Embedded/semi-custom. I think the issue is not whether AMD can get volumes or not, but whether they can generate enough cash inflow in a time frame tight enough to support a bleeding edge pipeline. And I think th answer to this question is no.
Did you ever build a server? I'm asking this because once in servers, and Shintai was referring specifically to Xeon, the price that the middle class can pay is not the issue (I love this unionist cliché, it's like back 50 years into the past), but TCO is, and for some cases even if AMD were to give its Opteron chips for free it wouldn't make sense to use AMD over Intel processors.
Last edited:
AtenRa
Lifer
- Feb 2, 2009
- 14,003
- 3,362
- 136
AtenRa
Lifer
- Feb 2, 2009
- 14,003
- 3,362
- 136
CHADBOGA
Platinum Member
- Mar 31, 2009
- 2,135
- 833
- 136
Servers are a good indicator of efficiency optimized chips. You still haven't addressed the fact that you are looking at an efficiency part vs. a performance part.
Pricing is irrelevant to technical merit. AMD doesn't want to sell so low. Intel could sell lower and make a profit. If you want to compare architectures you must compare the fully enabled chips either the performance variants (4770k vs. 8350/8370) or low power variants ( 8370E vs T or S variant i7).
It looks like it depends where you are getting the bench.
You can search around but any gains are likely due to other factors. If AMD was that close to intel they would have no trouble competing in the server market. If you look at mobile (a 4700mq will outperform a 4670k in MT generally) where intel fits in a 35W TDP (35W quad) demolishes anything AMD can fit in a notebook by more than a factor of two; you generally have go to the desktop 6xxx series to get close as FM2+ simply can't compete.
The 8370E is a step forward. Its good to see that. But if you undervolt/underclock both chips to a set power consumption the 4670k/4770k will step ahead.
We have numbers expressed in watts in the review i posted while you re relying in a general argument that has no technical basis.
Besides FX8370E is not only efficient it s also a performance part.
If price was irrelevant then the i7 would sell more than the i5 and the FX8370 is not that lowly priced, you perfectly know that its the price/perf that is the real comparison since we are in a same usage, that is DT.
As you can check, Hfr benches are the most standard ones, not that 7 zip, 3ds Max or Fritz are rarely used, but it s usual for some to negate benches scores once the result doesnt suit their opinion, yet you did ask for softs like 7zip when i started with Fritz, now it s the whole suite that is no more relevant.
The gains in power efficency were due to better binning and are easily understood when reading the review i linked, as for the 4700mq it was measured at 86W in a minimalist test plateform by Anand, get more serious numbers than branding a TDP wich is known to be fraction of the actualy consumed power, keep on computing its perfs watts based on a flawed number, i guess that its like the core M wich consume only 4.5W set appart when it sucks 15W, that is using 4.5W for the perfs/watt calculations of scores made at 15W TDP.
It s not all to reduce the voltage you must guarantee full stability, if Intel doesnt reduce the voltage more then it means that they wont risk instability for the sake of a few watts, so you re still searching around i see...
AtenRa
Lifer
- Feb 2, 2009
- 14,003
- 3,362
- 136
Performance and efficiency was not the problem for 6300 series Opterons.
FM2+ APUs can easily compete against Haswell Core i3. An 8-core SR CPU could easily compete(performance wise) against 4C 8T Haswell CPUs.
Well one is made at 32nm Gate-First PD-SOI when the other is made with 22nm FinFet. Im sure you would have the exact opposite results if FX8350 was made at Intels 22nm FF and Haswell at GlobalFoundries 32nm Gate-First.
If we want to compare at the technical level, just compare Intel Core i7 3820 (32nm server part) to FX8350 (32nm Server part) and see how those two behave. Those two are the closest SKUs technically.
Let me educate you.
See, can you stop spreading lies and mentioning Intel in an AMD thread now?
Last edited:
You manage to keep serious while writing this..??
The quotes you posted doesnt tell the story, use google translate
and read this article if you want to understandand something to the new power management scheme you re trying to explain :
http://www.hardware.fr/news/13889/idf-tdp-complique-core-m.html
Also you should be more cautious before branding other liars, using deffamation is just aknowledgment of your lack of arguments, i m looking at things technicaly speaking not on a marketing point of view like you seems to do, what i discuss are real numbers not hollow sentences sparsed with theorical numbers whose experience and test benches have proved that they are only marketing monikers.
