AMD EPYC Server Processor Thread - EPYC 7000 series specs and performance leaked

[scannall]

No sys admin ever got fired for buying Intel. AMD will have a lot to prove before anyone takes them seriously in the server space.

They used to say the same thing about buying IBM PC's. Until Compaq....

 

[frowertr]

They used to say the same thing about buying IBM PC's. Until Compaq....

I'm not saying that AMD can't find room. All I'm saying is that their 'burden of proof' will be extra high.

 

[scannall]

Intel spent 3.4 billion on R&D in Q1. AMD had Revenue's of like 900 million. Lets not pretend that in one product launch AMD is ready to ship an unlimited amount of CPU's for EPYC to outpace Intel. The fact is that AMD won't be a Rottweiler nipping at Intels ankle. It will be more like an Ant crawling up its shoe. It's not about sticking a finger in it's ear. It's about not throwing away Billions now because you might have to fight AMD later. It will be much easier for Intel to fight on the consumer products and force AMD to lower prices their increasing demand and forcing AMD to divert stock of dies for those products to keep them in stock.

A lot of that R&D is on process however. And other areas that Intel sells in, like memory, cell radios and network connectivity to name a few. They've been refining and fiddling with the same CPU architecture for a long time now. And why wouldn't they. Huge lead, and brand new architectures are a giant multi year and multi billion dollar crap shoot. With more duds than winners. Nobody on these forums, or at Intel thought AMD would ever get close enough to be in the ring. But they did. With a solid product that will sell well. I'm not sure if AMD has the capacity yet to pump out the volume to make Intel really take notice yet though. So I don't see Intel really slashing server prices yet, by a significant amount. Next year this time? Unknown.

One thing is known though. Intel needed a kick in the pants. Maybe that will shake up their internal structural problems. And AMD should finally become profitable again, and keep applying the boot. We all win, so yay for us.

 

[Topweasel]

A lot of that R&D is on process however. And other areas that Intel sells in, like memory, cell radios and network connectivity to name a few. They've been refining and fiddling with the same CPU architecture for a long time now. And why wouldn't they. Huge lead, and brand new architectures are a giant multi year and multi billion dollar crap shoot. With more duds than winners. Nobody on these forums, or at Intel thought AMD would ever get close enough to be in the ring. But they did. With a solid product that will sell well. I'm not sure if AMD has the capacity yet to pump out the volume to make Intel really take notice yet though. So I don't see Intel really slashing server prices yet, by a significant amount. Next year this time? Unknown.

One thing is known though. Intel needed a kick in the pants. Maybe that will shake up their internal structural problems. And AMD should finally become profitable again, and keep applying the boot. We all win, so yay for us.

That was my point. It wasn't Intel is better than AMD because they are spending nearly 15 Billion a year on development. My point was that Intel sales are strong enough that they spend 15 Billion a year in R&D. Intel isn't going lose Billions in revenue in competing in price on their super profitable server selection. Even if that means that AMD sucks up a billion in sales (1 Billion in Servers alone would be fantastic).

Even when AMD had a good lead on Intel they were never that competitive in price accept on the low end. They would bleed AMD by forcing them into slim margins on their best sellers. AMD fought back by trying to price their stuff within Intel's medium and high end consumer stuff. But realistically even AMD knows it can't win a price war. Price your stuff so that you sell what you can, slowly grow production. Get as much ASP as you can while you can. That is how AMD stays alive and that is why Intel won't drop their CPU's to a fourth of the cost just because one person thinks server CPU's shouldn't be that expensive.

4K for a CPU with that much resources over what Intel is charging 8k. Is perfect. They can soak up as much of the I need the resources but I don't want to spend 8k market. Everyone gets awesome value and in the end it's just expensive enough that it won't put pressure AMD can't compete with on Intel.

 

[scannall]

.

4K for a CPU with that much resources over what Intel is charging 8k. Is perfect. They can soak up as much of the I need the resources but I don't want to spend 8k market. Everyone gets awesome value and in the end it's just expensive enough that it won't put pressure AMD can't compete with on Intel.

Just to be clear, I don't think Intel makes a bad product. I don't. They have a very good product. They are just a bad value at the moment.

 

[raghu78]

AMD is just getting started with Ryzen and Epyc. GF is adding 14nm capacity in 2018. I think its obvious that AMD is confident of using up every wafer they can get. I think Intel will start feeling the loss of market share and revenues in 2018. Lisa Su told that server market share increase will take some time. But we should see a double digit server market share for AMD by late 2018. I think the real opportunity for AMD to get to >20% market share lies with 7nm Zen 2. Lisa Su mentioned that 7nm tapeouts are planned for later this year. GF 7LP is ready for customer tapeouts with risk production in early 2018 and volume production ramp by late 2018.
AMD has already wisely moved to multi die approach. So they should have no problems ramping Zen 2 server chips as the die size is likely to be well below 200 sq mm in die size. The rumoured specs for Zen 2 based server chips is 48C/96T. GF 7LP brings a 55% die shrink and with only a 50% increase in core count we should see a reduction in die size. Intel is expected to move to multi die with EMIB at 10++ in 2020. So their first gen 10nm server chips are still going to be monolithic. With Intel 10nm delayed and severe yield issues even with < 100 sq mm dies with their first gen 10nm process it remains to be seen how Intel is going to ramp 400 sq mm dies in 2019 , even on 10+ (which is likely to have process and yield improvements). There might be a chance that AMD get 7nm Zen 2 server chips out before Intel Icelake server chips. That would be some turnaround.

