AMD X399 !!!!!

Page 20 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

blue11

Member
May 11, 2017
151
77
51
None of this even you compute load out changes the fact that if I told that company that they could spend 10 million less dollars and get more CPU resources the Purchaser would go, well that's only 4% of the project's cost so lets just ignore it.
What you don't seem to understand is that buying more expensive CPUs saves money. The more CPUs you fit into one server, the less DRAM you need to buy. What's the most expensive component again?
 

ryzenmaster

Member
Mar 19, 2017
40
89
61
Why argue against strawmen? All I said was that 8ch memory controllers are inefficient and unnecessary without strong compute performance (i.e. SIMD).

Much like the SIMDs we just discussed, quad channel and 8 channel memory will not be useful in every scenario. Yet if we look at the ongoing trends, there are plenty of cases where high bandwidth does come in handy. Does the ever so annoying buzzword "big data" mean anything to you? Handling large data sets means you need all the bandwidth you can get. It does also seem to be a trend to increasingly cache databases and key value stores in RAM. The volume of content online is rapidly expanding and disk i/o cannot keep up with it. Coincidentally this also makes it easier for the marketing guys to actually sell Zen: all they have to do is to throw out some buzzwords like "big data" and how they have the memory bandwidth to handle it and it'll be easy sales compared to boring lecture about vectors and math and stuff that nobody even cares about.

Not all computation can be vectorized and not all data passing through RAM need heavy compute, so your fixation with SIMDs and how everything with the Zen must be failure if it can't beat Intel in AVX sound a bit silly to be honest.
 

blue11

Member
May 11, 2017
151
77
51
Does the ever so annoying buzzword "big data" mean anything to you? Handling large data sets means you need all the bandwidth you can get. It does also seem to be a trend to increasingly cache databases and key value stores in RAM.
In-memory databases, i.e. random access across multiple TB of memory, can not be read at anywhere near the limits of 4ch memory controllers, let alone 8ch. To make things worse, the memory latency of Zen will hurt those same memory databases, because of dependent reads in standard operations like indexed search.

Not all computation can be vectorized and not all data passing through RAM need heavy compute, so your fixation with SIMDs and how everything with the Zen must be failure if it can't beat Intel in AVX sound a bit silly to be honest.

Why argue against strawmen?
 

Topweasel

Diamond Member
Oct 19, 2000
5,436
1,654
136
What you don't seem to understand is that buying more expensive CPUs saves money. The more CPUs you fit into one server, the less DRAM you need to buy. What's the most expensive component again?
Tell that to Itanium buyers.

You completely Dodge my original point before you went on your smallest cost tangent.

My point was that A. AMD still has to earn goodwill in the Workstation community. B. The value in what is now EPYC is being able to offer more resources at value. They are trying to shake up the market. You do that buy offering a competitive product at a competitive price. Offering a CPU at Intel prices like let's say the 24 core at Intel $9k, or even based on theoretical processing power will do little to get the bean counters to switch platforms.

Saying I offer more at a better cost vs. performance is what gets guys to swap over. I mean outside the few of us that are still little bit fanboys from AMD's glory days with the Athlon or A64 days would get an AMD CPU that only matched Intel in every way. Not many because you feel safer (even if it's not always the case) when you are dealing with Intel. This is amplified in the Server market.

So no more expensive doesn't mean better and Intel wouldn't have 3 dozen server CPU configurations compared to the 7 or so desktop ones if people just bought the most expensive chips because they saved you money.
 

blue11

Member
May 11, 2017
151
77
51
Offering a CPU at Intel prices like let's say the 24 core at Intel $9k
You mean 22 cores for $4k? This is ignoring the fact that major customers don't buy anything at ARK prices. Nobody buys E7s for the 2 extra cores (which weren't even there in Haswell), but to run at 8S with esoteric reliability (e.g. CPU failure tolerance) features.

So no more expensive doesn't mean better and Intel wouldn't have 3 dozen server CPU configurations compared to the 7 or so desktop ones if people just bought the most expensive chips because they saved you money.

Obviously Intel creates multiple SKUs to capture all possible revenue along the demand curve, but the sales of the top bin SKUs is greater than the rest of the list. The server market works in a way nearly opposite to client, where the majority of sales are at the bottom end.
 
Last edited:

thecoolnessrune

Diamond Member
Jun 8, 2005
9,670
571
126
You mean 22 cores for $4k? This is ignoring the fact that major customers don't buy anything at Server List prices.

