Advanced Micro Devices Inc (AMD)

[rupert]

Opteron 250 benchmarked
But 939 boards not exactly ready

"TOP GERMAN MAG c’t has benchmarked the AMD Opteron 250 and there’s talk not only of dual systems but quad systems at its web site.........."





http://www.theinquirer.net/?article=16223

EDIT: This is the babblefish version I made of the German article.

Tandem teams AMD Opteron 250 against Intel Xeon 3.20 GHz with 2 MByte L3-Cache the 64-Bit-Serverprozessoren of AMD begin now with 2,4 GHz against Intels Xeons. The eternal race goes into a further round, this time into the workstation and server disciplines. AMD pushes with full Opteron force into the lucrative market of the binary systems, which Intels Xeons dominate to time. Good bench mark results and favourable prices gave much attention to the Opterons - enough anyhow, in order to induce Intel to the announcement AMD64-kompatibler processors. Also the past Xeons has Intel better against the competition prepared. Only for the expensive multi-processor (MP)Xeons developed Gallatin core still 1 or 2 MByte L3-Pufferspeicher is for some time available originally also for dual processor (DP)Xeons and gives these additionally to 512 KByte L2-Cache. A DP Xeon with 3,20 GHz and 2 MByte L3-Cache costs approximately 1200 euro. The Opteron 250 (2.4 GHz, 1 MByte L2-Cache) is not to be had in the German retail trade yet, might however at least 150 euro become more inexpensive (OEM list price 851 US dollar). For single as well as four to eightfold systems Opteron 150 and 850 are meant. In this test we compare the AMD Opteron 250 with the Intel Xeon 3.20 GHz with 2 MByte L3. over also diagram bench mark to accomplish to be able, we in each case two CPU tandems on workstation boards with AGP Slots tested. AMD recommends the Tyan Thunder K8W (S2885, about 480 euro), the Xeons was on nearly equal expensive SE7505VB2 von Intel (test into 18/03 c't). For two gigabyte memory from 512-MByte-Registered-DIMMs one must put on to time about 640 euro; there is not a considerable difference in prices between the PC2100R-Riegeln of the Xeons and the PC3200R-Modulen for the Opterons thereby. We used Mushkin DIMMs, which exhibit particularly short latencies (with 200 MHz as PC3200: 2,0-3-2, with 133 MHz as PC2100: 2,0-2-2), but because of their TSOP chips not the JEDEC standard corresponds. Workstations with equipment, which corresponds approximately to the test configurations, might cost at least 4000 euro. For a professional diagram map and a SCSI subsystem easily still 2000 euro are added. The fun is not only expensive, but also loud: The maximum waste heat production of these computers reached about 300 Watts and leads therefore to substantial fan noises (see in addition also page 81). Bench mark such expenditure is worthwhile itself only for strong computing power, and of it both twin couples offer enough. In most bench mark lie together their achievements closely, a view of the details shows however partially extreme differences. An example of it the results are accomplished in the CPU2000-Benchmark of the SPEC [ 1 ], under Windows XP (32 bits): The total results for firm (INT) and tasks of floating decimal point (FP) differ only in each case by a half per cent. But in individual tasks (further measurements under [ 1 ]) the differences between the processors up to 33 per cent amount to. The differences are still more glaring with the rate valuations, which measure the co-operation of CPUs in multiprocessor systems (scaling): Here the Opterons lies particularly in the floating decimal point discipline far in front (31 per cent), in single valuations gives it even differences of up to 112 per cent (with the Swim.171). Measurements on [ 1 ] with 64-Bit-Linux and the compiler GCC 3,3 show that the Opteron team still adds then with the CFP2000-Rate-Wert, in the CINT2000 rate however somewhat drop. The very good CPU2000-Rate-Werte is to be due probably to the NUMA architecture of AMD64: Each processor is directly tied up and communicated to fast RAM by hypertransport (6.4 GByte/s) with its CPU partner. Intels Xeon tandems against it must complete all accesses over a common FSB533 (4.3 GByte/s). Soon however the Xeons with FSB800 and DDR2-Speicher comes - it remains thus exciting. Performance data under Linux and Windows XP Professional. For servers with four and more processors plans Intel however first no FSB acceleration, and also with these computers the AMD64-Architektur shines: For the ProLiant DL585 HP announces a value of 770 Usern in the SAP SD bench mark; the Quad Xeon server IBM x365 lines up with MP Xeons with in each case 3 GHz and 4 MByte L3-Cache by 6,5 per cent behind it (720 user). Suns dual Opteron system V20z brings it with two Opteron to 248 to 410 user, nearly up (408) lies a HP ProLiant BL20p with 3,2-GHz-Xeons (with 2 MByte L3). Still hypertransport and AMD64 do not bring side and hyper+ Threading and high clock frequency on advantages so clear on the other side that one could kueren one of two processor architectures to the generally better product. Perhaps are missing also only suitable bench mark? Who stands before a purchase decision, must see completely exact anyhow. Particularly with professional servers, where one will use not constantly new, optimized compiled software for reasons of the operating and network security as well as because of the administration expenditure, one comes probably only with power measurements with the concretely planned programs of far (ciw) literature [ 1 ] official results of the SPEC CPU2000: www.spec.org/cpu2000/results/------------------------------------------------------------------------------- of comments: Achievement! (MacAthur, 29,5,2004 11:54)

EDIT: And this is my babblefish version of the article on the lack of 939 motherboards.

