how much core throughput will improve with SPR. Maybe Intel will make a big leap in performance from having a better 10nm node than Ice Lake. If they just achieve the improvement with SPR over Ice Lake-S that Tiger Lake had over Ice Lake U, then they'll see a significant clock speed uplift to start. So, it stands to reason that they should be able to maintain parity, at least to some degree, between SPR 4P and Genoa 2P.

 

[moinmoin]

I don't understand why you guys are being so critical of Intel in an Intel thread

I don't follow you, I posted my impression about an information I considered noteworthy in the one place that's for discussing said information. Do you prefer me posting that in some off topic places because people like you deem it "so critical"?

It's in Intel's own best interest to have the best possible yield on its certainly no longer fresh 10nm node, any competitiveness talk aside.

If you want to discuss Intel vs AMD competitiveness, feel free to create a thread if you want and those that want to contribute will contribute.

I disagree, every discussion about every product is always in the context of competitiveness in some way, you can't avoid that altogether (just look at any AMD thread back when it wasn't, or all the more exotic threads about RISC-V, Elbrus, Zhaoxin etc. pp.).

Anyway I didn't even mention AMD at all and I don't think my particular topic requires doing so, but I see people jumping on that tangent now so I'll see myself out again.

 

[Asterox]

Short summary, Steve can act or pretend that AM4 Renoir APU 4000 series do not exist.In reality, if you want buy Renoir APU you can buy it.Here in my country(EU region), Renoir APU is available for example R5 Pro 4650G.As far i see, R5 Pro 4650G is now also available at german retailer Mindfactory.

Renoir APU Vega 7 or 8=iGPU gaming performance as GT 1030+ddr5 memory

Procesori | AMD Ryzen procesori | ADM računala

ADM Računala Web Trgovina, prilaz V. Brajkovića 4, 10000 Zagreb - prodaja vrhunskih računala i opreme po najboljim cijenama!

www.adm.hr

AMD Ryzen 5 4650G 4,3GHz So.AM4 11MB Cache (Tray) - Sockel AM4 | Mindfactory.de

AMD Desktop von AMD | AMD Ryzen 5 4650G 4,3GHz So.AM4 11MB Cache (Tray) :: Lagernd :: über 30 verkauft :: 24 Jahre Kompetenz | Hier bestellen

www.mindfactory.de

Old 12/14nm AMD 3000 APU series, general power efficiency=CPU performance is very bad compared to 7nm Renoir APU.

Page 26 - Zen 2 APUs/"Renoir" discussion thread

Page 26 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

In today GPU market, or "where solid iGPU is important" R5 4650G has a pretty good price.

Page 10 - Discussion - AMD Cezanne/Zen 3 APU Speculation and Discussion

Page 10 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[Panino Manino]

It was strange to see this comparison without a AMD APU.
The new Xe GPU is good but these integrated graphics are too small. If he had included any AMD APU the question would change from "can you game on it" to "which games can you play on it".

 

[Shivansps]

Considering that AMD choose to neglect desktop APUs, im completely fine with what Steve did.

 

[Ajay]

Considering that AMD choose to neglect desktop APUs, im completely fine with what Steve did.

AMD is supply limited, they have to make some hard choices right now.

 

[LightningZ71]

Both Intel and AMD have the ability to make processors with substantial iGPUs in them and have them perform well. With the big crunch on retail GPUs and the dearth of good, low end options out there, and with CPUs nearing the DDR5 era, we are essentially a year or so away from having roughly the memory bandwidth of the 1050/rx550-560 available to CPUs, and lithography nodes small enough to make large on-processor caches doable at a reasonable price. There's no reason that both manufacturers wouldn't offer a processor with the ability to reasonably game at 1080p at medium to high quality by then.

 

[jpiniero]

Both Intel and AMD have the ability to make processors with substantial iGPUs in them and have them perform well. With the big crunch on retail GPUs and the dearth of good, low end options out there, and with CPUs nearing the DDR5 era, we are essentially a year or so away from having roughly the memory bandwidth of the 1050/rx550-560 available to CPUs, and lithography nodes small enough to make large on-processor caches doable at a reasonable price. There's no reason that both manufacturers wouldn't offer a processor with the ability to reasonably game at 1080p at medium to high quality by then.

And discourage a dGPU sale?

AMD appears to be going the other way; putting a basic IGP on the IO Die and calling it a day. Maybe at some point there will be an actual IGP chiplet.

 

[IntelUser2000]

Both Intel and AMD have the ability to make processors with substantial iGPUs in them and have them perform well.

Last time both manufacturers put a large GPU, it didn't work out for them. Look how the 5775C was and abandoning it with Skylake. How about AMD in the Llano days?

Yea sure they can put a large GPU, but would we pay extra for it? We want all the extra performance for practically free right? Same price as the current anemic iGPUs? But they won't do that.

The alternative is pairing the highest end iGPU on a Core i3/Ryzen 3, so you save on the "useless" extra CPU and is more balanced. I just don't see this happening either. Do you want to be stuck to the iGPU performance, or eventually upgrade to a much faster dGPU you've been waiting for?