FTFY. Discounts from the Vendor / VAR is made on the platform, not on the servers, and definitely not on the components. Your statement could have referenced CPUs, Memory, add-in cards, hard drive, or anything. It doesn't matter. There's a total platform sale, and the Vendor / VAR will add in a discount to get sales and hit specific revenue targets. Nothing to do with "ARK prices". That's the Vendor's job to deal with, not the end buyer.
 
  • Like
Reactions: IEC and beginner99

blue11

Member
May 11, 2017
151
77
51
FTFY. Discounts from the Vendor / VAR is made on the platform, not on the servers, and definitely not on the components. Your statement could have referenced CPUs, Memory, add-in cards, hard drive, or anything. It doesn't matter. There's a total platform sale, and the Vendor / VAR will add in a discount to get sales and hit specific revenue targets. Nothing to do with "ARK prices". That's the Vendor's job to deal with, not the end buyer.
The majority of server purchases are direct from component vendors, not through resellers like HP. Typically Google, Facebook, etc. create a server design with schematics and parts and hire an assembler like Pegatron, etc to put it together. Indeed, server purchases are so concentrated in a handful of major customers, that they have a great deal of leverage over Intel regarding pricing.
 
Last edited:
  • Like
Reactions: ozzy702

Topweasel

Diamond Member
Oct 19, 2000
5,436
1,654
136
You mean 22 cores for $4k? This is ignoring the fact that major customers don't buy anything at ARK prices. Nobody buys E7s for the 2 extra cores (which weren't even there in Haswell), but to run at 8S with esoteric reliability (e.g. CPU failure tolerance) features.

So now they wouldn't purchase the most expensive CPU? 4k 9k doesn't really make a difference or whatever the volume pricing ends up at. Right? But seriously make it 4k instead of 9k and slide it down. Instead of fitting everything to your small window and nitpicking points that doesn't matter yet. I proposed the Zen archtecture and the EPYC and ThreadRippers strength is to bring more resources in at value and you jumped at it. If it's 4k for the "realistic" Intel top Core dog. Than selling the 32 core EPYC at 4k brings tons of value. At 6k a lot of the value would disappear and it becomes more of a Packaging issue sure some companies might still jump on it because its nearly 50% more cores per box. But then you're paying the same for AMD product as you would an Intel and that will bother the OEM's a bit, but will be a big red flag to some customers. Even during AMD's heydey they still had to throw CPU's at vendors and customers to get them to hopefully order them in the future. So why sacrifice you biggest strength if you don't have to? Because whether or not perf/$ has direct impact on Server purchasing it still is Zen's biggest asset.

Obviously Intel creates multiple SKUs to capture all possible revenue along the demand curve, but the sales of the top bin SKUs is greater than the rest of the list. The server market works in a way nearly opposite to client, where the majority of sales are at the bottom end.
I do understand how it works and it's more fluid than well the top chips sell the most and it lowers from their but. Yeah the server market is top heavy. That said again if the solution was always to buy at the top then they would only ever offer 1 chip. It certainly doesn't happen that way.
 

Topweasel

Diamond Member
Oct 19, 2000
5,436
1,654
136
The majority of server purchases are direct from component vendors, not through resellers like HP. Typically Google, Facebook, etc. create a server design with schematics and parts and hire an assembler like Pegatron, etc to put it together. Indeed, server purchases are so concentrated in a handful of major customers, that they have a great deal of leverage over Intel regarding pricing.
This is the second time in nearly a month I have someone suggest that the server market is run completely on these BOM and not OEM systems. That's not the truth, not even close. There are a lot of huge data center companies that do account for a very large portion of server sales and probably the most lucrative in terms of singular contracts. But the there are 10's of thousands of companies in the US and hundreds of thousands of companies globally upgrading their servers each year that are not building bunkers storing exabytes worth of data. Deep learning and cloud server solutions are very big markets that are growing but you are ignoring billions being spent in the normal office space.
 

itsmydamnation

Platinum Member
Feb 6, 2011
2,743
3,074
136
Blue11 your arguments are terribly flawed. Zen has more memory channels a socket. You run to examples that fit your agenda without looking at the real world. You know what I see in the real world. That's lots of 10-16 core 2p with 384gb of memory, why? because 16 gb non lr ddr is cheap. So on Zen we can get 2p 24/32core 512 gb while keeping at 16gb dimms.

So in the mid range we are talking core advantage, memory capacity advantage, memory per dollar advantage, memory throughput advantage and pcie advantage ( nvme is getting very cheap).

You then have things like sve which in my market (government) will be a massive advantage.

Amd don't have to beat Intel everywhere, look at the average server running your average workloads and Zen Imo is very compelling.
 