 

[Doom2pro]

Wow that 8c/16t part... It's literally the shining corner of Infinity Fabric, allowing AMD to take garbage 8c/16t dies with only TWO functional cores they would normally throw away and gluing four of them together to create a $400 server part with all 128 PCIe lanes and all 8 DDR4 channels. Truly EPYC.

 

[stockolicious]

That's not really true. This is a different market for sure now, but back when AMD clearly had better server products they only managed max 20% of the server market.

every 5% market share is 900m in revs at this point and it was 25% - also the situation was different in that intel had ways of managing AMD's market share - and AMD did their own manufacturing back then which knee capped them too. From a server perspective they have never been in this good a place before from a business model perspective. They have a competitive server offering and some might argue an elegant one in that the MCM strategy was a last resort with no money to do anything but that - the infinity fabric has so far shown itself to be a savior. AMD doesn't do their own manufacturing anymore and has redundancy so there is less risk there. OEM's are dying for an alternative that has what AMD is offering a multi year roadmapp - I'M not saying AMD is some big threat to INTC but im thinking they are in the game big time. If you look at where INTC see's their future they are looking to different places, they know the monopoly in servers is going away.

 

[.vodka]

Wow that 8c/16t part... It's literally the shining corner of Infinity Fabric, allowing AMD to take garbage 8c/16t dies with only TWO functional cores they would normally throw away and gluing four of them together to create a $400 server part with all 128 PCIe lanes and all 8 DDR4 channels. Truly EPYC.

Yeah.

For AMD's sake we'll see if this slide is realistic

and holds up across the board or at least on the most important workloads -see (1)- because if it does even in these cases, they've literally hit gold...

Whatever the case, if they can get this to sell, that will be some very much needed and deserved money. Zen is a great base to iterate upon, unlike Faildozer.

 

[stockolicious]

Lets not pretend that in one product launch AMD is ready to ship an unlimited amount of CPU's for EPYC to outpace Intel.

your right - that will never happen - but it may turn out that with the redundancy in their manufacturing partners and an elegant MCM design they can have a material affect in the server HEDT space. Also, if we daydream a bit and see a future where servers need CPU's and GPU's connected in more efficient ways im thinking AMD will stay in the game there

 

[beginner99]

that 8 core is for datacenters that need lots of pcie/disks with little processing power so an higher tdp wouldnt make sense

memached server essentially.

 

[beginner99]

No sys admin ever got fired for buying Intel. AMD will have a lot to prove before anyone takes them seriously in the server space.

No sys admin ever bought Intel. They bought HP, Dell, ....

 

[beginner99]

Intel spent 3.4 billion on R&D in Q1

But as we can see there are diminishing returns because if you combine Intels and NV R&D to AMDs which competes with both, we can only say that AMD is very, very efficient. Because they have to be.

if corporations get large and get too comfortable, inefficiency will rise astronomically. In that R&D cost you have all the cost including non-technical people like project managers and as usually in large corps tons of in essence not really needed manager and coordinator positions. The company (10k+ employees) I work for is market leader world-wide in the specific niche. Yet the daily inefficiencies I face are astounding. It makes we wonder WTF our competitors are doing as they clearly need to be even less efficient. In fact in many cases less people / employees would lead to more stuff getting done: less interruptions, less meetings, less bean-counters,...It's scary how inefficient these orgs are and that's why it is for me not all that surprising AMD can keep up. They shed all that dead weight because they had to.

EDIT: It also means intel isn't doomed due to the threat of AMD and ARM. They can and will have to adjust as well. Their revenue might go down a bit while their expenses will go down a lot when they start shedding dead weight.

 

[krumme]

What i am most frightened about is that a new compettitor enters the scene with loads of capacity for a product that can do more or less the same as mine (ours). Its a nightmare.
I will eat eat my hat if thats not how Intel executives is thinking right now.

Now they have to fight gf samsung and tsmc 7nm capacity backed by Ibm basic process research sharing cost with apple and all phone users in the world. With a damn 200mm2 die that can be reused near 100% yield. What a mess!!! Can it be worse? Its hard to imagine.

Its far far worse situation from a business perspective than using a p4 to tacle a gf capacity constrained amd in 2004. Far worse. I am sure Intel will adapt as they have over the years.

