All of the 15h parts are limited to 2400MHz max. by the design. The "new" memory controller introduced in Steamroller is the worst piece of hack job I´ve ever seen. It is some off-the-shelf solution "integrated" (whisky tangoed) to the design by using some duct tape and hot snot. It´s functionality relies fully on vast amount of C code (AGESA).
The memory controller in Excavator designs is basically the same thing, the only major difference being the memory types it supports (DDR3/GDDR5 combo vs. DDR3/DDR4 combo).
The "quality" of the memory controller matches pretty closely the rest of the Steamroller design: nothing works (together). Steamroller is like the last ditch design with all of the scrap thrown into a same can and slightly stirred. Each of the different parts work well separately, but when combined together the outcome is just absolute junk. Normally this type of bugs and errata can be found in early evaluation silicon, not in something that gets released to the market.
I certainly hope (pray) that anyone who worked on Steamroller have had nothing to with the upcoming 17h family (or any other AMD product for that matter) :'(
Those prices are a little hard to swallow. $644 for a 14" one module A4-5150M??? I got a 13" Dell with Core i5 6200U for $499. Why would anyone get those HP laptops?
So perhaps another reason GDDR5 was included was because the DDR3 was so bad?
And AMD crossed their fingers one day they could use it.
Reason: it seems very mysterious to me AMD would include the option for GDDR5 as system RAM for an APU with only 512sp. This, unless, there was something wrong with the DDR3.
The HD7750 is a 512 shader GPU, and it needs GDDR5 to run at full performance. http://ht4u.net/reviews/2012/msi_ra...39.php&usg=ALkJrhi1G4TxkhzXnvN1ZfRJ3KdXukbpQQ That was only with DDR3-1600, much slower than Kaveri supports, but it also didn't have to share it with a CPU.
Do you have any estimates on how much performance Steamroller have lost due to this?
If we expect that a proper & properly implemented memory controller would reach similar frequencies (2666MHz+) that Intel IMCs do, then the performance penalty for the high-end Steamroller APUs in 3D is around 15%.
The core µarch is what it is and a proper memory controller would not make it any better in general. There are some cases where a proper IMC would improve the performance quite significantly (e.g data de/compression, SuperPI) which is latency & bandwidth critical. The memory latency on Steamroller is over 30% higher than on the previous generation at the same settings (MEMCLK and timings).
Excavators (CZ, BR) use almost identical memory controller IP as Steamroller does. The only difference is that Excavator has DDR4 Phy in place of GDDR5 Phy found in Steamroller. The DDR3 Phy is identical.
It is bat*hit crazy that the new memory controller has that much HIGHER latency. At 2133MHz CL12-12-12-2T (3400MHz CPU, 1200MHz NCLK) AIDA64 (5.60.3755) measures 112.4ns latency for the DRAM on Carrizo. Richland running with identical settings (excl. 1500MHz NCLK) gets 70.8ns.
Steamroller, which was the first design to implement the completely new memory controller has over 30% higher latency than Piledriver based Richland. For some reason the latency has increased even further in Excavator. While AIDA64 isn´t exactly the optimal method to measure the DRAM latency, based on real world applications (e.g. LZMA, RAR de/compression) it is clear that the memory latency is infact significantly higher than on the older memory controller. Quite an achievement, since the old memory controller was introduced in K8. Until it´s decommissioning it received only minor upgrades.
It's gotta be related to how they implemented some of the hsa technologies.Could it be possible that the reason why the Steamroller and Excavators IMCs latency is higher than previous generations is because they're IGP-first and throughput optimized? After all, GDDR5 supposedly achieve much higher bandwidth than standard DDR3 at the cost of higher latency, and Steamroller was going to support GDDR5. It makes some sense.
Also, anyone remember when Apple considered the AMD processors for the macbook air? Any chance this could ever happen again? All AMD needs to do to get back in the game is to cut Apple a sweetheart deal.
Also, anyone remember when Apple considered the AMD processors for the macbook air? Any chance this could ever happen again? All AMD needs to do to get back in the game is to cut Apple a sweetheart deal.
Was real at Llano era, reappeared again as a rumor with Raven Ridge(zen APU).
Raven Ridge is more like a hope, as there is no such product yet.
AMDs corporate vice president of worldwide marketing affirmed that Zen based APUs will be available 18 months from now
Q3 2017 D:
And most definitely no HBM around even at that point :'(
HBM on an APU is pretty much a fantasy
Do you have a source? Or good feeling about that prediction?Q3 2017 D:
And most definitely no HBM around even at that point :'(
Do you have a source? Or good feeling about that prediction?
Btw the hp elitebooks with carrizo are pretty discounted right now.
Heres a link (took this off google, no affiliate crap) - http://store.hp.com/us/en/pdp/cto-d...=L9Z80AV_1&ci_sku=L9Z80AV_1&ci_gpa=pla&ci_kw=
A hope? WAT?
Is real, just don't exist on official public roadmaps. Unlikely as a hope is that Raven Ridge will have integrated HBM.
RR is Time-framed to 1H2017, probably being the second Zen product after the new FXs. Zen CPU and Polaris GPU.
You can get a laptop with an Nvidia 860 for that much
Yes you could but this was a thread about Carrizo ()