October 22, 2003

Industry News and Site Updates - Josh

Unfortunately I am not in on the latest NVIDIA card releases, but it looks like Thursday morning will be the end of the embargo for the FX 5700/Ultra and FX 5950 Ultra reviews.  While Anand was able to show off the 5950, NVIDIA didn't allow him to do the same with the 5700 Ultra.  The 5950 looks to be merely a slightly reworked 5900 Ultra, with faster core and memory speeds, and that is about it.  The hints leaked about the 5700 Ultra are quite different.  It does appear that I was half right when I said both the NV 36 and NV 38 would be significant redesigns (in fact, it looks as if only the NV 36 is the only significant redesign).  We can expect a solid increase in clock speed (around 475 MHz if I caught the hints correctly), as well as a memory speed around 400 to 425 MHz DDR.  Most likely this core is based off the NV 35 core (vs the NV 30 core the FX 5600 series was based on).  As such, we can expect correspondingly better overall performance from the part (in most applications the FX 5900 was competitive with the Radeon 9800 series).  This could very well be a true competitor to the mighty Radeon 9600 Pro, and the recently released 9600 XT.  I guess we will find out tomorrow.

Speaking of the 9600 XT, I was a bit disappointed in one of the aspects of the design.  I really thought that the memory speed should have been increased with the core increase.  While going from a 400 MHz core to a 500 MHz core is nothing to sneeze at, leaving the memory running at 300 MHz DDR doesn't make a whole lot of sense to me.  Any increase in the memory speed would have a pretty significant impact on overall performance.  In tests where reviewers overclocked the memory, the performance of the 9600 XT really took off as compared to the 9600 Pro.  For the most part, the core is probably idling more since it can't get the data it needs from the memory.  If ATI had increased the memory even to 325 MHz DDR (650 MHz effective), much less if they had done 350 MHz DDR (700 MHz effective), the 9600 XT would have probably blown everything else away in its price range, and honestly would have cost at most $15 more with that single upgrade.  As it is, I feel that ATI did hobble the design to a great degree by setting the memory speed at 300 MHz DDR.  The jump in core speed will help in areas where memory bandwidth is not as important (such as complex shader ops), but in AA and AF performance, it is essentially in line with what the 9600 Pro can do.  Perhaps once NVIDIA releases the FX 5700 Ultra ATI will change its mind and do a slight revision of the 9600 XT with faster memory.  We can only hope!  Perhaps the other 3rd party manufacturers will produce more overclocking friendly boards with faster rated memory, so users can freely experiment with different speeds.  The current design appears to have memory that can barely reach 320 MHz due to the lower speed rating.

The site redesign is continuing along, and I need to get some artwork to the guys doing most of the technical stuff.  So far I am happy with what I see, and I hope to get the new design up and running.  I am also in negotiations with some new advertisers, so this will help the site overall as well as give it a bit more of a professional flair.  I too get tired of the popups and the Nexium ads, so hopefully this will all change within the next few weeks.  Originally I was hoping to have the new site up by now, but things never seem to go as planned.

I am desperately trying to finish my Athlon 64 3200+/Asus K8V Deluxe review, as this has turned out to be a very potent combination.  For now it appears like the VIA chipset is slightly faster than the NVIDIA nForce 3, but shortly there should be the release of the new 250 Pro series of chips from NVIDIA.  This looks to be a significant improvement over the nForce 3 150/Pro, with a faster running HT bus, as well as integrated SATA.  So for the time being, if I were going to invest in an Athlon 64 board, I would probably want to use the VIA based solution.  Of course the differences are pretty minimal, so if there are features you really like on an nForce 3 150 board, then by all means get it!  For those interested in the few extra frames or seconds of performance, then hit the VIA stand.  I am going to do my best to get my review up within the next two days!  After that, we should have some more nice content to share!  The next few reviews should be more audio in nature, as well as exposing an existing product to a new consumer base.

October 17, 2003

Abit nForce 2 NF7-M Motherboard Review - Josh

Keith had this one done quite some time ago, but due to all the happiness involving the site as of late, this has been the first chance we have had to post it!  Even though AMD has released the Athlon 64 chips to the market, they will be slow to trickle down into the mainstream and budget desktops.  Until that time, the Athlon XP will be the processor of choice for AMD fans.  Abit continues to address the lower end market with its integrated NF7-M, and due to its low overall price and solid feature set, this board should continue to do well.  Here is a quote from Keith's review:

     It is fairly obvious that NVIDIA hit a homerun with the nForce 2, and many numbers show that NVIDIA is the top Athlon based chipset company in the world.  Now that the Athlon 64 has been released, the nForce 2 and Athlon XP has been firmly pushed down into the midrange/low end market.  This is not to say that the nForce 2/Athlon XP combo is no longer a viable product.  The nForce 2/Athlon XP are in fact a very robust and high performing combination, and the price to performance ratio makes it a contender.  Let’s see how Abit does with its mainstream NF7-M motherboard.

