Jaguar is one of the best low powered architectures I've seen in a while. I've had plenty of success with jaguar (5350/5370) . I'd like to see a 8 core, 3ghz version released to consumers.
Any love with the K10 architecture today? An excellent lower-cost alternative to Bulldozer that's not clustered into threads at all.
No. The K10 derived Llano was clearly slower than the Piledriver based Trinity.
The K10 suffers from low single-thread performance today if not overclocked, but if you can find one a Black Edition one that runs over 4 GHz, then it's a superior choice than Bulldozer I think. I successfully found one Phenom II X2 570 that can run up to 4.2 GHz max on stock voltage. Disable one core before, it can go up to 4.4 GHz max. Sweet... Average is 3.9 GHz based on multiple Phenom 570 CPUs I've bought. Now only $19 each shipped from China. AMD processors maintain their resale value better than Intel, so the 570s won't crash down to $5 mysteriously at anytime soon.I have a couple K10 Athlon X4's and Llano systems running today. They're starting to show their age
Successor:
22FDX Node
1 Excavator-LPH22-Module // Two cores - No XOP/FMA4/CVT16 -- Full GPR/FP Renaming -- 5 GHz peak when no other cores are in peak use. -- In between 15h & 16h overhaul.
1 Catamount-ULP22-CU // Four cores - AVX2/FMA3 -- "Zen-Lite"
2 Vega-LP22 CUs // Equiv to ~256 or ~512 ALUs from 28nm generation.
6-8 MB L3 (Shared through the data fabric and buffers the single 64-bit DDR4(+ECC) 3.6 GHz(3.2 GHz))
Data Fabric includes an upgrade for AMD's HWP(Intel SpeedShift competition) support.
"Octo-core" APU => 2 BD Cores + 4 JG cores + 2 VG cores.
The previous post has the Seronx Certificate as "almost factual."Could you please label speculation as such? People often mistake it for actual facts.
It's lacking SMT which is why it was discontinued and hasn't been updated with modern instruction. Without SMT (or CMT) it's too hard to get good energy efficiency (at least at 3+ ghz). I can only imagine any non-zen architecture being continued if development is limited to a shoestring budget.No. The K10 derived Llano was clearly slower than the Piledriver based Trinity.
If they updated instructions to Excavator's set and also added some crude multithreading, then Puma would not just hands down beat Excavator in efficiency (at ~1.5ghz), but (using 22nm FDSOI) also Zen.Jaguar is one of the best low powered architectures I've seen in a while. I've had plenty of success with jaguar (5350/5370) .
So... BD will be still be alive but only in lower powered parts? Sounds weird... And BD design... Could fit better on something like ARM..The previous post has the Seronx Certificate as "almost factual."
The FDSOI designs will be a prelude the SSRW FinFET designs. So, anything new and improved in this non-disclosed product that will be taking up the BR/SR Refresh. Will happen to also appear in 7nm LP refreshes of 14nm products.
The K10 suffers from low single-thread performance today if not overclocked, but if you can find one a Black Edition one that runs over 4 GHz, then it's a superior choice than Bulldozer I think.
The K10 suffers from low single-thread performance today if not overclocked, but if you can find one a Black Edition one that runs over 4 GHz, then it's a superior choice than Bulldozer I think. I successfully found one Phenom II X2 570 that can run up to 4.2 GHz max on stock voltage. Disable one core before, it can go up to 4.4 GHz max. Sweet... Average is 3.9 GHz based on multiple Phenom 570 CPUs I've bought. Now only $19 each shipped from China. AMD processors maintain their resale value better than Intel, so the 570s won't crash down to $5 mysteriously at anytime soon.
Meh. I've had K10.5 chips before in the 4 GHz range and my 4.7 GHz Steamroller was generally preferable. K10/10.5 suffers from poor SIMD support, among other things.
Thuban was cool for awhile, but not anymore. It isn't 2010.
It's lacking SMT which is why it was discontinued and hasn't been updated with modern instruction. Without SMT (or CMT) it's too hard to get good energy efficiency (at least at 3+ ghz). I can only imagine any non-zen architecture being continued if development is limited to a shoestring budget.
