Intel Core i9 (Gulftown) - 6 cores, 32 nm: hands-on review

The Gulftown core

Like Conroe (Core 2 Duo 65 nm) and Penryn (Core 2 Duo 45 nm), Gulftown has a whole lot in common with Bloomfield. In short, we can say that the Bloomfield has been given two cores and some third-level cache love. The capacity ratio of the L3 cache to the number of cores remains the same: Bloomfield has 4 cores and 8 MB of L3, Gulftown has 6 cores and 12MB of L3. The L1D, L1C and L2 cache capacity remains unchanged. Also unchanged is the processor's “make-up” - in addition to six cores in a single block of silicon there is a block of L3 cache, a triple-channel memory controller and two QPI links. One of them is used to communicate with the IOH system and with the rest of the system (X58 desktop chipset or 5520 server chipset). The other one is used for communication with the second processor on dual-processor machines, and it will be disabled in the desktop and workstation version (similar to Core i7-900 and Xeon W35x0). Intel has not yet shown a silicon wafer with the Gulftown cores or the processor itself with the core uncovered. According to unofficial sources, its size should be approximately 21.1 × 11.4 mm. The core surface area is probably around 240 mm2, i.e. about 10% less than the Bloomfield core (263 mm 2). This means that if the yield on Intel's 32-nanometer process technology is as good as that on 45-nanometer technology, the manufacturing cost of a single Gulftown is less than that of a Bloomfield. But let us not count on Intel's six core price tag to be low. Firstly, Intel's throughput in the 45-nanometer process is much higher – only one plant (D1D in Oregon) is ready to produce 32 nm circuits. Another three are being converted. Meanwhile, 45-nm systems are produced in three facilities. In addition, the Gulftown prices will be tailored to their target consumers – prepare to spend about 1,000-1,200 dollars for the desktop version. We still do not know what commercial name the Gulftowns will be given – various sources report it to be i9 Core or Core i7-1000 series.

Microarchitecture changes

It is true that the Tick-Tock strategy implies a change in the microarchitecture and updating the technology process, but when changing processes, Intel also makes certain improvements in the core logic. A good example of this was the Penryn processor family – although built based on the Conroe architecture, they featured such improvements as a faster division algorithm and SSE4.1 instructions. The most important and most publicized change to the Westmere compared to the Nehalem is a new AES-NI set of instructions (Advanced Encryption Standard – New Instructions). There are seven new instructions, six of which have been designed specifically to accelerate AES encryption. Encryption using this algorithm requires keys to be generated first, which is performed by the AESKEYGENASSIST instruction. The AESENC and AESENCLAST instructions perform further encryption steps, while AESDEC and AECDECLAST do the decryption. The AESIMC instruction makes decryption key generation faster. The last PCLMULQDQ instruction (carry-less multiplication of two 64-bit operands) is not directly associated with the AES algorithm, but also speeds up the encryption process. Details of the AES-NI instructions and the AES encryption algorithm can be found on an Intel website dedicated to this subject. Similar hardware-based encryption mechanisms were applied long ago now by VIA in their C3 and Nano processors. The AES-NI and PCLMULQDQ instructions will also be implemented on all Westmere processors, including the desktop Arrandale and the laptop Clarkdale. They are also to be implemented in AMD’s future processor architecture (codename: Bulldozer).

The second of the changes relevant to users is the change in the operation of the memory controller. As in the Lynnfield processors (Core i5 and Core i7-800), the 1 uncore timing part memory clock speed limit does not apply any more. While the uncore in a Bloomfield has to work with a 4000 MHz frequency so that the memory can reach a clock speed of 1000 MHz (DDR3-2000), three times the memory clock is enough for the Gulftows and the Lynnfields. In our test item, the uncore operation frequency was 2400 MHz and could not be changed due to early BIOS versions. Intel has not confirmed yet whether the Gulftown processors will provide official support for memories faster than DDR3-1333, but it can be expected that with one module per memory channel, the fastest official speed will be DDR3-1600.

Intel Gulftown ES

The particular system we had the opportunity to test was a server version with default 3.07 GHz clock speeds and two QPI interconnects. The TURBO mode allowed two, three or four cores to be sped up by 133 MHz and one core by 266 MHz.