Last edited:

thecoolnessrune

Diamond Member
Jun 8, 2005
9,670
571
126
The majority of server purchases are direct from component vendors, not through resellers like HP. Typically Google, Facebook, etc. create a server design with schematics and parts and hire an assembler like Pegatron, etc to put it together. Indeed, server purchases are so concentrated in a handful of major customers, that they have a great deal of leverage over Intel regarding pricing.

Where'd you get that idea from? In Q4'16, everyone not the big 5 (Dell, HPE, Huawei, Lenovo, and Inspur) made up 42% of server shipments. That includes not only assemblers like Pegatron, but also small server manufacturers like Gigabyte. What you stated is factually incorrect. Outside of the US, like in the EMEA it's even worse. Everyone not top 5 only made up 25% of shipments.
 
  • Like
Reactions: tamz_msc

blue11

Member
May 11, 2017
151
77
51
Blue11 your arguments are terribly flawed. Zen has more memory channels a socket. You run to examples that fit your agenda without looking at the real world. You know what I see in the real world. That's lots of 10-16 core 2p with 384gb of memory, why? because 16 gb non lp ddr is cheap. So on Zen we can get 2p 24/32core 512 gb while keeping at 16gb dimms.

So in the mid range we are talking core advantage, memory capacity advantage, memory per dollar advantage, memory throughput advantage and pcie advantage ( name is getting very cheap).

You then have things like sve which in my market (government) will be a massive advantage.

Amd don't have to beat Intel everywhere, look at the average server running your average workloads and Zen Imo is very compelling.
It's funny that you mention the "real world," because I look there and I see the market increasingly moving to Cloud services, because the economies of scales from buying at the higher end and consolidating are quite compelling. The "average server" running an "average workload" these days is located somewhere in Amazon, occupying part of a 12-22-core chip (depending on age).
 

ryzenmaster

Member
Mar 19, 2017
40
89
61
In-memory databases, i.e. random access across multiple TB of memory, can not be read at anywhere near the limits of 4ch memory controllers, let alone 8ch. To make things worse, the memory latency of Zen will hurt those same memory databases, because of dependent reads in standard operations like indexed search.

The point here is that there is an increasing demand for more memory capacity and bandwidth as larger data sets are being passed through. You're saying it makes no sense, but I'm saying from marketing point of view it makes a lot of sense. If AMD are to have any success in enterprise markets they need to have at least some selling points. I would argue that higher memory bandwidth is something you can sell to enterprise customers. They do not need to and cannot beat Intel at everything, but they do need to become a relevant player in that market segment. More cores/threads per socket and more memory bandwidth are one way to distinct themselves.
 

thecoolnessrune

Diamond Member
Jun 8, 2005
9,670
571
126
It's funny that you mention the "real world," because I look there and I see the market increasingly moving to Cloud services, because the economies of scales from buying at the higher end and consolidating are quite compelling. The "average server" running an "average workload" these days is located somewhere in Amazon, occupying part of a 12-22-core chip (depending on age).

This also isn't true. Oppenheimer puts the 2016 estimates for cloud usage as overall server base at a little over 17%. Even as all cloud approaches the point in the next couple of years where they'll have the majority of compute capacity available (Microsoft, Amazon, and Google have over a million servers for instance), this overall capacity is still a shell of the actual used server base (where workloads are actually getting run), which is estimated to be between 75 and 140 Million Servers.

Also, Amazon has no 22 core chips offered. The most they offer is the the E5-2686 v4, and the E7-8880 v3 with 18 cores. Ironically the latter is used to break the memory bottleneck for VMs with very large memory footprints, by leveraging additional sockets. In those situations, would AMD's additional 512GB per socket that they're offering make a difference? Maybe. It would likely be doing so through additional memory slots, and that could be very compelling, as Cisco notes that most of their Virtualization platform shipments are with 256-512GB of RAM, populating all slots with much cheaper, smaller DIMMs. It also has the effect of minimizing faults, as while density is a major factor for virtualization, most Virtualization deployments target no less than 5 hosts, and after that it turns to even numbers when targeting 4-node blocks.
 
  • Like
Reactions: lightmanek

beginner99

Diamond Member
Jun 2, 2009
5,208
1,580
136
The idea that cpu cost is so small it doesnt matter becaue it eg isnt shown in a tco analysis from Gartner is pathetic. Every cost matters. All over. All the time. Every penny. You cost optimize your company all over or somebody takes your chair tomorrow.

But cost isn't entirely monetary. And switching hardware has a risk and hence cost associated with it as well. Saving $1000 heck even $5000 on the CPU for hardware that runs software costing millions just doesn't matter all that much. Managers are influence by brand value as well and most of them have never heard of AMD...
 
