Informazioni e notizie sui processori per calcolatori
 

AMD e' al lavoro, fix, aggiornamenti e non solo..

2fast 24 Mar 2017 09:28
Per quanto riguarda le FMA3 , ecco la dichiarazione di AMD :

"We are aware of select instances where FMA code can result in a system
hang. We have identified the root cause."

<http://www.guru3d.com/news-story/amd-preps-bios-update-to-fix-fma3-freezes-for-ryzen.html>


MSI introduce la tecnologia A-XMP per AMD Ryzen
per migliorare prestazioni e compatibilita' delle memorie:
<http://www.hwupgrade.it/news/skmadri/tecnologia-a-xmp-nelle-schede-madri-msi-per-cpu-amd-ryzen_67832.html>


Su LegitReviews approfondimento sulle memorie SR e DR
con la piattaforma Ryzen :
<http://www.legitreviews.com/amd-ryzen-single-rank-versus-dual-rank-ddr4-memory-performance_192960>

Secondo i test condotti con delle Corsair LPX ,
la velocita' indicata da AMD per le memorie
(tra parentesi) si e' rivelata piuttosto conservativa:

DDR4-2666 - Dual-Rank 4 DIMM (1866)
DDR4-2666 - Dual-Rank 2 DIMM (2400)
DDR4-2666 - Single-Rank 4 DIMM (2133)
DDR4-3200 - Single-Rank 2 DIMM (2667)


A proposito di FMA , memorie, fix e ottimizzazioni,
segnalo la nuova versione di y-cruncher
con eseguibile compilato per Zen :
<http://www.numberworld.org/y-cruncher/ *******

Alcune note del programmatore:

"... y-cruncher v0.7.2 has a new binary specifically optimized for AMD's
Ryzen 7 processors. The performance gain is about 5% over the
Broadwell-tuned binary and 15% over v0.7.1. It turns out that the
optimizations between v0.7.1 and v0.7.2 happened to be more favorable to AMD
Zen than to Intel processors. Nevertheless, this is not enough to make Ryzen
beat Haswell-E or Broadwell-E."

"It's unlikely that any amount of Zen-specific optimizations can make Ryzen
beat Haswell/Broadwell-E. The difference in memory bandwidth and 256-bit AVX
throughput is simply far too large to overcome. AMD made a conscious
decision to sacrifice HPC to focus on mainstream."

"Slightly more concerning is a system freeze with FMA instructions which
appears to be have been confirmed by AMD as a processor errata. Fortunately,
the source also says this is fixable via a microcode update. So it won't
lead to something catastrophic like a recall or a fix that disables
processor features."

"As for the Zen architecture itself. Here are my (early) observations:
- The FPU block diagrams that were released about Zen appear to be accurate.
The FLOPs benchmark is able achieve 4 128-bit CPU instructions/cycle if
there is an equal distribution of FP-adds and FP-multiplies. As expected,
FMAs have twice the cost since they need both an FP-add and an FP-multiply.
- 256-bit AVX instructions are handled efficiently enough that it seems to
be beneficial to use 256-bit instructions when there's no overhead to doing
so.
-Memory is huge bottleneck. Latencies are very high and dual-channel memory
is simply not enough to feed this much computational throughput.
-The Ryzen 7 can run FMA4 instructions even though the corresponding CPUID
flag is cleared. This is probably to enable some compatibility with code
written for AMD Bulldozer while discouraging further use of FMA4. On the
other hand, XOP instructions did not get this treatment so they will crash.
Ryzen's ability to run FMA4 instructions makes it possible to run
y-cruncher's XOP binary without crashing for small computations."

"For software developers, compiling code on the 1800X is about as fast as
the 5960X at stock clocks. But the 5960X has much more overclocking
headroom, so it ends up winning by around 15%. For a $500 processor, the R7
1800X is very impressive."



Ciao
Luca
Ermete@pc00 30 Mar 2017 14:44
Passano i giorni e aumentano le sorprese
https://www.pcper.com/reviews/Processors/Ashes-Singularity-Gets-Ryzen-Performance-Update
e questo incremento con poche righe di ottimizzazione del codice...

