OCAU News - Wiki - QuickLinks - Pix - Sponsors
|9th February 2017, 5:10 PM||#1|
Join Date: Jun 2002
Location: Griffin , Brisbane
AM4 Platform Preview + A12-9800 APU Review
So With a painful 3-4 weeks to go until Ryzen officially launches, what better time to have an early look at the upcoming AM4 platform in some detail, and at the same time, the 'Silently' launched Bristol Ridge APU Line (To be ' Officially' launched into the channel some time after Ryzen)..
First, a Big thanks to The_Stilt for his help in ironing out some issues during testing, and enabling me to simulate lower TDPs. Despite being tied up with Ryzen stuff he still found the time to help out.
AM4 - Finally A Unified platform.
Some 5-6 years go now Both AMD and Intel essentially drove a wedge down the middle of there respective Desktop platforms.. Intel, pushed there IGP-less Enthusiast line into an expensive, server derived platform, AMD, after introducing the FM1 platform for the first APU's, kept the (again, IGPless) AM3+ platform hanging around for the same sort of purpose.
Unfortunately, for both companies this has caused contention, particularly in recent years as increased core counts climb in demand, this platform segregation can be quite prohibitive, but is particularly bad for AMD, as the FM2+ line tops out at basically core i3 level on the CPU front.
Well for AMD, this all changes now.. AM4 is to be AMD's only desktop platform going forward, and will span everything from low power, small form factor APU's, through to the much anticipated, 95w 8 core, 16 thread Ryzen CPU's. You can plug in a Bristol Ridge APU one day, then throw in a video card, and turn it into a 8c/16th number crunching /gaming monster the next.
AM4 launched silently into the hands of OEM's alongside a range of new 7th generation Bristol Ridge APUs late last year. The reasons for going with such a release strategy is really only really known to AMD, but one can theorise a little..
Firstly, (and likely the main driving factor) is that Ryzen was some way off at that time yet , and Ryzen is really the "Killer app" here. The only thing you can plug into an AM4 socket right now, is the Bristol Ridge APUs - Nothing wrong with that, but so close to Zen launch, It's a little bit awkward, as they simply do not have the features nor performance to show off the new enthusiast Tier (X370) Chipset in particular.
Secondly, it's really only the bottom end A320 Chipset that has actually launched. This does not support overclocking, nor anywhere near the full IO capability
So what's new with AM4? Well, almost Everything - The socket itself is significantly different, All new chipset's, and importantly, it's AMD's first DDR4 desktop platform.
The Motherboard I have here is an Entry level ASUS A320M-C, using as the name would imply, the A320 Chipset. which is a competitor to Intel's H110. so it's down the bottom end of the pack.
At a glance it looks no different, but both pin pitch, and HSF retention holes are slightly different.. Note however the clip geometry is unchanged, so you CAN use old socket AM2 onwards coolers if they have clip retention.
This is the first time in 13 years that we've seen a Pin Size + Pitch change on AMD processors.. We saw a package size change with the little AM1 APUs, but they carried the same pitch.. And the difference is significant. No less than 1331 pins on the same package size.. and that's with space in the middle (Which is interesting in itself - The package can now fit a maximum of 1500 pins if desired. Who knows what the future may hold here, but the potential is there.
So does this mean more bent pins?...
Well, I don't really feel like testing this out but.. likely NO. Reason being, as you can see whilst the pin diameter is (I think) slightly smaller, and there's more pins to bend. The pins are also significantly shorter making them more resistant.:
being a smaller pitch however, it will take less distortion to cause fitment issues, so in the end they're probably similar in 'vulnerability' to mishandling.
New vs old - an Evolution of Packages.
AM4 brings a full spread of new chipsets with it.. touting lower power consumption, higher UMI bandwidth (unconfirmed), Gen2 USB3.1, nVME. Only the bottom two tier's have been released so far - X370 is being launched alongside Zen.
Also launching will be SFF (Small form factor) 'Chipsets' , these will not really be chipsets at all, but rather be the name given to feature-enabled SoC only functionality, much like how Bristol ridge operates currently in laptops, i.e no seperate chipset at all. continue reading for a bit of an explanation on how this works..
Physically the new chipsets are now built on TSMC's 28nm process, and have a nice big integrated heat spreader, which Are still having small, almost pointless heatsink's attached to them. Infact, I removed the heatsink and ran the system up and found the chipset only reached a fairly safe 85c, so perhaps in non-user accessible applications this chipset could run heatsink-less.