Gulftown processors are backwards compatible with the LGA 1366 socket. The only thing you need to do to run them on today’s motherboards is a BIOS update. We have managed to get Gulftown to work on three boards: Gigabyte EX58-Extreme, ASUS Rampage II Gene and ASUS P6T SE. Unfortunately, early versions of BIOSes with Gulftown support have their problems too. All of the boards had only two memory multipliers available (× 8 and × 10, or DDR3-1066 and DDR3-1333) and one uncore multiplier (× 18). In addition, the memory subsystem bandwidth is lower than expected. Similar problems could be encountered prior to the premiere of the first Core i7. We expect that this will improve in subsequent BIOS versions.

The latest versions of popular diagnostic programs do not have any major difficulty identifying the processor model and features. HWiNFO32 even displayed the presence of AES-NI instructions:

Test assumptions

The processor we had the opportunity to test is not a commercial product – it is an early engineering version which may be very different from the commercially available end-product processors. For this reason, we have not tested it with the default operation parameters. Instead, we opted for a clock-to-clock comparison. For the shootout, we took major quad-core processors available on the market. They all operated at 2800 MHz. On all platforms, memory clocking was DDR3-1333 and latency was 7-8-7-21 1T. Turbo Mode for Core i7, Core i5 and Gulftown was off. All the power-saving features such as EIST and Cool'n'Quiet were on. The scores labelled ‘Core i5’ have been obtained using an Intel Core i7-860 processor with Power HyperThreading enabled, meaning they are different from the real Core i5 (where the uncore clock was at 2133 MHz). Note that due to the early Gulftown BIOS versions and the early Gulftown processor versions, the results are only approximate and may not be representative of the final product.

Testing Rig

For the testing platforms, we used the following components:

The computer was running Windows Vista Ultimate 64-bit Edition with Service Pack 1.

Synthetic tests – Sandra

We started with synthetic tests in Sandra 2009:

In arithmetic and multimedia performance benchmarks, the results increase almost linearly with the number of threads. Gulftown is about 50% faster than Core i7.

What is important for wPrime is above all pure computing performance, and again – Gulftown ahead of Intel Core i7 by half.

In addition, scaling in the new version of the OMP Test, also featured in previous reviews, is almost perfect. The number of threads has a direct bearing on the results.

Then we checked the memory subsystem performance. Remember that the third-level cache and the memory controller on Gulftown operated at lower clock speeds than on Core i7 and Core i5.

It is evident at first glance that there is something wrong with the results. Apparently, even the latest version of Everest is still unable to measure the bandwidth and memory-access latency correctly. To make sure that the memory subsystem is working properly, we measured the time to count 16 million digits of pi using Super PI software:

Super PI is not exactly designed for comparing processors with different architectures, but it is very sensitive to any irregularities in memory access. Despite the greater L3 cache capacity, Gulftown’s score is poorer than that of Core i7. The difference is greater than that implied by the difference in uncore timing – it is caused by early BIOS versions that run 32-nanometer six-cores. The same memory slowdown effect occurred on all three LGA 1366 motherboards tested. It is expected that as we are closer to the release and commercial availability of the Gulftowns, the BIOSes will be refined and memory access will be a tad better.

Real-world tests – video encoding, rendering

Another benchmark was to render a test scene using Cinebench R10.

In single-threaded mode, all Nehalem-based processors obtain similar results. In multithreaded mode, the 32-nanometer six-core has about 33% more performance than the eight-thread Core i7. Why not 50%? The Cinema 4D engine takes a long time to complete a thread and start a new one, which means that for part of the time, not all 12 threads are being used. As the number of threads grew, scaling would keep deteriorating.

We checked how the six-core can handle ray-tracing under POV-Ray.

There are no surprises – ray-tracing is one of the applications very susceptible to parallelization of calculations. Scaling and thread number are close to ideal, at approximately 46%.

We also examined the speed of video conversion from MPEG-2 to x264.

In the first run, the workload is not distributed evenly among a large number of threads, but still, Gulftown gets a 20-percent advantage over Core i7. In the second run, in which tasks are assigned to threads more efficiently, once again, half more cores gives Gulftown a result better by half.

Real-world tests - Photoshop, file compression, encryption

Another benchmark was to perform 15 operations on a 38 megapixel image in Adobe Photoshop CS4.

Most of the filters and image transformations in Photoshop are not divided into multiple threads. More important than pure computing performance is memory bandwidth, especially witch such large images. Our Gulftown had its memory bandwidth limited by an early version of the BIOS, but was still ahead in the Core i5 test.

File compression and encryption

We tested simultaneous file compression and AES encryption. One of the few software applications that already use the AES-NI instructions is the latest version of WinZip. When creating archives In WinZIP 14, you can choose the compression type, encryption type and strength.

