| T.R | Title | User | Personal
Name | Date | Lines |
|---|
| 656.1 | all chips have bugs | WRKSYS::HOUSE | Kenny House, Workstations Engineering | Sun Mar 09 1997 09:21 | 16 |
| re .0 - "Are there any known problems with the PLX PCI9060-2 working
with a Pyxis based system?"
A quick look through the PCI mailing list archives shows that there are
at least 15 errata with the PCI9060-2 part, some of which affect data
reliability and performance. Specific details weren't available there,
so I'm off to their web site ( http://www.plxtech.com/ ) to see what I
can find. The -3 part was announced over a year ago, claiming to fix
lots of -2 problems in an allegedly pin- and program-compatible part.
Pyxis has its own problems, with simple workarounds, and I'm working
with John offline to see whether this card exposes any of them. A
slightly different spin on this customer's issue is being discussed in
WRKSYS::XL_PERSONAL_WORKSTATION note 289.
-- Kenny House
|
| 656.2 | s | SCASS1::MARIA | John Maria | Sun Mar 09 1997 20:26 | 9 |
| The vendo building the custom card has told my customer the design is
frozen. It seems silly that they would not at least try the -3 if it
is pin, and signal compatible.
I will ask the customer to pursue having the -2 replaced with a -3,
even if only for an experiment!
Thanks for working weekends!
John
|
| 656.3 | Just for completeness... | DECWET::LIVINGSTON | James W. Livingston, Jr. | Tue Mar 11 1997 14:05 | 9 |
|
it's worth noting that the Alcor chipset supports a memory path that's
256 bits wide, which could account for some better I/O performance, at
the same processor speeds. The greater bus width is what makes the
Alcor/Maverick/Brett cost more, to some degree. Mind you, the Alcor
chipset has its own problems.
Cheers,
James
|
| 656.4 | Pyxis is MUCH FASTER than Alcor | WRKSYS::SCHUMANN | | Tue Mar 11 1997 16:35 | 22 |
| re .3
Although the Alcor memory path is 256 bits wide vs 128 for Pyxis, Pyxis is
uniformly faster on all memory accesses. The measured McCalpin bandwidth for
Alcor/Bret is about 170 Mbytes/sec vs 292 Mbytes/sec for Pyxis (232 Mbytes/sec
with 2MB cache). The measured memory latency (including the wasted cache miss)
for Miata/Pyxis is 130ns vs 330ns for Bret.
see also
http://src-www.pa.dec.com/SRC/personal/berc/miata/index.htm
and
http://www.cs.virginia.edu/stream/
The only performance metric where Alcor performance will exceed Pyxis
performance is on DMA read bandwidth for long (>64 byte) accesses. That's
because Alcor supports DMA read prefetch and Pyxis does not. The
max sustained DMA read bandwidth on Pyxis is estimated at 160 Mbytes/sec,
whereas Alcor can presumably do over 200 Mbyte/sec. This performance
difference impacts very few real-world applications.
It's bad enough that Aspen is spreading anti-Pyxis FUD. I would hope that
our own employees would refrain.
|
| 656.5 | Where can I get info? | SWAM1::POIANI_MI | | Sat Mar 29 1997 12:27 | 7 |
| I've seen data stating the Pyxis can actually do 1.3GB per second
sustained memory bandwidth, probabally without cache. What's the
bottleneck.
Is there any white papers or tech docs on Pyxis?
MP
|
| 656.6 | peak vs sustained | WRKSYS::SCHUMANN | | Sat Mar 29 1997 15:54 | 13 |
| There's STREAMS benchmark data at
http://www.cs.virginia.edu/stream/
The cacheless Personal Workstation 433a does 292.6 Mbytes/sec on COPY, which is
competitive with other vendors' high-end workstations. Pyxis is actually capable of
slightly faster COPY performance, if using CPU frequencies 466, 533, 600, etc.,
which permit the memory bus to run at exactly 66 MHz, rather than the 61 MHz
clock rate used on the 433a.
The "peak" bandwidth of the Pyxis memory bus is about 1Gbyte/sec.
--RS
|