IMO AMD is really going to shine when they release the server version of Ryzen (Naples). Ryzen was great and certainly delivered on performance and price but it still gets beat by Intel in applications which are no heavily multi-threaded. In the server space, there are many almost entirely parallel applications where this advantage will be even better.
Intel's gross margin is 60%. And That is including consumers and lower end product like Pentium. I am willing to bet their Server product are at least 70%+ margin. Which means a lot of room for AMD to strike.
The problem is, what if Intel react and lower their Margin. Hence their Mobile and DC First for 10nm. Having said that, lowering Margin is easier said then done. There will be immense pressure from shareholders and investors.
It will be interesting to see how this story unfold. Of course GF is also working on 7nm ( similar to Intel 10nm ), but Intel's 10nm still has a year lead.
I don't know about the PC Pentium chip, but on laptop Celerons and Pentiums Intel must have like 100%-150% margin.
Those are not Core chips anymore (since post-Haswell). They are Atom chips, which Intel used to make for $30-$50, which the company is now selling for $100-$150.
Making such rip-off chips is one of the reasons why they merged the mobile business and the PC business, so they can show higher profitability "overall" for the division (especially since Intel used to subsidize the mobile chips, in order to compete).
Intel's process advantage used to be more like 3 years. It's now down to 1 year. This happened for two reasons: the other foundries reached FinFET, too (Intel was 2 years ahead just from FinFET alone before), and Moore's Law slowed down first for Intel, allowing the rest to catch-up (and it's now slowing down for them, too).
However, even with Intel's one year advantage, AMD could use Intel's chip building strategy to best it. What do I mean? I'm talking about how Intel will build server chips first on 10nm, and PC and laptop chips later.
If AMD starts building 7nm laptop APUs ASAP before Intel puts out its 10nm laptop chips, AMD could actually be seen as the "better option" overall in laptops (not just in value, in which they will be the winner anyway, but in pure performance).
Intel has also signaled in other ways (other than the 10nm strategy) that it will start fading out of the PC market to focus on "more profitable" server/machine learning markets, because there's not much for it to gain from the shrinking or at best stagnant PC market where it has 90% market share. However, AMD could grow from 10% or whatever it has now to at least 50% in the next 5-7 years, if it plays its cards right.
The only question is how interested they are in the PC market, too, because they may also chase the more profitable server/machine learning markets first. I'm not going to say with confidence that this is a mistake. They may need to focus on that at least partially to become profitable (which they haven't been for years), but I think it would be a better strategic move for AMD and for its image, if it started "conquering" the PC market first, before worrying about "what's the more profitable strategy."
If I were at AMD, I would probably worry more about Qualcomm in the PC market for the next few years (unless Windows on ARM is still a catastrophe, but I have a feeling it won't be, and that regular users in emerging markets won't notice a difference at that Celeron/Pentium level of performance). And the reason for that is that I suspect that even with AMD's sever cost-cutting over the past few years, Qualcomm's ARM chips may still be more profitable to build at a higher margin, which may mean that Qualcomm's chips could offer better value in budget laptops than either AMD or Intel. But that should be revealed to us in the next year or two at most.
And that's still a huge understatement. Intel's strategy for long term viability (that is, investing a ton on fab research) depends on those huge margins.
Who's going to buy and deploy those chips though? Don't AWS, Google, Azure use Intel exclusively and I assume have lock-in contracts? The exception are GPU's, which AMD (NVIDIA) has some cloud penetration[1] but still lots of work left to do.
If Backblaze is an indicator of large cloud provider spending habits then AMD will be given a thorough examination. Their annual HDD Reliability Reports show that while they may favor a vendor, they definitely do not put all of their eggs in one basket.
Google were buying opteron back in 2006.
Granted their stack is more larger and much more complex now, they might consider amd again if the performance and the price is right.
At the very least, it should force some deep discounts by intel.
Actually at google's scale I'd expect them to also be looking heavily at Power/Performance which AMD seems to have an edge on in the consumer Ryzen's. That alone might be worth it for them to start buying.
Frankly, Ryzen's efficiency improvements are marketing nonsense. The only way they hit their TDP numbers is at single-core loads, under a full-CPU load they have fairly similar performance-per-watt to Broadwell-E.
Of course that's on consumer chips with higher clocks for gaming - the math might be different when we get it down closer to 2 GHz.
Anecdote time: my 5820K (4.13 GHz all-core at stock voltage) pulls 92W during a Handbrake encode. It pulls ~140W during a Prime95 run. I think that's roughly comparable to what 6C Ryzen is pulling at equivalent clocks - if not less.
