|
|
 |
|
|
|
Among the innovations from Asus, we note the arrival of a new series of graphics cards: Formula. At launch the series is made up of an overclocked Radeon HD 4890, a Radeon HD 4770 and a GeForce 9400 GT. They are equipped with a cooling system that looks like an F1 car. A bonnet that allows Asus to create high and low pressure zones to better manage the flow of air. It remains to be seen if it is as effective as it looks.
 |
| |
|
Asus’ famous GeForce GTX 295 MARS Limited Edition was of course on show at Computex. We had the opportunity of seeing a dismantled version and noting that the PCI Express switch is in fact an NVIDIA NF200, contrary to what we first thought. To recap, the card is equipped with 2 GT200s in the same configuration (240 SP and 512 bits) and clocks (648/1476 MHz) as the GeForce GTX 285. Also, each GPU has 2 GB of 1200 MHz GDDR3.
The card needs 2 8 pin power supply connectors and is over 3 cm (even a little more with the SLI connector), the standard height for graphics cards. In terms of connectivity it has 2 DVI connectors and an HDMI connector, like the double PCB GeForce GTX 295, rather than just the DVIs on the mono-PCB.
  |
| |
|
We were able to see 3 new, more affordable, Nehalem derivatives in action from Intel: Lynnfield, Clarksfield and Clarkdale. To recap, the first two use a similar organisation to the Core i7s although there are now 2 memory controllers instead of 3 so as to make room for 2 PCI Express 8x controllers (or 1 16x). The QPI bus has also been discarded. Lynnfield is the desktop version while Clarksfield is its laptop equivalent.
The TDP of these processors will be significantly down on the Core i7s but Intel didn’t go into any detail on this. The manufacturer did however say that there will be a higher margin in Turbo Mode for these processors, which means that motherboard manufacturers will have to implement it correctly rather than simply overclocking in load so as to emulate it as some are currently doing. The Core i7s are however unlikely to be in competition with them and will keep their advantage in terms of basic clocks and overclocking. Note that the difference in performance between desktop and laptop versions, at the same base clock, could be significant because Turbo Mode will have less room for manoeuvre in the laptop versions.
8 threads on a laptop will soon be possible.Intel is waiting impatiently for Windows 7, which will finally be more adapted to execution with all 8 threads. Windows 7 will be able to park cores so that they are completely inactive when they aren’t needed. The cores will be progressively reactivated according to load. Next, and this is linked, the Windows 7 scheduler will abandon automatic migration of threads from one core to another. Currently in Windows, when a single thread loads the processor, it in fact moves constantly from one core to another which, on top of reducing performance levels, stops the cores from staying in idle. This was a choice made by Microsoft to optimise load balancing on server versions of Windows and it hasn’t been modified since. With Windows 7, if one thread is loading a core effectively, it won’t switch. The new shceduler will also take hyperthreading into account so as not to use it until all the cores have reached a given load level. Intel is expecting to be able to increase its advantage over AMD with Microsoft’s new OS. To finish this little tour of forthcoming Intel CPUs, we were also able to observe the Clarkdale in action. To recap, it is the first dual core CPU with an integrated graphics core. This isn’t in fact much of a technological exploit but simply the integration of CUP dies with the northbridge in the same packaging. It will be interesting to see what performance this solution gives, given that it sheds part of the new Nehalem architecture, such as integration of the memory controller in the CPU, because it will be in the northbridge. Intel demoed this new entry level solution with Bluray. 
The Clarkdale platform on demo. |
| |
|
We managed to interview Tom Sonderman, Vice President of Manufacturing Systems and Technology of Globalfounderies who presented the objectives of the new foundary, previously a division of AMD. Globalfoundaries clearly aims to establish itself as the major competitor to TSMC and try to attract its clients.
In order to do this Globalfoundaries is keen to draw attention to the assembly of a very qualified team, drawn from the talents of numerous players in the semi-conductor industry. They are also highlighting the reliability of their fabrication processes, developped over time on production lines.
This strategy has been put into place because AMD did not have the means to set up a pilot line for the exclusive use of the development of new processes. The disadvantage is that this now monopolises part of the production capacity and can sometimes prolong development time. On the plus side however, the foundary will be able to move over to new processes when they are truly ready and avoid being trapped in unproductive lines. This is, then, a way of limiting risks but also a way of introducing little improvements from time to time.
With regard to these new processes, AMD is currently finalising 40 and 45 nm bulk, ie. not SOI. They will be used to attract customers who do not want to complicate things with SOI technology that can be complex to put into place. These processes could also be used to attract current TSMC customers who aren’t happy with the problems that it is currently experiencing with its 40 nm line. Getting new custom on board with these processes is not however Globalfoundaries’ priority. Rather, they want to show that they master the technology so as to convince potential customers to opt for their services in the future.
And here we’re talking 32 nm technology, bulk and SOI, with production starting in 2010, probably in the second half of the year. In 2011, Globalfoundaries will be offering a small improvement to this process that will go to 28 nm, bulk only.