 

[imported_ats]

Intel's current top of the line Xeon. http://ark.intel.com/products/96900/Intel-Xeon-Processor-E7-8894-v4-60M-Cache-2_40-GHz

24 cores. 32 pcie lanes. 4 channel memory. $8898. I'd say Epyc makes a pretty compelling argument. 32 cores. 128 pcie lanes. 8 channel memory. For less than half the cost.

Except that design can scale up to 8+ sockets while Epyc appears to be limited to two sockets (and is already an 8 node system at 2 sockets).

Also Intel will be replacing their whole server lineup this year with significantly higher performance parts.

 

[tamz_msc]

Except that design can scale up to 8+ sockets while Epyc appears to be limited to two sockets (and is already an 8 node system at 2 sockets).

Also Intel will be replacing their whole server lineup this year with significantly higher performance parts.

It isn't limited to 2 sockets - it is intended for up to two sockets, as 2P systems are maximum volume.

 

[imported_ats]

It isn't limited to 2 sockets - it is intended for up to two sockets, as 2P systems are maximum volume.

It is limited to 2S until someone actually demonstrates it going beyond 2S. E7s are available up to 64 sockets. So far AMD hasn't demonstrated any intention or capability for EPYC beyond 2S.

 

[tamz_msc]

It is limited to 2S until someone actually demonstrates it going beyond 2S. E7s are available up to 64 sockets. So far AMD hasn't demonstrated any intention or capability for EPYC beyond 2S.

They aren't targeting anything more than 2S, Infinity Fabric is advertised as scaling up to 2S. It is what it is.

Just like Intel hasn't demonstrated any intention of capability for HEDT CPUs on consumer platforms supporting ECC memory.

 

[scannall]

Except that design can scale up to 8+ sockets while Epyc appears to be limited to two sockets (and is already an 8 node system at 2 sockets).

Also Intel will be replacing their whole server lineup this year with significantly higher performance parts.

OK, fair enough. Here is an Intel $4,000 ($4115) 2S Xeon. Priced the same as the top end Epyc. E5-2699 v4.
That gets you 22 cores/44 threads. 4 channel memory. And 40 pcie lanes. Versus 32 cores, 64 threads, 8 channel memory and 128 pcie lanes. The Epyc has quite a bit more horsepower per rack than the same priced Intel option. And with so many pcie lanes, the OpenCL, GPGPU potential is huge.

 

[LTC8K6]

OK, fair enough. Here is an Intel $4,000 ($4115) 2S Xeon. Priced the same as the top end Epyc. E5-2699 v4.
That gets you 22 cores/44 threads. 4 channel memory. And 40 pcie lanes. Versus 32 cores, 64 threads, 8 channel memory and 128 pcie lanes. The Epyc has quite a bit more horsepower per rack than the same priced Intel option. And with so many pcie lanes, the OpenCL, GPGPU potential is huge.

That's the old Broadwell, though. Difficult to make comparisons when we don't yet have solid info on Skylake-SP.

 

[frowertr]

No sys admin ever bought Intel. They bought HP, Dell, ....

Oh gimma a break. You really cant read betwen the lines of what I was saying? Ok, "No sys admin ever got fired for choosing Intel as the CPU".

Better?

 

[krumme]

That's the old Broadwell, though. Difficult to make comparisons when we don't yet have solid info on Skylake-SP.

Still 14nm and we have spec int numbers from Intel on the hedt line with 15/10% perf uplift. Hardly any surprices here. Its tdp constrained. What is more unknown is epyc efficiency and as power and cooling cost is a major factor in tco its still a wildcard.

 

[LTC8K6]

Still 14nm and we have spec int numbers from Intel on the hedt line with 15/10% perf uplift. Hardly any surprices here. Its tdp constrained. What is more unknown is epyc efficiency and as power and cooling cost is a major factor in tco its still a wildcard.

Yes, but we were talking about the number of PCI-E lanes and "horsepower", and pricing, which as far as I know, we don't know for SL-SP?

 

[dark zero]

A lot of that R&D is on process however. And other areas that Intel sells in, like memory, cell radios and network connectivity to name a few. They've been refining and fiddling with the same CPU architecture for a long time now. And why wouldn't they. Huge lead, and brand new architectures are a giant multi year and multi billion dollar crap shoot. With more duds than winners. Nobody on these forums, or at Intel thought AMD would ever get close enough to be in the ring. But they did. With a solid product that will sell well. I'm not sure if AMD has the capacity yet to pump out the volume to make Intel really take notice yet though. So I don't see Intel really slashing server prices yet, by a significant amount. Next year this time? Unknown.

One thing is known though. Intel needed a kick in the pants. Maybe that will shake up their internal structural problems. And AMD should finally become profitable again, and keep applying the boot. We all win, so yay for us.

That is job for Qualcomm.... AMD didn't had the luck of being helped, unlike VIA and the Chinese... They are now with Lenovo

 

[Jan Olšan]

E7s are available up to 64 sockets.

Not true, E7s are 8-socket maximum - the 8800 SKUs that is. Where do you think the first "8" comes from?
64 sockets would only be possible with extra glue logic, at which point you can have more than 2S with Epyc too...