When Nvidia released its latest video card, TITAN X, the ultra-high standard hardware and the price of 999 US dollars (about 6230 yuan) attracted a lot of attention. For those who pursue the ultimate graphics performance, this current strongest Nvidia graphics card is absolutely unmistakable.

The new Titan X has added a new Maxwell GPU, GM200 (also known as Big Maxwell). From a macro perspective, The GM200 can be seen as the product of pushing the GM204 to the limit. Compared with the G204, it is larger, faster, and naturally consumes power. A total of 8 billion transistors occupy an area of 601 square millimeters, making it the largest GPU in Nvidia's history. This is also the first time that Nvidia has used every inch of available space to push out the strongest performance for a long time, In addition to Maxwell's efficiency advantages, its strength cannot be underestimated.

Generally speaking, GM200 has a 50% improvement over G204 - it has 50% more CUDA cores, 50% more bandwidth, 50% more ROP, and even 50% more area. After joining GM200, TITAN X has 3072 CUDA cores, 192 texture units, and 96 ROPs in total. Recall that the performance of the GTX980 using GM204 was easily superior to that of the GTX Titan and GTX 780Ti using GK110, and this performance advantage was basically due to the improvement of Maxwell's architecture - after using the Maxwell architecture, Compared with the Kepler architecture, the performance of the CUDA core has been improved. In short, the performance of the newly produced TITAN X over the existing TITAN and GTX980 is a certainty.

In order to feed the GM200, Nvidia uses a 384bit memory bus to drive 12GB GDDR5 video memory at 7GHz. This paper specification is slightly lower than GTX TITAN BLACK. However, given Nvidia's efforts to research and develop memory compression technology on Maxwell, the existing 336GB/s memory bandwidth can transmit more data. As for the size of 12GB video memory, it is also the largest among NVIDIA video cards at present - this is very consistent with the tradition of NVIDIA video memory. 12GB video memory should be quite abundant in the next few years. At the same time, there is also a relatively massive 3MB L2 cache between the functional unit of the GPU and the memory bus, maintaining the same 32K: 1 cache/ROP ratio as Maxwell 2. This increase in cache size can naturally reduce the workload of the memory bus.

In terms of frequency, the basic frequency of TITAN X has been raised to 1GHz (1002MHz according to Nvidia's software tool test), and the maximum can reach 1075MHz. Compared with GTX TITAN BLACK, there is an increase of 100MHz (that is, about 10%). Naturally, this setting is also to improve performance, but compared with GM204 (GTX 980), which has a smaller total area On the contrary, the frequency drops slightly, so the overall performance difference at the maximum frequency is about 33%.

As for power consumption, Nvidia has not disappointed us either - according to Nvidia's official data, The TDP of TITAN X is 250W, which is the same as that of the previous generation of single core TITAN series. Compared with the GTX980, the power consumption of TITAN X is 51% higher, which is understandable. After all, there are so many transistors there. In general, the power consumption of TITAN X is quite large, which is the price of obtaining top performance energy.

As for comparison with competitors, to be honest, TITAN X has no competitors, because the last time AMD released a high-end graphics card was in 2013, and the GTX 980 has been stronger than the R9 290X. As for TITAN X, it just further expands Nvidia's leading edge. At present, there is no single core GPU in the market that can compare with it.

Therefore, the only real threat to TITAN X is the multi-core settings such as GTX 980 SLI and R9 295X2. In particular, the price of the latter is $699, which makes TITAN X slightly uneasy. However, in terms of stability, the single core GPU is still better than the multi-core GPU.

Nvidia is also well aware of this. In the absence of competitors, plus the ultra high-end positioning of TITAN X, Of course, the price of TITAN X is also quite expensive, maintaining the previous listing price of $999 for GTX TITAN and GTX TITAN BLACK. From our perspective, we naturally hope that TITAN X can lower its price a little. It would be great if it were about $700. But for Nvidia, if it sells $999, many people will buy it. Why not make money?

As the core of TITAN X, the GM200 is certainly worth seeing. To be honest, it is very conscious. To sum up, its excellence lies in its non excellence.