Besides why not talking of Intel since we re comparing CPUs and TDP definitions of thoses CPUs wich are directly competing products.??.
Can you please provide the tests/benchmarks confirming Core M constantly uses 15W during benchmarks? Also feel free to explain how they're able to fit the imaginary 15W TDP Broadwell-Y in a fanless 7mm thick 10 / 12.5'' package. I'll be waiting.
That s not the thread but well, i ll answer to this point since Shintaidk is saying the same thing as you although on a more direct and less lenghty manner.
It s true that TDP can reach 15W only during 10ms but what if the duty rate is 90%..?
What if i let only 1ms duration before each 10ms pulses as was likely the case when CB11.5 was benched.?
The CPU will reach 90% of its peak 15W TDP, let say 12W during 30 seconds for our exemple, but if your cooling apparatus has enough thermal inertia and that the subsequant power usage is say 2W then you can integrate the power usage over a period of 2mn30s wich will yield 4W power usage averaged on this period, that s how TDP is estimated in these new power management schemes, the thermal inertia of the whole device is used as some reserve where you can borrow excess power that will be paid back later when a lower power usage will allow it.
We dont know if AMD is using this kind of PM, what is sure is that Mullins is limited to about 1.5Ghz when fully loaded, estimations are not easy as it could be at 4.5W as well as at 6.5W at such a frequency, so far there has been no accurate numbers which would allow for 10% precision estimation.
Last edited:
What is then? Cause they ain't selling.
Mobile?
The top 8350 (server part) is the 6328 which runs at 3.2/3.8 at 115W and at $~575. Not sure why you are comparing to the 3820 in servers when the competing part would be the E3 Xeon series. Unless you need the quad channel RAM or the 40 PCI-E lanes. But for a simple box (1P system) that would be the competition.
Even in AT's Haswell EP article they mention that the 6300 series isn't competitive vs. sandy bridge.
Sure, but its a low power chip (hence the E).
It costs intel the exact same to build an i5 as an i7. All things equal it possibly costs intel less to build a 4770k die than it does AMD to make a 8350 die. Take your interest as a consumer and stand back from the problem. Evaluating a product on purely technical merit includes the performance characteristics of the product and the price it costs to build. It has nothing to do with the price it sells at because a superior product will command a better price.
7-zip is fine. Fritz suffers from many of the same problems as 3d-PM, namely that there is a massive loss of performance between PII and the construction cores.
PII x4 980 = 9009 points
a10-7850k = 7563 points
PII is 20% faster clock for clock.
When you actually look at the tests you see that the systems are not configured the same.
http://techreport.com/review/24879/intel-core-i7-4770k-and-4950hq-haswell-processors-reviewed/7
Idle power is 12W higher (37W) on the mobile platform because its not a production platform with power optimizations. Also that is a most power hungry SKU. A 4700mq will draw substantially less (without edram as well).
Yes, but intel (and AMD) generally seem to have a lot of extra volts. This may be the reason why Core M is hitting those speeds.
witeken
Diamond Member
- Dec 25, 2013
- 3,899
- 193
- 106
Makes sense, else Intel would have given it a 15W TDP, else it wouldn't have been possible to put in a thin tablet.
So it s not a bench favourable to the FX but it is more to Haswell since it has 12% better score than a C2D so what would be its irrelevancy.?.That it favour not enough the Intel chips.?.
You noticed that there s two other chess game benches to ponder this one, just that the FX seems to appreciate them more than it does for Fritz...
Idle power will be reduced but not the 55W delta, also they say 92W peak power while the graph show that peaks are above 100W and that the averaged power is more 95W than 92W, i guess that they wanted to be sympathetic to the CPU provider.
Intel is more agressive in matter of binning given their bigger production, for AMD it doesnt make sense economicaly to fine grain the binning more than they currently are as there would be risk to not produce enough of scarce anyway parts to just recoup the increased costs of binnings and inventories.
