AMD announce EPYC

[Journey]

Epyc, just launched for Datacentre, this is Naples.

Link to the AMD EPYC product page - http://www.amd.com/en/products/epyc

Research paper showing hte benefits of EPYC single socket soluiton. - http://www.amd.com/system/files/2017-05/TIRIAS-AMD-Single-Socket-Server.pdf

EPYC Features

• A highly scalable, 32-core System-on-a-chip (SoC) design, with support for two high-performance threads per core
• Industry-leading memory bandwidth, with 8 channels of memory per EPYC device3. In a dual-socket server, support for up to 32 DIMMS of DDR4 on 16 memory channels, delivering up to 4 terabytes of total memory capacity
• Complete SoC with fully integrated, high-speed I/O supporting 128 lanes of PCIe® 3, negating the need for a separate chip-set
• Highly-optimized cache structure for high-performance, energy-efficient computing
• Infinity Fabric coherent interconnect for two EPYC CPUs in a dual-socket system
• Dedicated security hardware

“Today’s single-socket server offerings push buyers toward purchasing a more expensive two-socket server just to get the memory bandwidth and I/O they need to support the compute performance of the cores,” said Matthew Eastwood, senior vice president, IDC. “There are no fully-featured, high-performance server processors available today in a single-socket configuration. EPYC changes that dynamic by offering a single-processor solution that delivers the right-sized number of high-performance cores, memory, and I/O for today’s workloads.”

 

[Journey]

AMD are on course to cause a serious shift in the enterprise space with this, it looks immense and that 8-channel memory, blooming heck. Welcome back AMD!

 

[Journey]

Indeed, very quick to market from announcement.

 

[Journey]

Looks like we've got 64c/128t Epyc CPU's on the way 1H '19 - sounds almost silly, 64 cores on a single CPU.

 

[Journey]

Looks like Patrick has issued a correction on servethehome article - 48 core next, 64 core will be later

Ah, indeed, Starship first, then Rome, so 1H '20 for 64 Core, damn, will have to make do with 48 cores for about a year.

 

[Journey]

Forgot to link this from about a month ago, but here's a supposed leak of a 7nm 64-core EPYC CPU.

https://hothardware.com/news/epyc-7nm-rome-benchmarks

Seems to look almost legit, and given the launch is due in the next month or two then I can't see why these results wouldn't be real.

 

[Journey]

I'd say with the new node, and the IPC increase I don't see that data as being inaccurate, 64c/128t it's a heck of a lot of computational power in a very small space.

 

[Journey]

Also if history is anything to go by looking at the gpu space, I think intels gpu's will be utter pants, intel do great hardware (subjectively of course given how insecure their parts currently are) but have never been able to do software so imo they will probably have great hardware but be let down by drivers.

It's a different world now though, for GPU's especially with compute/AI/machine learning taking advantage of the GPU's. I think they have realised they actually need to make some serious inroads before Nvidia/AMD leave them behind selling out of date processors, when the rest of the world is using more and more GPU power. I'd be very surprised if they didn't invest heavily in their software team this time around, and collaborate closely with developers, and companies like Google/Amazon/IBM yes they are competition but if a product from another manufacturer can make you better then you use it. I bet AMD have some Xeon servers, or at least did somewhere.

 

[Journey]

Even more interesting and not at all related to EPYC is the newly announced Intel Cascade Lake 48 core CPU, sans SMT of course. It's a multi die configuration like EPYC since they are stuck on 14nm+++++++++ and should be out in 2019, probably about 300w TDP at 1.8GHz.

 

[Journey]

Meanwhile back in the EPYC CPU thread...

You can't argue that seeing what they produce with their backs against the wall will be really interesting. At first they were all like "They are just gluing cores together, dont accept that", now a year down the line it's like "Oh **** we better learn the recipe for that magical IF glue".

The most interesting bit for me will be memory config and how their solution scales, anything less than near perfect ryzen like scaling and performance will be seen as missing the mark. The benchmark for an mcm x86 design has been set and now it's Intel's turn to show some engineering prowess. It is time for the big boys to lay it on the table and in my opinion all we need to know is how short they fall. Sadly I just don't think their first iteration will fall in the same ballpark for all metrics but being Intel you really would be a fool to completely write them off.