Zweispurig developed AMD Athlon 64 3800+ and Fx-53 for the base 939 Nichtmal one year after the introduction on the market of the Athlon 64 presents the third base format already to AMD. That sounds annoying and nervt to Upgrader, proves with the view of the technical background however as meaningful. Completely besides the Athlon makes the life still heavier for 64 3800+ Intels Pentium 4. AMD had developed two platforms for the two variants of the 64-Bit-Prozessors: The server processor Opteron with two memory channels needs Sockel-940-Mainboards and buffered memory. The Desktop variant Athlon 64 has only one memory channel and runs on more inexpensive Sockel-754-Boards. AMD wanted to establish the advantage of the two memory channels also within the Desktop range and therefore to the Athlon 64 FX offered - for the base 940. Because it needed special memory (registered PC3200R), did not run it not in all Opteron boards. Besides the boards were rather aligned to the server as the Desktop market: The current savings technology Cool?n?Quiet is not missing, some boards has AGP, but but Pci x or Onboard SCSI. Fast this FX the call had to be away, only a renamed Opteron, a notloesung. The new base 939 is one for Highend PCS custom-made platform: It supports like the base 940 two memory channels. Owing to skillful layout the obligation to buffered memory is void, whereby also Cool?n?Quiet functions again. The base 939 is however not dualable. Architect two processors for the new base AMD presents at first, to the Athlon 64 3800+ and the Athlon 64 Fx-53. Both use two memory channels and run like the Opteron 150 (see c't 12/04, P. 20) and in March the introduced Fx-53 for the base 940 with 2,4 GHz. Because the number of the memory channels is omitted as demarcation of the FX series of the Athlon 64, AMD has now the L2-Cache as small difference auserkoren: While so far the Athlon-64 and mobile versions were equipped with 512 KByte or 1 MByte depending upon model, each 939er-Athlon will have only 512 KByte and each FX 1 MByte L2-Cache. Besides AMD accelerated the Hypertransport Frontsidebus from 800 MHz to 1 GHz (HT1000), transfers it now maximally 4 GByte/s into both directions. That support three chip-corrode: Nvidia nForce3 250, the SiS755FX and VIA K8T800 pro. Overhauling maneuver so far exists however no hardware, which reaches this speed: Agp-8x gets along with approximately 2 GByte/s and uses already only rarely, and even the fastest Athlon-64-Chipsatz Nvidia nForce 250Gb binds the other periphery (PCI, non removable disks, LAN) also straight times maximally 400 MByte/s on (see ct 12/04, P. 138). Pci x transfers 533 MByte/s, to Sockel-939-Boards not to be found will however probably have. HT1000 nevertheless offers air for PCI express, which hunts as diagram map interface in the 16X-Version 4 GByte/s per direction by the conductive strips. Pilot lot boards of MSI and Asus with the Via Chipsatz were to us at the disposal. At first they ran disappointing slowly, then different bio its positions - above all switching the 2T COMMAND off rate - lifted the bench mark results on the expected order of magnitude. Only a test of production stages boards will uncover the final performance of the 939er-Plattform. After the past measurements the Fx-53 in the base 939 due to the unbuffered memory modules about two per cent works faster than in the base 940. In running with Intels Pentium of 4 extremes edition differ most test results around less than five per cent. With some 3D-Benchmarks the Fx-53 has clearly the nose in front, the PÊE profited however by the Cinebench 2003 from the Hyperthreading and by the memory throughput from 2 MByte large L3-Cache. The 3800+ profits to 64 3400+ (1 MByte L2-Cache, 2.2 GHz clock, base 754) in the comparison to the Athlon more strongly from the second memory channel and from the higher clock, than it loses due to the smaller Cache: Many applications do not add more clearly, than alone the higher clock explains, not one program work more slowly. The designation 3800+ suggests more speed than the 3400+, which after the past measurements only with few applications the case is however twelve to per cent. Before the Pentium 4 with 3,4 GHz the 3800+ with most applications, only remains the domain leads the binaryable of hyper+ Threading. The Athlon 64 exists in three designs for the base 754 (left), the base 940 (center) and the base 939 (right). Cool?n?Quiet did not function yet perfectly: The two Sockel-939-Prozessoren throttled their clock on 1,2 GHz, however no more shifted up to their full clock. With a memory module both processors functioned, worked then however naturally clearly more slowly - with FR C C dropped back the 3800+ even behind the 3400+. Mixed components did not run reliably; AMD recommends, only modules with same capacity to use organization and CAS Latency parallel. With the assembly with 4 GByte only about 3.5 GByte can be used due to the faded in PCI address area also in the 64-Bit-Modus - a restriction, from which so far all x86-Chipsaetze suffers. Track switching the Sockel-939-Plattform combines the advantages of the two past platforms for AMDs 64-Bit-Prozessor: Two memory channels for quite inexpensive PC3200-Speicher provide for high speed and support Cool?n?Quiet, the boards leave themselves for Highend PCS massschneidern. The smaller L2-Cache brakes the Athlon 64 hardly out, the advantage of the two memory channels weighs more strongly. The 939er Fx-53 outdated even the Opteron and proves as the fastest AMD processor. Also the Intel competition the AMD processors leave with many applications behind itself, only the hardly available and expensive PÊE keep up. Agp-8x and PCI must exist probably still at least a half year long, before PCI express Chipsaetze for the Athlon are available 64. All too heavily waiting should not fall however, brings nevertheless ATI and Nvidia their current diagram chips also in a AGP version (see c't 11/04, P. 130). Nvidias chip set nForce3 ties up besides non removable disks and Gigabit LAN so fast that the whole 133 MByte/s of the PCI are available for extensions. Thus at present few reasons exist not to seize with the desire for much performance to the base 939 - if production stage board and processors available are (jow) -------------------------------------------------------------------------------