AtenRa
Lifer
- Feb 2, 2009
- 14,003
- 3,362
- 136
Lack of MG&A at the time of release. AMD was in its biggest transformation process in 2012.
Kaveri mobile is doing fine against mobile Core i3/5 Haswell even at low TDPs.
I meant that if you want to talk technically, then you have to compare FX8350 which is a server part against Core i7 3820 that is again derived from a server part. Both have large L2/L3 caches and 40 PCI-e lanes with 6x or more SATA-3 etc.
For server SKUs you have to compare Opteron 6328 115W TDP (same die as FX8350) against Xeon E5-2643 130W TDP (same die as Core i7 3820)
Without knowing the wafer price and yields, die size alone doesnt mean anything.
Last edited:
Phynaz
Lifer
- Mar 13, 2006
- 10,140
- 819
- 126
Yeah, it's his SOP. He hates Intel so much he can't stop himself from railing on about them, even in an AMD thread.
There are certain benchmarks that are very poor and some that are okay. Some are also good. An ideal benchmark will not show undue gains or losses due to architecture. It should not be unduly above or below the average performance gains on a variety of applications.
Fritz is a little weird. If you take haswell and compare it to the baseline P3 @ 1 Ghz (480 Kilonodes) you see that haswell has about 57% more IPC.
i5-4430 = 9023 points / 4C /3Ghz = 752/480 = 157%
Its a terrible benchmark that is no indication of modern processor performance.
Also about hardware fr. I don't know if its the motherboards or what but just looking at the SB -> HW generation idle powers are higher (what on earth) and the peak powers are nearly identical. Their power numbers are weird and different from everyone else.
As long as the benchmark is updated that is fine. But if you examine it and find its not then there is a problem.
I'm not sure about houdini but here are some publisher benchmarks.
http://www.cruxis.com/chess/manual/index.html
Doesn't seem like much scaling between lynfield and ivybridge once you take the clockspeed difference into account.
Please, they aren't biased, they are taking the average. Dropping 12W brings the average to 80W which is still high for a notebook but reasonable. Lets also not forget that I was taking about lower power level chips (4700mq), the 4950hq is, on average the same as the 8350 and faster than the 8370E.
Yeah, okay :whiste:
The CPU gets destroyed against the 35W competition. The GPU is good but it can't stand against a 35W dual let alone a 35W quad.
While true these only matter if they are going to be used. Cache only matters if it helps performance. For a lot of workstations LGA 1155 is sufficient unless you are running a ton of expansion cards or SSDs, even among workstations this is relatively uncommon. If your specific case requires it then you are 100% correct if it doesn't then why spend money. This is the same argument you see people trying to build an equivalently priced imac to hackentosh and someone insists on a thunderbolt capable motherboard even though very few use thunderbolt.
The 6328 runs at lower clocks than the 8350. Sure you can compare it to the E5 but the E5-2643 is about 4x more ($884 vs. 215) than the E3-1230 and has identical CPU performance. Sure it has more lanes, cache and I/O but nobody is going to buy it over the E3 unless they need those lanes, etc.
Which is why I said probably. They are both fairly mature now and 22nm greater (likely) cost is balanced by the fact that intel doesn't pay foundry margins.
Going forward Zen looks to (probably) take the jaguar core and scale it up. For that reason it is unlikely it is a terribly high performance high frequency product but rather focuses on cost and efficiency. Smaller core sizes and greater cache efficiency are important in the server market. IPC will increase not doubt about that (given how jaguar is almost clock for clock as strong as steamroller). But AMD doesn't have a terribly large budget and the amount of time spent on this product is less than BD so as mentioned before the scope and breadth of zen is probably going to be smaller than expected. In my opinion it will likely recycle parts of jaguar and PII and take design cues from the construction cores.
AMD has to play to their strengths to win.
This is a red herring, this is a variant of the same argumentation John Fruehe used to write about on his blogs when he was tap-dancing about the value of AMD server offers. As AMD market share now shows, the argument was flawed to the bone.
Xeon E5-2643 exists for the cases you have workloads that needs tons of single thread performance but at the same time you require tons of PCIe lanes, which means core count aren't a priority. It's a niche SKU in the first place. But what if your workload is more typical for server environments, e.g., core scale matters a lot for you? Then we are talking about packing more cores for the two sockets, meaning that you won't take the 4C Xeon, but likely the 6C or the 8C versions. That is the Opteron competition, that is the performance/watt it must beat and it doesn't.