Globalfoundries did not come to Computex empty handed!All these developments will take place in Fab 1 in Dresden that will dedicate a priority module to SOI production and another to bulk. Globalfoundaries told us that each module will be fully equipped however so that, although each will have its priority, each will also be able to do some of the work of the other, depending on demand for the respective technologies. Fab 2, which will be built in New York, should go into production in 2012, starting at 28 nm but designed for 22 nm. In terms of the relationship with AMD, Globalfoundaries was first of all proud to announce the arrival of the Istanbul, the 6 core Opteron, which shows its capabilities in terms of quality of execution. Tom Sonderman nevertheless said that AMD’s influence in Globalfoundaries would gradually be reduced. First of all because AMD needs to concentrate on its products and secondly because its share in the new foundary (34% but 50% of control) will gradually fall in proportion to investments from its partner ATIC. The contract between the parties states that AMD sees itself offering to make an equivalent investment so as to conserve its share but according to Globalfoundaries this is unlikely to happen. This is a way of organising the separation gradually so that AMD’s CPU team has the chance to move over to the “fabless” way of working. This should be to the advantage of both companies, making them more agile and focussed on their core activity. AMD, for example, will be able to be more meticulous in the development of its CPUs so as to avoid having to carry out numerous tests and revisions. When the manufacturer also owns the foundary, it is easy to fall into this. If Globalfoundaries has to carry out these tests however, it will encourage AMD to think twice before asking for them! This pressure can only be of benefit to AMD’s engineers, according to the foundary. Globalfoundaries left us with a pretty good impression overall. With pragmatism and effective execuction, the foundary shouldn’t have too much trouble in attracting the services of numerous chip designers, including of course AMD and NVIDIA for GPUs. Globalfoundaries' arrival moreover comes at an auspicious moment: TSMC is currently having a great deal of difficulty in terms of production processes and this is proving a problem for AMD and NVIDIA in particular. On top of the fact that Globalfoundaries will simply bring increased competition, there can be little doubt that TSMC’s failures will also sway AMD and NVIDIA when they come to sign their future contracts. |
| |
|
Announced by NVIDIA just before Computex opened, we have been able to get a closer look at Supermicro’s new server. In the 1U format, it houses two Tesla cards, 2 CPUs and 3 hard disks. It is also a full server as opposed to the Tesla S1070 that only contains accelerators and therefore needs to be connected to a master server.
 In its default configuration, the server has a bi-Xeon 5500 platform (Nehalem) and 2 Tesla M1060 accelerators. It can support up to 48 GB of memory as well as 8 GB in total on the 2 Tesla cards. Supermicro told us that a different platform is possible, either mono-CPU or Xeon from the previous generation or Opteron. You can also use Tesla C1060 cards in place of the M1060, with the difference being their format. The first are standard cards with a fan while the second only have a large cooler over which a sufficient flow of air needs to flow, the case on this server. Supermicro said that the server should be available within a couple of weeks; putting it on sale is a priority for the company as numerous orders have already been placed, which underlines the interest in these massively parallel accelerators. It is also a very important product for NVIDIA, now working with a recognised player in the server market, in terms of pushing Tesla. As for pricing, in its basic configuration, it will come in at between 8 and 12,000 euros depending on the margins taken by distributors. Note that Supermicro is preparing a second Tesla server in the 4U format for 4 Tesla cards. |
| |
|
AMD has made the most of Computex to demonstrate its first DirectX 11 GPU, expected to come on stream this autumn. For AMD, DirectX 11 support is relatively simple as its current GPUs already support almost all its functions. Only slight adjustments are needed to make the architecture compatible with the forthcoming Microsoft API.
AMD’s relatively basic demo was based on a tessellation using DirectX 11, that standardises this function supported by Radeons since the HD 2000 series. AMD took the opportunity of reminding us of its previous more complex demos that were based on DirectX 10 and were used to market the Radeon HD 2000, 3000 and 4000.
 AMD has not then really given us anything particularly new or impressive. The important thing is that AMD already has a functioning DirectX 11 GPU and should, if everything goes well, be the first on the market with compatible products. This could well happen in time for the launch of Windows 7 on the 22nd October. The other important info from AMD concerns the size of the first DirectX 11 chip. TSMC has in reality given AMD a wafer of the first DirectX 11 chip engraved at 40 nm. This is an indirect way of giving us the load down on this chip as the die is relatively small. Evergreen, code name of this GPU, measures approximately 180 mm². Very small indeed. This would seem to indicate that this GPU will only contain a little over a billion transistors, or only 100 million more than the RV770/RV790.  Although 180 mm² may seem small for a 256 bit bus, we think this is the configuration AMD has gone for. In terms of processing units, they may stay at 160 or go up to 192 vec5. This strategy would allow AMD to place itself on the market rapidly and conquer the €150/250 segment. A bi-GPU card of course is still a possibility for the top end, but this will not perhaps be enough against the NVIDIA GT300 that looks like being an enormous GPU and giving a big jump in performance on the previous generation. This said, AMD is also preparing a bigger GPU which is likely to come a bit later. Of course TSMCs contribution is still an unknown – at the moment problems with the 40 nm process are slowing availability of the Radeon HD 4770. By going for a small die, at least at first, AMD is limiting the risks here, because it will be simpler to make. With a big die however, the situation may well be more of a problem for NVIDIA if TSMC doesn’t get to the bottom of the problem quickly. Whatever the case may be, AMD should have a few months headstart. This should allow AMD to supply developers with its hardware before its competitor and therefore become the platform of reference for DirectX 11, with all the advantages that this implies.  |
| |
|
|
Copyright © 1997- Hardware.fr SARL. All rights reserved.
Read our privacy guidelines.
|
|