The cat cores may be the closest thing to the old K8/K10's. But they are way more area efficient and clock much lower than K10's.
I really like Seronx's idea or prediction of a fused cat-excavator cluster. Excavator for decent sparse thread and puma for the high area and energy efficiency and number of threads. The OS can easily taskset lower priority threads to lower clocking cores.
An excavator pair plus puma quad (2+4) budget APU for lower power profiles would be a cheap test project of what can be done on FDSOI and body voltage biasing for both the CPU side and GPU side.
On 28nm ballpark 9 mm2 per XV core (w/o L2) and 9 mm2 for the XV module's shared L2. This would be ~27 mm2 for the core part of the Stoney module. Puma quad would be under 4x3.1mm^2~13 mm2, and these cores would be ultra efficient on the upgraded 22nm FDSOI, especially under 1.5ghz. It would be worth adding a little area and new instructions to the Puma cores so the don't lack intructions that Excavator has. Also a shared L2 cache may add ~10mm2. This may grow the area but it still would be not much more than 50mm^2 for a total 6 cores. (On 22nm this would shrink down to under 40? wild guess a 120 nm2 APU if you add 40 for uncore and ~45 for gpu). Derivative products could include 2+2 APU and 2+0 and 0+4 APUs.
If such a test is successful follow on projects could be a similar Zen-puma cluster, or mobile 14nm FDSOI super low power GPUs.
What benefit would this theoretical device give AMD over their current planned lineup? Right now they already have Zen, Polaris and Vega IP on 14nm FinFET, and could easily use these to offer a 2 core APU to address those same market niches. Whereas your proposal requires a hell of a lot of CPU design work to integrate XV and Puma into a single cluster in an efficient fashion, along with porting AMD's entire IP stack to an entirely new process with more expensive wafers.
Initially, i thought that bulldozer would be viable as a jaguar replacement.
But i think that an updated Jaguar architecture (with knowledge learned from the piledriver and on)without SMT but with better simd support (wider paths ,execution units and less use of microcoded instructions) and the Zen architecture will be what is happening.
2 architectures. Jaguar and Zen.
No SMT, means there is some time that there are stalls during execution and that is good for power dissipation, no execution is lower consumption. Less dark silicon required. I think that Jaguar is small and cheap enough to just add more cores, no need to redesign it with SMT, also, the fabric to connect all 4 core modules already exists. But i do wonder how much effort it takes to take existing jaguar to a smaller process and how much in clockspeed it would gain. Jaguar is there when it needs to be as inexpensive as possible. When performance is required, Zen is there.
Power consumption is another biggest concern with K10. To get 4.5GHz quad-core, it requires 1.7 V stable.
It's between 1.6V to 1.7V on average with Phenom X4 at 4.5GHz, but the K10s do really require more power consumption in order to match Bulldozer's equivalent single-thread performance score. My Phenom X2 (2 cores) runs at 1.4V at only 3.9 GHz max.I'd be interested in seeing wattage vs frequency plotted over 4ghz. ~160W?? or almost as much as FX piledriver octacores consume at 4.5ghz full load?
We get it. You hate AMD. You joined up just to bash Ryzen/Polari.
Those are some really good points.
I think they should keep the microcoded instructions to conserve area because the small area of the cat cores would be the main reason that the cat cores may not go extinct yet and outlive dozers.
.
Well, avx might come in handy and wider paths and instructions that need less cycles to complete would give jaguar an enormous boost. It seems to become more and more used. low amount of cycles for completion might help a lot. I do not see much in adding SMT.
agreed, and zen's boost frequencies would have a much greater range than jaguar/puma even at that TDPIMHO at the point of adding AVX2 to Jaguar, you may as well use the Zen core, perhaps minus the L3 cache. Zen is already very efficient, it should be more then able to scale down to 25W for 4C/8T with lower clocks.
agreed, and zen's boost frequencies would have a much greater range than jaguar/puma even at that TDP