One thing that can be somewhat confusing is that with AM4, some chipset functionality normally dedicated to the 'south-bridge' is also available from the CPU - All APU's/CPU's are true SoCs (System on a chip) that is to say, they do not need any additional chips in order to operate, and interface with typical devices. . So the total chipset functionality is achieved by adding the SoC functionality to the chipsets, so will vary depending what you put in the socket. The below diagram illustrates this
As you can see from the Feature set table at top of this section, it's quite a step up towards Intel's chipset, surpassing it in regards to standard features available, but still lacking in the general purpose PCI-e lane department on the higher end models .. This is somewhat of a compromise though, as the less features available standard, the more GP lanes MB manufacturers will have to use to supply this functionality, E.g USB 3.1 G2, which only AMD's new chipset natively supplies. Whilst not illustrated here, it's worth mentioning that Intel's B150 and above's GP (general purpose) lanes operate in PCIe 3.0 mode, whereas the B350 and Z370 are still 2.0.
I've also done up an I/O lane equivalent diagram to make comparisons easier. This is based on Intel's IO lane diagram from their Datasheets. Intel's shared i/o lane approach can be difficult to compare without this style of diagram
Focusing on the budget models, AMD A320 certainly has the upper hand vs H110. The Chipset-less models sacrifce a lot for simplicity, but still support Nvme, and USB3 directly from the SoC (i.e the APU/CPU).
The Motherboard: ASUS A320M-C
This is clearly an entry level board, as made obvious by it's A320 Chipset.
Currently this board is only available to OEM's, but has popped up on an European website selling CPU MB Bundles.
Which is handy as Looking at the prices on this website one can get an idea of the approximate value of this combo, to give some perspective:
i3-6100 + H110M-A: $189Euro
A10-9800 + A320M-C: $199Euro
A10-7870K+ A88XM-A: $211Euro
It's still difficult to isolate the induvidual cost here. But gives a good idea. With higher volumes and more time on the market one would expect to fall pretty much in line with the i3 combination.
It's a bit of an odd beast, with PCI slot, and quite a large footprint, whilst still mATX , it has a lot of empty space. I have a suspicion this is a prototype physical layout, i.e not optimized for space or aesthetics.
Typical of entry level boards, we have 4 SATA, only 2 RAM slots, but the inclusion of an M.2 port is a bonus.
on the rear, HDMI output , and all 4 USB 3.0 ports.
VRMs - Improved for AM4.
The updated VRM spec is one of the interesting parts of the AM4 platform.
Due to the fact boards such as this will support up to 8core, 14nm CPU's (albeit likely limited to 65w), the Current requirements will still be very high in comparison to Intel (Since they only offer 4C parts on socket 1151) . The reason why is because (due to it being a low clocked 8 core part) Core voltage will be low, therefore,even @ 65 TDP, current will high. I'll give a somewhat made up example:
Please note these figures are pulled out of the air for illustrative purposes!
4 cores, 4ghz @ 1.42v 95w:
95w / 1.42v = 66.9A
8 core, ~3Ghz @ 1.05v
65w / 1.05v = 61.9A
Therefore despite the 2nd example having a far lower TDP, it has nearly as high current draw.
So compared to FM2, and Intel's Socket 1151, it's clear VRMs will have to be beefier, and the challenge will be how to reach the best compromise as to not eat into BOM costs for entry level motherboards, and more importantly for SFF/ITX form-factors, not eat into precious board space. Don't be surprised to see motherboards only able to support 65w Ryzen , perhaps even only up to 6 core variants.
My guess is that SFF AM4 boards will also have limitations on what you can run- Likely limited to 65w TDP Octocore's.
The other factor here is much tighter Regulation requirements for AM4. maximum droop specification will be much tighter than predecessors, this also means higher quality FET's and potentially more board space to route high current traces.
Ultimatly this situation has Pro's and Con's. On the plus side it means Budget level boards will have reasonably high quality VRM's - they're basically forced into being halfway decent, not just because of the current requirement,s but the increased regulation tolerances. The downside however is, I speculate this platform won't under cut Intel price wise like has been the case in the past. This is easy enough for AMD to offset with the APU and CPU pricing though. So I guess we'll see how this pans out.. It will ultimatly be up to AMD to set platform costs based on CPU pricng.
Back to this Board, and this already seems to bear out in practice - It's clear the VRM is strong compared to both the ASUS equivilant H110 and A68 Board's . But it is still most likely limited to 65w,
To start with This is a 6 (4+2) phase design,( controlled by a mysterious ASUS digi controller).
Somewhat strangely, only the two CPU-NB phases are heat-sinked (removed in the above picture) . I suspect this is a result of increased NB power requirement on AM4, but do not have access to Datasheets to confirm.
the FET configuration is a pretty hefty (For an entry level class board) A single 4C06B High side, but Two 4C09B's on the low side of each phase. Compared to both the other boards here which both have a lower phase count (5 in total) and only a single low side FET per phase,
so whilst this is a pretty weak VRM compared to high end boards, looking at it in perspective, it's impressive. Will be interesting to see the implemntations on retail motherboards.