2fast wrote:
> Per quanto riguarda le FMA3 , ecco la dichiarazione di AMD :
> "We are aware of select instances where FMA code can result in a system
> hang. We have identified the root cause."

>
<http://www.guru3d.com/news-story/amd-preps-bios-update-to-fix-fma3-freezes-for-ryzen.html>

> MSI introduce la tecnologia A-XMP per AMD Ryzen
> per migliorare prestazioni e compatibilita' delle memorie:

<http://www.hwupgrade.it/news/skmadri/tecnologia-a-xmp-nelle-schede-madri-msi-per-cpu-amd-ryzen_67832.html>

> Su LegitReviews approfondimento sulle memorie SR e DR
> con la piattaforma Ryzen :
>
<http://www.legitreviews.com/amd-ryzen-single-rank-versus-dual-rank-ddr4-memory-performance_192960>

> Secondo i test condotti con delle Corsair LPX ,
> la velocita' indicata da AMD per le memorie
> (tra parentesi) si e' rivelata piuttosto conservativa:

> DDR4-2666 - Dual-Rank 4 DIMM (1866)
> DDR4-2666 - Dual-Rank 2 DIMM (2400)
> DDR4-2666 - Single-Rank 4 DIMM (2133)
> DDR4-3200 - Single-Rank 2 DIMM (2667)

> A proposito di FMA , memorie, fix e ottimizzazioni,
> segnalo la nuova versione di y-cruncher
> con eseguibile compilato per Zen :
> <http://www.numberworld.org/y-cruncher/ *******

> Alcune note del programmatore:

> "... y-cruncher v0.7.2 has a new binary specifically optimized for AMD's
> Ryzen 7 processors. The performance gain is about 5% over the
> Broadwell-tuned binary and 15% over v0.7.1. It turns out that the
> optimizations between v0.7.1 and v0.7.2 happened to be more favorable to AMD
> Zen than to Intel processors. Nevertheless, this is not enough to make Ryzen
> beat Haswell-E or Broadwell-E."

> "It's unlikely that any amount of Zen-specific optimizations can make Ryzen
> beat Haswell/Broadwell-E. The difference in memory bandwidth and 256-bit AVX
> throughput is simply far too large to overcome. AMD made a conscious
> decision to sacrifice HPC to focus on mainstream."

> "Slightly more concerning is a system freeze with FMA instructions which
> appears to be have been confirmed by AMD as a processor errata. Fortunately,
> the source also says this is fixable via a microcode update. So it won't
> lead to something catastrophic like a recall or a fix that disables
> processor features."

> "As for the Zen architecture itself. Here are my (early) observations:
> - The FPU block diagrams that were released about Zen appear to be accurate.
> The FLOPs benchmark is able achieve 4 128-bit CPU instructions/cycle if
> there is an equal distribution of FP-adds and FP-multiplies. As expected,
> FMAs have twice the cost since they need both an FP-add and an FP-multiply.
> - 256-bit AVX instructions are handled efficiently enough that it seems to
> be beneficial to use 256-bit instructions when there's no overhead to doing
> so.
> -Memory is huge bottleneck. Latencies are very high and dual-channel memory
> is simply not enough to feed this much computational throughput.
> -The Ryzen 7 can run FMA4 instructions even though the corresponding CPUID
> flag is cleared. This is probably to enable some compatibility with code
> written for AMD Bulldozer while discouraging further use of FMA4. On the
> other hand, XOP instructions did not get this treatment so they will crash.
> Ryzen's ability to run FMA4 instructions makes it possible to run
> y-cruncher's XOP binary without crashing for small computations."

> "For software developers, compiling code on the 1800X is about as fast as
> the 5960X at stock clocks. But the 5960X has much more overclocking
> headroom, so it ends up winning by around 15%. For a $500 processor, the R7
> 1800X is very impressive."

> Ciao Luca

Links
Giochi online
Dizionario sinonimi
Leggi e codici
Ricette
Testi
Webmatica
Hosting gratis
   
 

Informazioni e notizie sui processori per calcolatori | Tutti i gruppi | it.comp.hardware.cpu | Notizie e discussioni hardware cpu | Hardware cpu Mobile | Servizio di consultazione news.