Simply select any AES encryption – hardware-based encryption acceleration will work immediately if you have an AES-NI enabled processor.

In encryption-free compression, Gulftown is not much faster than Core i7. In addition, WinZip compresses using a maximum of two threads, so the Gulftown’s entire advantage is related to its greater third-level cache capacity. When you turn the encryption on, Gulftown’s advantage becomes evident. Enabling encryption on non-AES processors results in compression times longer by 160-200%. On Gulftown, this causes compression to take 120% longer. Let us remember that in addition to encryption, the processor has to deal with file-packing – in typical cryptographic applications the advantage of using AES instructions will be much greater.

Synthetic tests – 3DMarks

And finally, what probably most readers pay attention to – 3D graphics tests. As usual, we started with 3DMark06.

You did not expect anything else, did you? When generating three-dimensional graphics, the two additional cores have virtually no impact on the performance. On physics and artificial intelligence benchmarks, the scores are better, but they too do not scale linearly with the number of threads.

Then, we measured performance using 3DMark Vantage.

Gulftown’s score on physics and AI tests is impressive, but its weight in the final result is only 0.15. In real games, the gain with a multi-threading processor is even smaller.

Games

3D benchmarks continued – games.

The four extra threads give a slight advantage to Gulftown over Core i7. As shown in the simulation of particle effects in the Source engine, some parts of the game code can be parallelized with good effect. But Left4Dead based on the same engine clearly shows that this is of little help in gaming – the number of frames per second is determined by the rest of the code, unfortunately most often single- or dual-threaded. The LGA1366 processors’ low score on Left4Dead is an anomaly resulting from an error in the Nvidia drivers and does not occur any more with the latest driver versions.

Power consumption

We measured the power consumption of the entire test rig. Keep in mind that it is different from the one used in the recent processor reviews and the results are not comparable with the older ones. The processor, the main character of this article, is not the final product – it is an early engineering copy running on a board that provides unofficial support. The Power consumption may not represent the Gulftown processors we will see appearing in stores in a couple of months.

Our engineering copy was powered with 1.11 V under load and 0.95 V at rest.

Gulftown has about half more cores and half more cache than Core i7, and under load it draws ... about 24 watts less. As you can see, the 32 nm process technology has perfect effects on the power efficiency of processors, if you can pack 50% more transistors into a much smaller energy budget.

Operating temperature

We also measured the processors’ operating temperature. The temperature was measured using a digital thermometer with a K-type probe located on the side of the IHS.

Gulftown is one of the coolest processors we have tested so far. The temperature under load is comparable to processors that have two or three times less cores.

Overclocking

Of course, we could not have missed overclocking. Unfortunately, limited time and issues with early BIOS versions did not allow us to carry out a full test. However, we can show you a few screenshots to whet your appetites:

In view of the recurring questions in the comments, we also decided to play Crysis:

The voltage in the screenshots is overstated – in fact it was at 1.35 V. Based on our short experience in overclocking a 32-nanometer six-core, we can venture a conclusion that the temperatures will not be a limiting factor for maximum overclocking in the air. What could be a limit is rather the voltage – the maximum frequency for stable operation scales with voltage much better than on Core i7. Unfortunately, we have no knowledge of degradation on these processors under the influence of high voltage. Around 1.4 V is probably the highest reasonable value of when air- or water-cooling. We will try to come back to overclocking this processor (including extreme overclocking).

Summary

The Gulftown processor can be briefly summarized using a military concept: shock and awe. Twelve threads and huge bandwidth put on a show of incredible computing performance. It is just that if Gulftown is the shock and awe, most readers are the citizens of a neighboring country that does not take part in the conflict. In domestic applications, with gaming on top, the four extra threads are of little importance to the efficiency and comfort. Core i7-900 is not the best processor for gamers and the new six-core is not much faster clock-to-clock. Gulftown is a processor for servers and workstations, not for gamers.

The average John Doe will probably be much more impressed by the 32 nm technology it was made with than the processor itself. Smaller, faster and more efficient transistors will allow Intel to build a system with significantly lower power consumption and higher power efficiency than today’s processors. Made using the 32-nanometer process technology, the dual-core Clarkdale processors are likely to appear on the shelves sooner than the Gulftown, and are bound to give a much mightier shake to the market than the high-end six-cores. We hope this article has given you some idea of what we can expect from Intel in the near future.