I just built a Xeon E5-2650v3 ES (10C at 2.3 GHz all-core) system last week for video encoding ($60 mobo from MicroCenter, $150 processor off eBay, threw the rest together from parts). It pulls 60W doing the same encodes, and hits roughly the same framerates. Single-thread performance is pretty bad though.
Remember that Ryzen is built on a process that is optimised for mobile applications. The frequencies it is being sold at are well above its optimal power efficiency, and the limited overclocking potential agrees with this.
In the Anandtech forums there are some benchmarks of underclocked, undervolted Ryzen and the performance/W is frankly astonishing. Well beyond anything Intel produces, and yes that includes 10C 2.3GHz Xeons, even modern ones.
Moreover, the consumer Zen parts have the local power plane regulators fused off. Now, I don't know if this is due to design bugs or product differentiation, but if they do have them enabled for Naples that is another big efficiency bump waiting to happen.
TLDR: you can't extrapolate the performance of server parts based on desktop parts, because of different silicon profiles and server-specific power features.
Yep, that was just me bragging about my ultra-dirt-cheap video encoding build. I'm not implying in any way that that would be scalable for a datacenter.
Although in a lot of ways it does appear to be more stable than Ryzen given its memory problems. Especially regarding ECC. I trust a relatively mature platform like Haswell-E much more than something like Ryzen where even basic memory functionality is touchy and ECC is entirely unvalidated. Given a budget build I know which I'd trust.
I was more aiming at TDP figures there. Although I realize anecdotes are not super reliable data, I put the Xeon box together over the weekend and I was literally just looking at the power data last night to see how it was doing.
I was worried about cooling at first, before I knew the TDP numbers it would hit. It's on a Hyper T4 and it didn't break 40C during an overnight x264 encode so I'm pretty happy with that.
1. A lot of Intel consumer platforms (which Ryzen is) had early issues with DDR4 on Z170; there's lots of forum posts about that.
2. AMD said it was not going to be validating ECC for consumer processors.
3. Haswell-E you're referencing officially only supports up to DDR4 2133Mhz memory. Depending on your configuration Ryzen officially supports more (2666, 4000)
4. You're comparing a $1200 chip to $500 chip at retail.
5. You're comparing the TDP on processor running at 1 to 1.5GHz more.
6. ES samples tend to run at lower clock speed then official final processors with hopefully lower temperature as well.
This just isn't remotely an apples to apples comparison. I'm curious what the comparison will look like once there's HPED AM4 chips.
On the flip side I personally would not trust a ES chip personally. The chips are know to contain unfixed errata addressed in further versions of the silicone and a lot of motherboard do not contain microcode updates for those.
It depends on the stepping of the ES chip - in the case of the 2650v3 there's an M0 stepping... and an M1 which is the same die as the production version.
It does appear to be the M0 stepping based on the eBay auction. I don't suppose you have any place you can pull errata for stepping M0 so I can know what buttons not to push here?
In CPU-Z I am getting something different from the seller's published CPU-Z screenshot - it's showing stepping "R2" ("retail"?) - however the name string is also still showing as "ES".
I've attached a dump from CPU-Z that shows CPUID 0x0 and 0x1 values. From the processor value (0x000306F2) and the Intel data I could dig up - it would appear that this is actually the SR1Y1 stepping. Which is apparently the M1 stepping. Possibly actually a QS rather than ES.
Not holding out hope overall given that it was sold as the M0 stepping but I guess you never know, maybe I got lucky. There was no pic on the eBay auction, and I would have to pull the cooler back off to check what's printed on the processor itself.
(anecdotal evidence: the ~140W package TDP I observe with Prime95 on a 6-core (analogous to the 6850K) is just about right for 178W total system power at 80% efficiency, so their numbers ring true here. I just ran another Prime95 SmallFFT test and I came up with 133W package TDP on HwMonitor64.)
Basically AMD is low-balling their TDP here (they rated the TDP for like 1-2 core loads with maybe a mild load on the other cores, is the official word IIRC) and Intel is high-balling their TDP, something like mixed-allcore with a decent amount of AVX thrown in.
In practice: both 8-core processors at default clocks are around 125W at a full normal (non-AVX) workload. Both 8-core processors are around 200W with an aggressive overclock. The 6900K tends to win modestly in most cases in either scenario (CineBench is like AMD's best-case). The general take was that the 1800X was a win on price vs the 6900X, not an overall win on performance or efficiency.