First of all, from the perspective of semiconductor manufacturing process, the GM200 has maintained the setting of 28nm. Although 28nm has entered its fourth year, it naturally starts to encounter bottlenecks. For bottlenecks, there is no need to say more. For manufacturers, it means that they need to spare no effort to squeeze the last bit of performance and efficiency of 28nm before upgrading the manufacturing process.

  First of all, Nvidia uses the Maxwell architecture to replace the Kepler architecture, which improves the energy efficiency ratio without changing the manufacturing process. Nvidia's GM204 does not increase the power consumption without increasing the chip area, so the 294 mm2 and 3.5 billion transistors of GK104 become 398 mm2 and 5.2 billion transistors on the top of GM204, and after increasing the area, Nvidia can also place a considerable number of high-performance Maxwell SMMs.

On the other hand, Nvidia has also made GM206 (GTX 960), cutting GM204 into two. Although the performance of GM206 is not much better than that of GK106, it greatly reduces energy consumption and takes into account manufacturing costs.

However, for TITAN X, these two paths are impassable, because GK110 itself has a huge area (551 square millimeters), and it is obviously unrealistic to force 33% of the area to increase like GM204. If the architecture is replaced by Maxwell without changing the direct area, the performance improvement is also very limited, so Nvidia chose the third path, The third way makes the GM200 very interesting.

First of all, Nvidia chose to abandon the high-performance FP64 unit, because Big Kepler's graphics performance is already considerable, but it still spends a considerable part of the chip area on the FP64 CUDA core, which allows Nvidia to use the same GPU design in the entire product line, but this also means that such a design is a technology learned by rats, Especially when facing a new wave of 28nm products and hoping to use the power saved by switching to Maxwell architecture in more useful places, it is meaningful to do so. So on TITAN X, Nvidia has created an enlarged version of G204, not the successor of Big Kepler.

The total area of the GM200 is 601 square millimeters, which is the excellence of the GM200 - it is just large. It does not add any function that makes Tesla and Quadro series users drool. It is a GM204 that has increased by 50%. That is to say, the design of the SMM is the same as that of the GM204. Each SMM has 128 FP32 CUDA cores, a 512Kbit registration file, and only four FP64 ALUs, The proportion of original FP64 converted is only 1/32, and the extra space of those reduced FP64 is full of the ALU of FP32 and other things. As for how much space was saved, Nvidia refused to announce.

In other words, this "purification" means that Nvidia can maintain the same power consumption on the basis of GK110, improve the performance by 50%, and increase the chip area by only 9% (50 mm2). The GM200 also seems to have fulfilled Nvidia's long cherished wish - the last GPU using this design idea is the G80 (GeForce 8800GTX), However, the G80 does not have any FP64 related hardware at all. It can be regarded as the "purest" GPU in 9 years.

In this way, the GM200 has achieved superior performance in the FP32 category (as for FP64, FP32 is used for compatibility), and we are quite sure that Nvidia will not launch products with performance that exceeds or approaches that for quite some time in the future.

thus, GTX TITAN X It should not be that the original GTX TITAN is a product targeted at consumers (consumers in production activities). That year's TITAN is not only Nvidia's flagship, but also their entry-level FP64 computing card. TITAN X will certainly not have such a dual identity this time. After all, GM200 sacrificed FP64 performance in exchange for graphics and FP32 performance. In short, TITAN X is more inclined to the consumer market of entertainment graphics cards than TITAN.

But now the question arises again. Is Nvidia's practice of clearly distinguishing the functions of the GeForce/Quadro and Tesla product lines constrained by the manufacturing process and forced to compromise, or is it the basic idea of the whole product line in the future? Perhaps in the future Pascal architecture, NVIDIA will return to the previous design of "everything can be" with the help of the 16nm FinFET process. Personally, the GM200 is quite likely to be the cornerstone of the GeForce series and GTX TITAN series in the future.

From the appearance of TITAN X, TITAN X follows the design of the TITAN series, adding a redesigned magnesium free all metal housing, a high-speed fan and a grille type heat sink. As a result, this 250W high-performance graphics card is quieter and more elegant than the previous product.

Open the shell and you can see the huge radiator inside. The radiator and GPU are connected through a copper heat conduction chamber to obtain the best heat conduction efficiency. The outer surface is covered by a black aluminum substrate, providing support and protection for all internal components.