Platform test - Enter A12-9800
With all the hype surrounding Zen, it's a little hard to get particuarly enthused about the release (if you could call it a release) of these APU's, but nonethless, they're an interesting last hoorah for the 'construction cores' (i.e bulldozer derived). Finally bringing the Excavator architecture to the desktop. On top of that they bring some 'behind the scenes' technology on board that will carry through into Ryzen and Zen derivites.
Bristol Ridge vs Kaveri/Godvari:
In a nutshell, Bristol Ridge brings the following to the table over Godvari:
* Excavator CPU cores: An evolution of Steamroller. Larger (128KB L1D cache, Larger BTB, and general core enhancements which bump IPC up by approximatl 10%. Which in isolation, is certainly significant. These cores originally launched in Carrizo Laptop processors in 2015, so they're not new.
* Tonga based GPU - This GPU architecture was originally seen in the R9-285, 380 and 380X and FURY desktop cards. Generally more efficient, and importantly for APU's that suffer the unfortunate situation of having no dedicated memory, features more advanced Delta color compression - which according to AMD reduces bandwidth requirments by around 10-15%.. To turn that around, it in theory should provide bandwidth limited performance of the older architecture with 10% less physical bandwidth.. so With DDR4-2400 , we should be seeing equivilent of Kaveri/Godvari @ DDR3-2666 - a 23% increase in effective bandwidth, and should make up the bulk gains in gaming.
* Intergrated Southbridge functionality (see platform diagrams above)
* Globafoundries process - Carrizo and Bristol ridge are both manufactured on an an enhanced 28nm Bulk /Planar process
* Advanced power management - Shadow P-States, Reliability tracking, means They can operate closer to true stability limit.
It does comprimse on a few things along the way though:
* L2 Cache drops to 1MB per Module. Down from 2MB, which hurts in certain applications.
* High density Libraries and GPU orientated Metal stack - Being a mobile-orientated part, this results in less die space, and vastly improved perf/watt at low frequencies - but quickly puts a lid on high frequency scaling.. It's ultimatly the reason Carrizo never launched on the desktop. AMD have pulled a bit of a rabbit out of a hat on this one though with Bristol ridge, with frequencies above that of the fastest 65w Kaveri chip - which it will be up against in the following tests.
This illustrates the effect of the above mentioned High density libraries / Metal stack. The core has shrunk in size from its predecessor, without an actual Node decrease.
As a bit of a side note, in regard to Zen, it's currently unknown if they have continued with this approach or not. But given the frequencies achieved with Bristol Ridge it is possible.
So I'm pitting this newest APU against a couple of Relevant contenders.
Intel Core i3 6100 - Now technically this how now been superseded by the Core i3 7100, with a speed bump , however it also recieved a price hike, so the 6100 remains at a similar price point to AMD's 7860K
AMD A10 - 7860K - Conveniently, this model is the highest performing 65W APU, which is important, since Bristol ridge is 65w TDP only.
As you can see from the chart, its clock speeds are lower than the new Bristol Ridge flagship, so one would expect a decent boost here when combined with the IPC uplift of the A12's Excavator cores.
Something else that's new to the APU lineup is the presence of 35w Energy efficent models. it's been a long time since AMD have offered 35w in a desktop package. Not since the old Athlon64 EE models in fact. Popular at the time with HTPC's and the like.
Without being able to actually try one of these, the next best thing is to simulate it by limiting TDP. A big thanks to The_stilt for making this possible.
Please note though this SKU's P states are NOT optimized for 35w. So an actual 9800E would almost certainly perform slightly better .
Time to throw some benchmarks at this setup
MB: ASUS H110M-A
RAM: G.skill Ripjaws DDR4 2666 @2133
HDD: Sandisk Ultra SATA 256gb
MB: ASUS A68HM-PLUS
RAM: G.skill Ripjaws DDR3 2133 @2133
HDD: Sandisk Ultra SATA 256gb
MB: ASUS A320M-C
RAM: Generic DDR4-2400 (single rank)@2400Mhz
HDD: Sandisk Ultra SATA 256gb
Operating System: Windows 10 home 64 bit
Clockspeed Analysis (65w)
Throughout the demanding POV ray Run , all cores bounce between the 3.8Ghz base clock, and 4Ghz.
Unfortunatly I don't have the same graphed data for 35w TDP, but typically clock speeds will bounce between 3 and3.4ghz , and this is what I mean by not optimized for such a TDP. idealy there would be more P states to choose from in this range, which would improve performance.
And straight away, compared to its predecessor, it does not disappoint.. setting new levels of performance for AMD in a 65w TDP.. Unfortunately it's competitor is the Skylake i3, and it does not look nearly as good. Still, it's putting up a good fight in several of the threaded App's here
With Fritz we see the only regression in performance. It's rare, but as with any compromise made to improve overall performance, you'll have an outlier like this.
