PCTuning fórum

yuri.cs píše:Zatajene zmeny u PD:
-dekodovani 2*double instrukce/cyklus

Ty double instrukce za cyklus umí zřejmě i Bulldozer.

je zajímavé, že Bulldozer je v AVX rychlejší jak PD. (zřejmě design flaw)

\\ jo máš recht, už jsem to našel. Bulldozer umí jen (1-1-1-1) a (2-1-1)

del42sa píše:Ty double instrukce za cyklus umí zřejmě i Bulldozer.

Dle jeho mereni ne.

tenhle závěr je opravdu "povedený":

Therefore, there is no advantage in using 256-bit instructions when the bottleneck is
execution unit throughput or instruction decoding. The poor throughput of 256-bit stores
makes it a disadvantage to use 256-bit registers on the Piledriver.

While Intel processors have a large penalty for mixing 256-bit AVX instructions with non-
AVX XMM instructions due to a mode switch (see page 130), there is no such penalty and
apparently no mode switch on the AMD Bulldozer and Piledriver.

256-bit memory writes

256-bit memory write operations are exceptionally slow on the Piledriver. So slow, indeed,
that it is better to not use 256-bit registers at all on the Piledriver.

BD analyze by Agner
PD/JG analyze by Agner

nějaké plány do budoucna:

http://www.amd.com/us/press-releases/Pa ... eleases%29

“Hierofalcon” CPU SoC

64-bit ARM-based platform
up to eight ARM Cortex™-A57 CPUs expected to run up to 2.0 GHz
two 64-bit DDR3/4 channels with error correction code (ECC)
10 Gb KR Ethernet and PCI-Express Gen 3
enhanced security with support for ARM TrustZone® technology and a dedicated cryptographic security co-processor
sampling in the second quarter of 2014 with production in the second half of the year.

“Bald Eagle” APU/CPU

R-series; x86-based embedded processor available as both an APU and CPU featuring up to four new “Steamroller” CPU cores within a 35W TDP.
Graphics Core Next GPU architecture and HSA enhancements
configurable TDP
first half of 2014.

“Steppe Eagle” APU SoC

G-Series APU SoC platform
enhanced “Jaguar” CPU core architecture and AMD Graphics Core Next GPU
increased performance-per-watt both at a lower TDP
above 2 GHz.
first half of 2014.

“Adelaar” discrete GPU

Radeon GPU based on Graphics Core Next architecture
2 GB of graphics memory.
multi-display support
DirectX® 11.1, OpenGL 4.2 and both Windows and Linux.
first half of 2014 with seven years of planned supply availability as an MCM, mobile PCI express module (MXM) and standard PC graphics card2.

dále desktop Kaveri by mělo přijít koncem roku 2013 dříve než notebookové modely. (přesně opačný scénář než s Llano a Triniti/Richland)

http://translate.google.com/translate?s ... er-2013%2F

Divny je ten “enhanced Jaguar”. Neni to jenom pridane richlandovske Turbo, ale taky to neni zadne nove Jadro. Za to Kaveri ono richlandovske Turbo ma - configurable TDP.

Maly ARM bude mit oba radice DDR3 i DDR4. Zajimave.

http://investorshub.advfn.com/boards/re ... d=91844647

potom Flankerovi ty jeho zdroje kecaj nebo to není pravda. Bohužel zatím oznámené roadmapy podporují tu první teorii

Seronx tvrdí, že SR bude mít takt okolo 3 GHz na 28nm

Steamroller doesn't maintain the high frequency engine. The high frequency from Bulldozer was largely needed because of the VMT front-ends. Steamroller has SMT front-ends which allows both cores/contexts to fetch/dispatch/retire. More execution units with an improved core design will always trump a high frequency with a poor core design. Also, the Bulldozer/CMT model concept is about pushing as much cores(logic+cache) into one area while improving perf/watt.

A10-5757M(RL-A1) vs A10-6757M(KV-B0) ((KV-B0 might not be production stepping))

CPU:
2.5 GHz vs 1.8 GHz
SCT 3.5 GHz vs 2.3 GHz

CPU:
FP:
RL-A1: 2.5 * 2 * 8 => 40 DP GFlops/80 SP GFlops
KV-B0: 1.8 * 2 * 16 => 57.6 DP GFlops/115.2 SP GFlops
Int:
RL-A1: 2.5 * 4 * 4 => 40 AGLU Ops
KV-B0: 1.8 * 4 * 8 => 57.6 AGLU Ops
--
FPSC:
RL-A1: 3.5 * 1 * 8 => 28 DP GFlops/56 SP GFlops
KV-B0: 2.3 * 1 * 16 => 36.8 DP GFlops/73.6 SP GFlops
IntSC:
RL-A1: 3.5 * 2 * 4 => 28 AGLU Ops
KV-B0: 2.3 * 2 * 8 => 36.8 AGLU Ops

Effective theoretical comparison. Which gives KV-B0 a 36% advantage over RL-A1 with the same wattage. If you get into detail, more execution units can actually give you lower wattage if not all units are active.

We don't know if Kaveri is Sea Islands that is included in Volcanic Islands nomenclature or if Kaveri is Volcanic Islands. So, the comparison of 6CU VLIW4 to an unknown GCN arch version of 8CUs is up for debate.
Last edited by Seronx; Today at 12:08 PM.

Seron blábolí a je mimo realitu...

http://diit.cz/clanek/amd-bald-eagle-hi ... le-adelaar

AMD představila své plány pro příští rok zaměřené na dění v embedded segmentu. Společně s nimi se ale dozvídáme také několik informací, které se můžou týkat i desktopu…