You can read the entire review here.

AMD Announces Earnings - Josh

AMD beat estimates by a significant margin this past quarter, posting revenues of $954 million, where most analysts expected AMD to earn around $866 million.  AMD attributes this to strong PC processor sales, with much higher ASPs, as well as very high demand for its flash products.  This really shouldn't be much of a surprise, at least about the processors.  For the past several months AMD has kept the Athlon XP 3200+ at the $460 mark at retail, and its even the low end Barton processor (the 2500+) is still over $110.  The very fact that AMD has kept its processor pricing stable, even through the increased competition from Intel with its Pentium 4 "C" models, shows that AMD is experiencing great demand for its processors.  The sales of Athlon 64 parts probably didn't hurt either, though its impact should be minimal due to its release at the tail end of last quarter.  Still, current prices for the three top of the line processors from AMD are all above $400 (Athlon XP 3200+ is $460, Athlon 64 3200+ is $440, and the high end Athlon 64 FX 51 is around $760).  Sales for all three products are strong, especially for the AXP 3200+ and A64 3200+.  The Athlon 64 FX 51 is in very limited supply at this time, most likely due to AMD not wanting to really push the Socket 940 architecture, since it will soon be taken over by the more consumer level friendly Socket 939.  Socket 940 will continue for a while though, and it will still be the basis for Opteron class products.

With the restructuring of the flash business with Fujitsu (the whole FASL deal), AMD's breakeven point is now quite a bit higher than before.  The good news from that is AMD can include more of the flash business on its balance sheet.  Many now expect AMD to post a significantly higher revenue for the current quarter due to greatly increased market demand for its high end processors, as well as the holiday buying season.  With results like this from AMD and Intel, many think that the hardest part for the PC industry is over, and more companies should be seeing a rebound with their bottom line.  With strong demand for its flash products, as well as greatly increased demand for the Athlon 64 processors (and the non-eroding price of their Athlon XP products), AMD could have revenues as high as $1.2 billion.  This would probably include a net profit of $150 million or so.  Even though AMD says the break even point would be around the $1 billion mark, I could easily foresee that number go higher due to the higher cost of producing significant quantities of Athlon 64 product to meet demand.

By Q1 of 2004, AMD is hoping to have a 50/50 mix of A64 and AXP wafer starts.  AMD also claims that their 90 nm process is now in the engineering sampling phase, which could mean consumer level wafer starts by Q2 2004 (though AMD merely says sometime in 1H 04).  Next summer should see the introduction of the first 90 nm Athlon 64 parts.  At that time, the die size of the A64 will be significantly smaller, and AMD can start to really pump out the chips.  At that point, the Athlon XP may go through another shrink and be considered their very low end part (I would estimate that the current Barton, which is around 105 mm sq. with the 130 nm process, could shrink to around 60 to 70 mm sq. with the 90 nm process).  This would be nice for AMD, as it could get around 400 dice per wafer.  If yields for this product are good, then this would be a very solid, low end processor.  With that size, it might even be a strong candidate for high powered handhelds (clocked down to 1 GHz or so), as well as the thin and lights at higher speeds.  With the way AMD ramps new processes, if the 90 nm process looks robust enough, I would say that by early Q4 2004 nearly all wafer starts at Fab 30 will be 90 nm.  By this time though, the death of Socket A will probably be near, as AMD will transition the Socket 754 to the budget sector, and keep the Socket 939 as the midrange and high end product.  There may be a Barton class product that has an integrated memory controller that will allow it to work on Socket 754, but this hasn't been totally confirmed.  Die savings would be impressive over the current Athlon 64, as the Barton class product would only have 512 K of L2, as well as none of the extra registers and transistors that give the Athlon 64 its 64 bit functionality.  Still, it would take a significant amount of engineering manpower to do such a thing, and may in fact be counter productive to AMD's 64 bit stance.  If AMD wants its 64 bit processors to succeed, it needs to push 64 bit from top to bottom.