To sum up, the processors are within 10% power consumption on any given workload which is pretty much neck and neck as far as I'm concerned (normal variability). It's a good result for AMD given their Bulldozer track record but it's not an obvious win in efficiency. Again, of course the picture for server chips at 2.5 GHz might look very different than the picture at 4 GHz but that's speculation.
I don't know how well I trust your sources. Tom's Hardware shows typical 1-7% idle load for a Windows machine using Ryzen 1800x is about 14 watts. And your link shows power consumption idle for the same Windows at 70+ watts from .bit-tech.net and the other pcper shows almost 40watts. So you can see how that's a little contradictory. To be clear Tom's Hardware notes the numbers are for cpu only - not sure about your links.
The problem is that most PSUs have really terrible low-end efficiency. 80-Percent Plus's efficiency standards only kick in at 20% load, so the 1500W PSUs in the test systems sites use (eg your link - Silverstone ST1500) are massive overkill and you're running them at like 1% load. At those levels efficiency trails off to pretty much nothing - there is more or less a constant load at very low levels (i.e. efficiency approaches 0% near 0W load)
Looking at package-TDP only gives you a totally different picture. Not invalid necessarily - but different. It's easier to tease out the system under test but it's more difficult to get a holistic picture.
(As such, it is much much easier to cook the books when all you have to do is lie through a sensor readout than when you have to lie to a meter at the power plug. This is usually a game that gets played with GPUs, i.e. how the RX 480's "GPU-only" TDP excludes power consumption from memory, losses to VRMs, etc)
Also - servers are pretty much never at 0% load. Any system that has CPUs at 0% load consistently was over-allocated in the first place. Ryzen does idle a little bit better than Broadwell-E - although factoring out the PSU inefficiencies near the zero bound it's probably more like 25W for Broadwell vs 15W for Ryzen.
Right now HwMonitor64 says my 5820K is idling at the desktop at 25W FYI - "as reported by the CPU".
Google at least has hedged Intel domination by building PowerPC servers themselves. I doubt any major provider would have a 100% lock-in contract. It would be more likely they would contract for XXXX servers per year, where XXXX is some percentage of their likely demand.
This is all speculation of course, and if anybody actually knows that would be awesome. I think it is totally believable and reasonable that Intel would have exclusivity clauses in their contracts with the "big 3" cloud providers.
The PowerPC thing isn't speculation[1]. Intel doesn't have the market power to demand exclusive contracts - the big 3 can just say "no" and what would Intel do?
Intel didn't need exclusive contracts: they simply had a much better CPU, cheaper.
I suspect you've asked a bad question. It should be obvious that anyone who does know such a thing could not share it since it would violate any kind of NDA or basic 'secret sauce' / leverage based competitive advantages agreements they would have surely signed as even basic employment.
They might, but I'm sure there are a ton of large companies that run their own datacenters or at least operate their own bare-metal servers (e.g., Dropbox, OVH). It would at least be a start for AMD.
There are also still plenty of businesses that have their own hardware on the 0s - 10s of servers scale. They are going to be much more cost sensitive, so AMD could really shine there.
Separate to cloud providers, there is a whole market for home servers, especially storage/NAS devices. AMD could tap into this, not only the datacenter.
At the kinds of scale the big cloud operators buy and build at, I bet each watt of TDP probably costs them six figures a year. If Ryzen has real advantages in the server room, they'll figure out a way to deploy it, I doubt they're foolish enough to lock themselves into one supplier irrevocably.
I'm excited and looking forward to Ryzen. I can't wait to see how well it performs for us as a CDN server. However, I'm concerned about the heavily NUMA nature of even a single socket system (4 dies per package, see http://www.anandtech.com/show/11183/amd-prepares-32-core-nap...). Depending on the bandwidth of the on-package interconnect, things could either be just fine, or it could be a non-starter.
Part of the problem is FreeBSD just not being very NUMA aware. A less obvious part of the problem is even if FreeBSD was NUMA aware, we'd still have problems with files read off storage on node 0 having to exit the system via a 100GbE NIC on node1, etc.
It's going to be tough for AMD to claw their way back into this market anytime soon. The market for low-end to mid-range servers has been on the decline for a while and the explosive growth from cloud providers is starting to level off too.
IDC seems to think so. I think they have a pretty good track record. From a user perspective it feels right to me because it's easier than ever to scale on demand and otherwise manage cloud resources efficiently.
Who says they aren't going to be high end servers? If their server lineup is anything like Ryzen it will well surpass what Intel is offering in performance.