The fact is, at 32nm Intel can build a 8C/16T Xeon chip that can sustain high clocks, AMD can't, their 16T chips have to pay a huge penalty in clock speed in order to achieve a reasonable TDP. In order to counter that you must go 4P with AMD to reach the performance of Intel 2P servers, and that's bad for TCO. For specific cases AMD would have to pay you to use their servers instead of Intel-powered chips. This is why Bulldozer is a failure on servers, anything else is a smoke screen.
Cerb
Elite Member
- Aug 26, 2000
- 17,484
- 33
- 86
That 1-2x cost difference is probably made up for within a few months of running it, if not within the first month. When they had a good product, lower prices made people used to one system willing to switch. IE, going from Netburst or Core 2 type Xeons to Opterons, instead of just buying whatever their Dell rep was pushing that day. If that's not the case, it barely does any good, except to show that they are desperate for sales. ShintaiDK is spot on, there: they could, and still can, barely give them away.
A very small number of users will be doing things that require lots of RAM per socket at modest costs, and/or have work limited primarily by threads*clocks, and have them under load often (so idle won't make much difference). For everyone else, upfront CPU costs are a very small part of their server buying decisions, and not a big part of the total cost. Power almost always matters, and sometimes performance does. Spending another $3k per server, FI, if you need a beefier CPU, is not much, when any server taking a full-depth U or more is going to cost you in the tens of thousands just have to have in the rack and powered for 3-5 years.
Cerb
Elite Member
- Aug 26, 2000
- 17,484
- 33
- 86
Only at unusual clock speeds. At those they are typically sold at, they cannot. The advantage AMD has against an i3, though, is that CPU performance is only an issue for folks that post here, wanting maximized performance on a budget. Enough software is sufficiently well multithreaded now that 2(C/M)4T is tangibly superior to 1(C/M)2T or 2(C/M)2T, even to regular users. OEMs realize this, and are putting Kaveris where they had fast Pentiums and slow i3s, even in notebooks.
But, at regular prices and performance levels, Intel can easily sell an i5 for more than an FX-8xxx chip, because it's better, all things considered. Ditto for i7. If they really could compete, they would, and either AMD's prices would be higher, or Intel's lower.
AMD has to live with a lagging process. That is not a suitable excuse for performance metrics of their processors. AMD failed against their own prior CPUs. Meanwhile, they have been competitive, even with inferior processes, here in 2014, with Jaguar and Kaveri (Intel isn't shivering in fear, or anything, but OEMs clearly like the chips). They've done what they can to make lemonade from lemons, but take off your fanboy glasses: they made lemons. I was hoping for BD to be better, myself, but it wasn't. Even as it has improved, it has remained too little too late.
With restricted R&D, and ignoring some markets at first, the very good Bobcat was a fair success, but not what it could have been, and the same with Jaguar. If they can lead, rather than follow, markets, with whatever they have coming out, and they are making sensible designs (speed demons need to stay back in the 90s!), they may have some hope. IMO, they really need to follow Qualcomm, and work directly with companies that may implement their products, to both come up with compelling products, and be able to sell a predictably good amount of them.
Sure they are. The E3 is a desktop socket. If you're buying a server or workstation not using the desktop socket, you won't get to buy desktop CPUs. If you're buying a cheap one using desktop CPUs, you're stuck with a 32GB RAM limit, and inferior ECC.
The plan is servers, targeting data centers.
See: http://www.amd.com/en-us/products/server/opteron-a-series/a1100
AtenRa
Lifer
- Feb 2, 2009
- 14,003
- 3,362
- 136
SteamRoller Modules can easily compete in MT loads even on Clock to Clock against Haswell 2C 4T CPUs.
If i5 Haswell at 22nm FF wasn't better than two year old 32nm FX, then there would be a serious problem for Intel.
And still there are workloads that FX is still faster than any i5 today.
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes + WCL Discussion Threads
- Started by Tigerick
- Replies: 24K
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 23K
-
-
Facebook
Twitter