I don't think it will be anywhere near as efficient in PPW compared to Zen2 cored EPYC on 7nm, and that is the killer especially where TCO is concerned. Also two large 24 core monolithic dies in one package, is that going to need a new socket, more than likely. I'd never write them off, they must have something in the bottom drawer for a rainy day.

 

[Journey]

It's basically Athlon 64 round 2 isn't it? Genuinely interesting stuff being put on the table, those early signs of that true "next gen" slapped in front of you for the first time. Epyc 2 is shaping up to be that r300 (gpu reference ), the thing that absolutely crushes everything that came before and makes the industry stand up and take note. The thing is we all know how this goes, history says AMD innovate, intel take it on the chin perfect a similar design and run away with it. Is it really any different this time?

Intel are in a slightly different predicament due to the process node issues, that might delay the steamroller for about a year, so 2022/23.

 

[Journey]

Already mentioned 2 pages back.

 

[Journey]

Well given what has just been shown Intel better bring the goods with 10nm server CPU's or they are going to be bleeding market share by the end of 2019.

 

[Journey]

Threadripper has one major disadvantage - its very high power consumption - compare the 65W 2700 with the 180W+ 12-core.

The Ryzen 7 2700 sticking to the 65w TDP only boost to about 3.4GHz on all cores, the 2920X is more like a 2700X with 4 extra cores, and a slightly lower clock speed on all cores at once. 180w/12 = 15w per core 105W/8 = 13.125w per core, and the 65w/8 = 8.125w per core, you can have high clocks and lower power all at the same time.

 

[Journey]

I've taken the liberty to do a comparison shot of the dies, and put an overlay of the Rome die (2x8 cores) over the Naples Dies (1x 8core + I/O)



Pretty impressive.

 

[Journey]

It's not that hard to figure out is it?

Retail pricing of the R5 1600/2600 is USD $140-160 it's a six core CPU, using an 8 core die. Moving to a 16-core CPU isn't that big of a push at the highest end of the product stack for the enthusiast space. £499 again, like the 1800X was on release but this time for a 16c/32t CPU with 20+4 PCI-E lanes, and dual channel RAM, with a performance of that to outdo the competition. Again don't forget they sold the TR4 1900X which was only an 8-core CPU on the HEDT platform, giving access to those features that AM4 doesn't have. They could happily start the HEDT platform at 16c then go to 24/32/40/48 etc, allowing scope to make it impossible to chose Intel be it for number of cores, speed, or price.

I'd be highly surprised not to see a 6/8/12/16 in the standard desktop space, with a view to abandoning 4 cores to all but the very entry level devices.

 

[Journey]

Yep. While i was at work today i was thinking a lot about "Horizon" last night. To be honest, if Ryzen 2 is as good as AMD claims, they would be stupid even to sell 6 core it makes much more sense in my mind to just start with 8 in the first place.

I think 6 cores is a great starting point, and will allow them to use partially faulty dies (potentially), and if you consider they could be bringing them in at $99 and hopefully with a nice APU attached it would greatly disrupt the OEM market, and the low end desktop, more so than the 2200G has so far. It would be competing against Intel's i7 8700 or the i5 9600 but at a huge cost advantage and overall be better in power, speed, and security, a quadruple whammy.

 

[Journey]

I don't disagree with your latter thinking there Journey, but i also think that AMD won't actually have enough partially faulty dies to make the process worthwhile. In other words, why cherry pick such small dies, makes more sense to me to dump them and pick the best.................after all, they will never have to pay for them in the first place if they don't come up to spec.

Indeed, but even imperfect products can be sold, and TSMC/AMD may have an agreement in place while the 7nm costs are high that allows both parties to benefit. TSMC get paid a little towards the losses, and AMD get ultra cheap product to use at the lower end of the market place.

 

[Journey]

Yep, ok, we will wait and see lol.

Yep, it's pretty much all guess work at the moment, but we've both been around long enough to know you can always be surprised one way or the other.

 

[Journey]

You'll probably find that the £400+ SKU's would become an Ryzen 9, rather than Ryzen 7.