An Athlon 64 X2 By Any Other Name…
This week AMD released the long awaited Kuma, and while performance is surprisingly competitive with what Intel has in the market, the real story is how this strange little processor came to market.
The road to Kuma has been a long and winding one, but ultimately worth the wait. Sorta. This is honestly one of the oddest releases in AMD’s history, and the background of this part is probably far more interesting than the performance it brings to AMD’s dual core offerings. But before we head into the strange story of the processor codenamed Kuma, let us take a look at the technical specifications of the product.
The Athlon X2 7750 Black Edition. Coming to a motherboard in Ryan’s lab soon.
The Athlon X2 7750 and 7550 are based on the Phenom architecture rather than the classic Athlon/Athlon X2 series of chips that preceded them. This means higher IPC than the older Athlons, as well as some new features that could be interesting to some users. Probably the most interesting feature is that of hardware virtualization support, which is a major upgrade in AMD’s dual core market. Another feature is official support for DDR-2 1066 memory, which should bring some interest for the product from budget enthusiasts.
While AMD has codenamed this product Kuma, it appears as though it is just a standard B3 revision Phenom X4 with two cores disabled. The transistor count and die size numbers that AMD has given with this launch are identical to the full Phenom X4 and X3 products already on the market. As such the two working cores each feature 128 KB of L1 cache (split between data and instructions), 512 KB of L2 cache, and the full 2 MB of L3 cache shared by the two cores. “Kuma” is still a 65 nm product with a TDP of 95 watts. The wattage is not exactly impressive as compared to its older 65 nm Athlon X2 brethren, which mostly come in at 65 watts as well as the low power 45 watt versions. Considering that the Phenom 9950, which runs at 2.6 GHz, has a TDP rating of 125 watts, then the 95 watt rating does not look quite as bad.
The 7750 runs at 2.7 GHz and is available in both regular and Black Edition flavors. The BE of course is unlocked, so people wanting to try their hand at overclocking can give it a good shot with minimal complications. The 7550 runs at a slightly more sedate 2.5 GHz. Both products have their “northbridge” functions running at 1.8 GHz, which gives the HT 3.0 connection a speed of 3600 MT/s.
Performance appears to be right around the Intel Core 2 Duo E8200 processor (2.66 GHz) level, if a little slower in most applications. This is certainly good news for AMD, as they have not had a processor be competitive in the dual core arena in some time with what Intel has to offer. The big shocker with this release is that the 7750 is to be priced around $79 (in 1K quantities), which means that it competes with the Intel E5200 in price, which is clocked at 2.5 GHz and has 2 MB of L2 cache total. In head to head comparisons, the 7750 does outperform the E5200 in a large majority of applications that I have seen tested. To consistently outperform the 7750, the E8200 would be the best choice, and it is around $80 more expensive than the 7750.
Overclocking seems to be about average when looking at the latest B3 revision Phenoms. Most are reporting going easily to 3.1 GHz and 3.2 GHz, and with a small amount of effort getting to 3.3 GHz. Anything above that though is a crapshoot, and people with more extreme cooling will be able to hit 3.4 GHz to perhaps 3.5 GHz. Getting 600 extra MHz with little effort is a positive selling aspect, but considering that most of the 45 nm Intel products can hit 4 GHz with few issues, AMD has a ways to go.
While not exactly accurate… close enough.
Looking past performance and specifications, I find it far more interesting to trace the development of this product. When the original Phenom news was released, there was mention of a dual core version that would eventually be released, but it would be 6 to 9 months after the initial quad core Phenom introduction. It was also going to be a separate dual core design with a unique die as compared to a quad core die with two cores disabled. As we were approaching the release of Phenom, there was a lot of discussion as to why AMD pushed the native quad core instead of focusing on the dual core variant, where bugs might be easier to detect and eliminate. For many it just seemed more logical to work out the issues on a smaller die with a smaller transistor count than do the same thing with something far more complex. It also would have been easier to fabricate as initial yields would have likely been higher due to the smaller die size. Power consumption and clockspeed issues would also have been minimized due to the smaller design and lower transistor count.