The single threaded Deficit has shrunk considerably. owing to a combination of IPC increase, and a rather aggressive 4.2Ghz turbo clock. It comes at a cost though - ST power consumption is quite high.
Lame encoding show's an almost perfect bridging of the gap between Kaveri and Skylake here. It will be interesting to see how Zen holds up in these more legacy workloads.
GPU Performance - 'Preliminary results'
As mentioned above, the newer Tonga based GPU architecture's biggest benefit to this IGP is better use of bandwidth, but on top of that, Bristol Ridge boosts clockspeeds to an impressive 1106Mhz, so @ DDR4-2400 we should be seeing at least the 23% extra effective BW turned into performance, unfortunatly though there some roadblocks in testing this to it's full potential...
unfortunatly there's a couple of issues problem with this system which hinder GPU performance. i'm still in the process of diagnosing, but at this stage appears to be an early BIOS issue.
First of all, GeAPM was initially enabled (this is a setting that forcibly throttles CPU cores to minimum power state when the GPU is active). I was able to disable this with some help, but was then left with another issue.. the DDR4 Kit supplied with this combo is Single Rank, which in regards to GPU performance at least, is the equivilent of a normal Dual rank DDR4-2133
Ultimately I will have to revisit IGP performance in more detail, once I have retail board(s) in hand, with more up to date BIOS' and Memory support So take these results as somewhat preliminary.
In the meantime bear this in mind, that these results are potentially lower than they should be. Particularly 3Dmark as I know it is capable of higher based on results from some other pre-production boards that have floated around the web.
First up - 3Dmark
3Dmark is an example where I know performance is slightly off the mark. Given the DDR4 issue .
Resident evil 6's integrated benchmark on the other hand showed good gains.. albeit at barely playable frame rates
Dirt 3, you can see a continuing to walk away from its predecessor as detail settings go up which interesting., and showing impressive gains in this particular game. almost twice as fast as the i3, and is the difference between borderline unplayable, and for casual gaming at least, 'smooth' game-play.
testing @35w again shows performance nipping at the heals of its predecessor
Using my standard method for lower power CPUs here.. A pico PSU powers the platform, and DC input voltage and current is measured real time. This eliminates the inefficiencies and power factor errors of measuring AC input power.
one of the areas AMD platforms have slipped behind in is idle power. Both chipset and CPU/APU power management had previously lagged behind. Looking at these results suggests that is behind AMD now. Idle usage is a clear match for Intel's systems. The SoC only x300 platform will likely shave even more off of this.
this is where things get a bit ugly. aggressive turbo modes eat every single bit of that 65w tdp up, unlike the i3, which with only cpu load, and with absence of any turbo only chews what it needs, well below its rated max tdp. Similar plays out under a single threaded load. With a max turbo of 4.2Ghz
The load power of the 9800 @35w limit says it all about the vastly better performance/watt achieved when running the cores at lower clocks.
IGP Gaming workloads
and lastly , IGP gaming workload power paints a far better picture of Bristol ridge at 65w, even better at 35w.
Despite a huge deficit in process tech.
Overclocking, Undervolting? - Not on this board!
Normally, i'd be all over this, but unfortunately it's certainly not an option on this entry level board.. Whilst OCing would be kind of pointless anyway on this SKU, it would be interesting to see how much further Perf/watt could be pushed. I'm not entirely sure this plays out with the new Power management on these SKU's. Ditto ryzen, but all will be revealed soon enough.
Wrapping it up.. for now.
Hopefully this gives a bit of an insight to the upcoming AM4 ecosystem.
As for the Bristol ridge APU, As a 65w TDP product, it lags particularly in performance/watt on the CPU front, and frankly this is expected being primarily a mobile-centric design. but for anyone doing a bit of casual light gaming, These APU's continue to have appeal. Running Bristol ridge @ a 35w limit was a real eye opener. 'adequete' CPU performance (only slightly lagging behind the 65w Kaveri predecessor) is met with somewhat outstanding GPU performance /watt (despite the platform issues ). on the CPU front at least, a true E series SKU would likley perform better than these preview results, since its P states are optimized for 35w.
I guess one can only hope AMD decide to make the official E series APU's available, as such a TDP is far more suited to this architecture, at least when limited to just two modules like this. Still, as an all rounder, the standard models certainly are an improvement over Kaveri in every respect.
As much as Ryzen will distract me and many others for some time, I look forward to revisiting these with a more mature platform, running Dual rank Memory and in particular the little SFF x300/A300 chipset.
Last edited by mAJORD; 20th February 2017 at 9:43 PM.
|Join OCAU to remove this ad!|
Sign up for a free OCAU account and this ad will go away!