September 24, 2003

AMD News and Insights - Josh

As you have all read, AMD has released the Hammer to the consumer market.  The Athlon 64 3200+ features the single 64 bit memory controller, while the higher end Athlon FX 51 features the full 128 bit memory controller (essentially dual channel).  Two things are very evident.  The Athlon 64 3200+ performs on par with the Pentium 4 3.2 GHz, basically being a very true competitor to that part.  Pricing for the Athlon 64 3200+ is significantly lower than the P4 3.2 GHz, and the ability to upgrade to a 64 bit operating system in 2004 makes it a very tempting investment.  These processors are flying off the shelves as we speak, and most online retailers are selling out very quickly.  I know for a fact that larger firms like Tech Data have quickly exhausted their supply of the parts, and are anxiously awaiting AMD to deliver more to them.

The other thing mentioned in nearly all the reviews is the dislike of using registered DIMMS for the Athlon FX 51.  The Athlon FX is essentially a remarked Opteron 246 with the extra HT links disabled.  This processor is honestly the fastest processor a user can buy for the desktop.  While Intel tried to derail AMD here with the release of the Pentium 4 3.2 GHz Extreme Edition (with the 2 MB of L3 cache), reviews show the FX 51 beating out the P4 3.2EE in the majority of tests.  A user will pay a hefty price for this part though, as each chip is well over $700 and registered DDR 400 DIMMS are not so cheap either.  Add to that the FX 51 is out in even fewer numbers than the Athlon 64 3200+.

Now for the interesting stuff.  Each Athlon FX and Athlon 64 is a large die for AMD to produce.  193 mm square to be exact.  Compare this to the Athlon XP 3200+ which sits around 105 mm square (IIRC).  Now, at 105 mm square, AMD has enough wafer production to supply approximately 30% of the processor market.  By doubling the size of the processor, the amount of dice per wafer goes down by 2/3.  The current Athlon XP is experiencing excellent yields.  Using simple calculations, each 200 mm wafer at the Dresden fab can produce up to 290+ dice.  Using a healthy yield for such a product (lets say 80%) and 232 good dice can be harvested.  With a 193 mm square die, only 130 Athlon 64 cores can be produced, and out of those 130 only around 55 of them are good.  If Fab 35 were to produce 100% Athlon 64/FX at a wspw of 5000, then we would only expect 275,000 good die to be produced.  Of course Fab 35 isn't producing 100% Athlon 64s.  In fact, I would guess that only 1000 wspw were given to the Athlon 64/FX/Opteron, and that might in fact be too high as it stands.  So, we are seeing AMD deliver well under 50,000 chips a week.  50,000 seems a lot at first, but when looking at the overall market, it will hardly make a dent.  Fab 35 continues to make the other processor products for AMD, and they are busy converting part of the Fab over to 90 nm production (which takes away space and production capacity).  So, currently AMD is not able to really address more than 20% of the market due to production constraints as is.  If they were to transfer all production to Athlon 64 then they would hardly be able to address 15% of the market.

90 nm is where the Athlon 64 needs to go, and all indications point to AMD producing such products in Q1 2004.  This of course could be a lot of wishful thinking on our part.  Intel, which has a lot more fabs and research money, has had a very hard time getting their 90 nm process up and running.  While Intel has told the world that they are on schedule and things are going great, it has been a very hard struggle for them, and it appears that internally they are around 6 months behind schedule with the delivery of 90 nm parts.  Prescott will see the light of day this year, but supplies will be VERY constrained.  Insiders are expecting a rollout much like the original Pentium 4 when it was released (as in not a whole lot of supply for the first 3 months).  AMD could be even farther behind, but no one is talking about it.  Luckily the Athlon 64 appears to have some headroom, at least to 2.6 GHz on the current 130 nm/SOI process.  AMD is still tweaking that process, and we will see faster models hit the floor soon.  Unfortunately, the 90 nm jump is a tough one, and AMD is adding SOI on top of it, so delays really should be expected in a 90 nm AMD part featuring SOI.  This is probably one of the reasons why AMD has teamed up with IBM because the technology jump is so incredibly tough, and both companies could use each others' help.  Intel could still be in trouble with its 90 nm process, but since they have two 300 mm wafer fabs to work on it, the impact will not be as big as it could be for AMD with its 200 mm Fab 35 (AMD's only processor fab).  AMD appears to be working on foundry agreements, but products from these collaborations typically run significantly slower than parts from the primary fab.