AMD had a tough time with Phenom, and the multiple revisions required to even get to release quality B2 silicon (nobody outside of AMD knows how many A stepping revisions there were, but we certainly know there where 3 major revisions with the B stepping), and many question if they would have run into the issues they had while pursuing the less complex design and applying those lessons to the native quad core. After the B2 revision/TLB disaster for AMD, word about Kuma was confused and mixed. There were initially talks of releasing Kuma in Q2 of 2008, possibly around a May/June timeframe. Then there was some very specific rumors about Kuma being cancelled altogether and the only dual core Phenom based product would be at 45 nm and would be at least a year away from what was supposed to be the initial release.
So at the beginning of Summer of this year, all talk about Kuma had essentially died. Most assumed that the processor would not hit the market, and it was in fact cancelled. Then a funny thing happened. About a month later new rumors about dual core Phenoms started to hit the street. First indications and unofficial plans called for a dual core product that would be clocked around 2.3 GHz and would be labeled as the Athlon 6500+. Some websites even leaked some supposed part numbers and prices, but these products obviously never came about.
Now we finally see what AMD has been up to, and the results are actually quite surprising. The 7750 and 7550 parts are very competitive in their price ranges, and they give AMD’s dual core offerings a much needed boost. The bad news about this release is that the cores being used do look to be native quads with two cores disabled. We also have the not so friendly 95 watt TDP for a part running at 2.7 GHz (though that particular rating is “worst case”).
The big question here is if AMD has been stockpiling partially defective cores, or are these products simply being disabled to fulfill a market category. Historically speaking, one would expect to see these partially defective cores showing up at the beginning of the product’s introduction. ATI had done this with the R300 chips (Radeon 9700 was the full version, and the Radeon 9500 had one half of the shader pipelines disabled), and it was a success for them… until the fab started improving yields on fully functional parts and ATI had to disable perfectly good GPUs to fit into a market category, thereby losing out on potential revenue by not selling the GPU as a fully functional part at a higher price. Considering that the 7×50 series is based on the B3 revision, and we have been seeing the quad and triple core products since Q2 of this year, it is hard to imagine that these dual core Phenoms are partially defective in some way.
The above die is of the 45 nm quad core Phenom, so we wonder if AMD will still salvage dies by turning them into dual core variants? In this case, I honestly wonder if AMD would not be better off doing a native dual core with 3 MB of L3. That would be a significantly smaller product than the 260 mm square full quad core.
One other possibility is that these are primarily fully functional quad cores, but the power draw and clockspeed limits for individual cores might make them unsuitable for quad or triple core use. An example would be two of the cores can run at 2.7 and 2.8 GHz with tolerable power draw, but the other two cores have a limit of 2.3 GHz with a much higher power draw than expected for that clockspeed. These would be chips that would be unsuitable for the 65 watt 9350e or the 8650, but by disabling the two lesser cores a perfectly fine dual core could be used that runs at 2.7 GHz with acceptable power draw.
Something else to consider is how much actual real estate would AMD have been able to save per die by creating a new native dual core design with 2 MB of L3 cache? If you scroll up to the second picture in this article, you can see how much die space each core takes up compared to the other “non-core” elements. These elements are the memory controller, Hypertransport controller(s), and the large L3 cache. If you take two cores out of there, and reconfigure things, the total die space would probably be reduced by about 1/3. I am betting AMD took a good long, hard look at development costs vs. die space savings, as well as yields and how the product will be placed and priced in the market. Once the final answer came back, it likely was not all that pretty and AMD was looking at actually saving more money in the long run by just disabling cores on fully quad core dies.
AMD certainly has not hit a homerun with the 7×50 series of chips, but they are addressing a market that is still important. Most computers sold today are still dual core machines, and the price/performance for most applications users run still favors dual core products heavily. With AMD offering new products with improved functionality and better performance, we now have better choices coming this holiday season. AMD also has an advantage with their integrated chipsets over Intel, both in price and capabilities. So that, hand in hand with the new processors, may sway more than their fair share of users to give the 7×50 series a chance.