  • Like
Reactions: Ajay

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,683
136
Saving $1000 heck even $5000 on the CPU for hardware that runs software costing millions just doesn't matter all that much. Managers are influence by brand value as well and most of them have never heard of AMD...
So Intel can triple the prices and no manager will care. Interesting dynamic.
 
  • Like
Reactions: Drazick

Atari2600

Golden Member
Nov 22, 2016
1,409
1,655
136
What you don't seem to understand is that buying more expensive CPUs saves money. The more CPUs you fit into one server, the less DRAM you need to buy. What's the most expensive component again?

Well.... Naples can put 64C or 128T on a 2P server. It can also carry more memory on that blade than the best Intel alternative.

So, who will save the most money again?
 
  • Like
Reactions: Drazick

beginner99

Diamond Member
Jun 2, 2009
5,208
1,580
136
So Intel can triple the prices and no manager will care. Interesting dynamic.

Obviously within limits.

But the analogy is more that if GM wants triple the price for wheels compared to Ford doesn't make the whole car 3 times more expensive. If you not these wheels they served you well and reliably, why change?

All I'm saying is just delivering a good CPU is far from enough to succeed. It should offer something special and as has been said more memory channels actually helps a lot not because of BW but because of total amount of memory per socket. Low usage applications running on separate VMs don't need much CPU but lots of RAM. Databases also profit from tons of RAM.
 

Atari2600

Golden Member
Nov 22, 2016
1,409
1,655
136
Saving $1000 heck even $5000 on the CPU for hardware that runs software costing millions just doesn't matter all that much. Managers are influence by brand value as well and most of them have never heard of AMD...

I would be very surprised if any purchaser in a company at that level of expenditure wasn't aware of every single market player.
 
  • Like
Reactions: Drazick

krumme

Diamond Member
Oct 9, 2009
5,952
1,585
136
But cost isn't entirely monetary. And switching hardware has a risk and hence cost associated with it as well. Saving $1000 heck even $5000 on the CPU for hardware that runs software costing millions just doesn't matter all that much. Managers are influence by brand value as well and most of them have never heard of AMD...
Ofcource hardware cost is miniscule.
But every cost can be measured monetary and is done so. Its a processional job. (Eg waste time waiting for server response)
My point was driving a business every cost counts.

Nearly 20 years ago i was a management consultant doing tco benchmarking on the organization total IT solution and benchmarking it.

At the meetings outlining the process and later presenting the result was typically the CEO. Often the CFO and sometimes the often unwilling CTO that was forced into this benchmarking.

Most of the view on IT on that time was centered around cost. Therefore IT was ironically often placed under the CFO management. A pretty defensive move but it shows how obsessed you are with controlling cost.

You have a pretry strong eye to where you get most for your money. All over. Intel cant charge twice as much for their Xeon line even though eg ddr prices are a major cost (same could be said for cooling, space whatever). It would hurt their own revenue and profit. Sale would crater.
 

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,683
136
But the analogy is more that if GM wants triple the price for wheels compared to Ford doesn't make the whole car 3 times more expensive. If you not these wheels they served you well and reliably, why change?
But this is not about just one small change in operating costs. They cut costs wherever possible, because they add up. As krumme outlined in the post above, it is deeply embedded in top management thought process. Sure, it may be a bad choice in isolated cases, but in the bigger picture it pays off.

There's always a price threshold where decision makers will start asking questions. Whether that's the case here or not, remains to be seen.
 
  • Like
Reactions: Drazick

Veradun

Senior member
Jul 29, 2016
564
780
136
Infact I think you will find that there is sufficiently large enough portion of the Market that buys 20k and cheaper servers maybe even business that primarily run off of 10k or less servers.

My servers are around 11k each. 48 HECs and 384GB each.
 
Last edited:

Atari2600

Golden Member
Nov 22, 2016
1,409
1,655
136
I have to admit, with X399, I'm not so interested in needing the fastest CPU AMD can produce right *now*.

I just want to be on the platform for if I need it later on - and the enhanced I/O will increase system longevity.

Thus, a relatively cheap 8 core chip with high clock would be of great interest to me. If and when I need the horsepower for work - I'd upgrade the CPU to the latest and greatest available *then*.
 
  • Like
Reactions: Drazick

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
What you don't seem to understand is that buying more expensive CPUs saves money. The more CPUs you fit into one server, the less DRAM you need to buy. What's the most expensive component again?
I can't just read your amok any longer, I have to ask: how many % of workloads benefitting from more CPU power versus less RAM at the same time????