The Athlon 64 design is excellent, but the future is still very shaky for AMD due to the complexity of the part and the possible problems migrating to the 90 nm process.  The integrated memory controller for the Athlon 64 is probably the toughest part to design and get working on the processor, so any changes made to the current and future processes could have a devastating effect on the functionality of the memory controller.  If on-die memory controllers were easy to put on high performance parts, then many other companies would have done it by now.  There are reasons why products such as Geode never took off, as well as other products featuring the integrated memory controller.  It is much simpler to design one to run at 200 MHz than have one running full speed at 2 GHz.  It appears that the design is very solid for the Athlon 64 though, and hopefully for consumers AMD has its 90 nm process under control and available for Q1 2004 production of the A64 parts.

NVIDIA News and Insights - Josh

Many were expecting NVIDIA to take the wraps off of the NV36 and NV38 parts at Computex, but in fact not a whole lot was mentioned about them.  There were samples being shown at Computex, but none of them were running, and none of them benchmarked.  It now appears that the release will probably happen soon, but I can't say when exactly (within the next 2 months obviously).  There are some things that we can probably gather about the design of these products, and a few that we can guess on.  The most likely piece of technology at work here is that the NV38 and NV36 will most likely be the first products based on the IBM/LowK 130 nm process, and the first NVIDIA chips to come out of the East Fishkill Fab in New York.  These should run faster than the TSMC fabricated parts, and probably produce significantly less heat per MHz.  Of course, with the design of the new cooler on the NV38 samples shown, it will be running at a very high speed, and produce a goodly amount of heat.  If I were a betting man, I would guess the clockspeed for both parts would be in the 550 MHz to 600 MHz range for the Ultra editions.  Considering that the NV30 from TSMC ran at 500 MHz, this is not a wild guess.

There are also some indications that these parts could well be redesigned enough to partially make up the performance discrepancy between NVIDIA parts and ATI parts.  The NV30 was well underway and nearly finalized when the DX9 standard was agreed upon.  NVIDIA originally went at DX9 alone, and thought it may be able to force Microsoft to make its own design the basis for DX9.  During the initial stages of DX9 development, NVIDIA removed itself from the group developing the technology standard.  Only when DX9 was far into the development stages did NVIDIA rejoin the group.  By then DX9 was nearly finalized, and ATI had a very good idea what it would be like (and had been working on the R300 core since the basic inception of DX9).  NVIDIA was left with an underperforming part in floating point fragment programs, and they knew it.  The NV35 development showed that NVIDIA did realize it made some significant mistakes with the NV30, and the NV35 was designed to work around those problems.  NVIDIA is a smart company, and when the final specifications for DX9 were made official, NVIDIA knew they would run into problems in the future.  This was over 1.5 years ago, and during that time design changes to the NV36 and NV38 could be implemented to help this situation.  My belief (and it is only a belief) is that the NV36 and NV38 parts will be much better PS 2.0 performers than the NV30, NV31, NV34, and NV35 parts.  I do not believe that NVIDIA had enough time to change around the entire design to be as fast as ATI's R3xx series in such situations, but I believe that the new parts will at least be a lot more competitive using standard ARB code paths for Pixel Shader 2.0 operations.  This is only speculation, but some evidence pointing to this does appear to be out there.  In the meantime NVIDIA is doing major damage control with their products and the apparent lack of performance in PS 2.0 operations.  This damage control includes heavy driver and compiler level optimizations, working with developers to implement more NV3X friendly code, and keeping reviewers stocked with FX based cards to keep the good word out.  I could be very wrong here and the NV36 and NV38 parts are merely speed upgrades, but NVIDIA is due for a major product design change this fall.  It could be that the FX 5950 is merely named so as to confuse the competition, and in fact NVIDIA will release a major new part.  I guess we shall see.

The one thing we can bet on is that the NV40 will not have the problems the NV3x series has now.  PS 2.0 performance problems will be a thing of the past with the NV40, as it is supposed to have PS/VS 3.0 hardware.  Memory bandwidth will also not be a problem as GDDR3 will be available in late Q1 2004, and that promises around 51 GB/sec bandwidth on a 256 bit bus.  The NV40 is starting to look like the answer to NVIDIA's problems as of late.  ATI is also very far along with the R420, but it may not be as advanced architecturally as the NV40.  Still, not enough is known about either part to give further insight.  Needless to say, 2004 looks to be more competitive than 2003 was