Further, we finally saw the card body, with a large GM200 GPU in the center, surrounded by VRAM chips, and other various components, The PCB design of the GM200 does not completely follow that of the GK110, but the changes can be counted by one hand, that is, the GPU and VRAM are on the front, and the other components are on the back (the back of the TITAN X is completely exposed). In addition, there are some minor modifications that affect the air flow to improve heat dissipation.  

Like GK110, Nvidia has followed the design of 6+2 phase voltage regulation module, where 6 corresponds to GPU and 2 corresponds to VRAM, which means that TITAN X also has a little room for overclocking - the maximum allowable power consumption of Nvidia is 275W. To be honest, this is not a big space for overclocking maniacs. Based on our evaluation of TITAN X, The maximum high-frequency voltage is 1.162V, so it should be feasible to increase the voltage after releasing the TDP limit. The maximum voltage of our card can reach 1.237V.

As for the interface, TITAN X, like GTX 980, has one DL-DVI-I, three DisplayPort 1.2, and one HDMI 2.0, with a total of four ports. HDMI 2.0 is enough to drive a 4K resolution display. If you want to drive multiple 4K displays at the same time, TITAN X also provides SLI play.

Next, start the performance test. The test is as follows:

The system built is as follows:

Field 4:

The first one is Battlefield 4, the FPS game launched by DICE. The following frame rates are recorded in the single player mode. According to our experience, the frame rate in the multi player mode should be halved, so if you want to hold the multi player mode, you must reach at least 60fps in the single player mode.

After the anti aliasing of the frost engine in Battlefield 4 is turned off, the frame speed of ultra-high image quality in 4K resolution has reached 58.3fps. Although it does not reach the standard of 60fps, the gap is not large. So in the multiplayer mode, TITAN X should be able to maintain a frame rate of about 30 fps, and even if the frame rate drops, it will not be too outrageous. If it can accept medium image quality, then the performance of 94.8 fps per person/45 fps per person is naturally very smooth.

In terms of performance, TITAN X is about 33% better than GTX 980, which is almost the same as the performance gap we estimated before. Although the execution unit of TITAN X is 50% more than GTX 980, the frequency is 12% lower than GTX 980 after all; Compared with GTX 780 Ti and TITAN, the performance advantages under 4K are 50% and 66% respectively. After all, it is unrealistic to double the performance improvement under the same manufacturing process. However, NVIDIA can increase the performance in such a way on the basis of maintaining the power consumption and increasing the area slightly. As always, it is worthy of our admiration.

Compared with the competing R9 290X/290XU, the performance advantage is about 50%~60%. At the top of the scoreboard, the R9 295X2 and GTX 980 SLI show the performance advantage of multi-core GPU settings; Finally, for those players who are satisfied with 1440p, TITAN is the only single core video card with a performance of more than 60fps. You only need to know this.

Island Crisis 3:

This game, "Island Crisis 3", read as "Graphics Crisis 3", is still a video card killer. In those years, Island Crisis 1 was the template for the performance requirements of the game in the next few years, but in the future, Island Crisis 3, it still is.

Middle Earth · Mordor Shadow:

This game is characterized by the high occupation of video memory. In this game, the graphics cards of 3GB video memory have fallen down, and only those with more than 4GB video memory can still fight. The multi-core GPU also performs generally here, showing the consistency and difference of performance between multi-core settings and single core settings.

Civilization · Space:

Dragon Century · Trial:

Talos' Law:

Isolated Island Horror 4:

Attila Total War:

Starting motorhome race · motor sport:

Synthetic performance:

Computing performance:

Power consumption, heat generation and noise:

First of all, starting from the voltage, we found that NVIDIA is quite conservative in setting the voltage, which should be the result of power consumption control. The voltage of TITAN X is only 1.162V at high frequency, much lower than that of GTX 980. The voltage difference is reasonable corresponding to the main frequency difference.

Now that we talk about the dominant frequency, let's compare the frequency differences between the TITAN X and the GTX 980. It is not difficult to find that the TITAN X cannot increase its operating frequency at will due to the limitations of power consumption and heating. Although the highest dominant frequency in each game is different, there will be a 50-100MHz difference between the TITAN X and the GTX 980.

In terms of static power consumption, TITAN X performs very well. Even if 8 billion transistors are inserted, the static power consumption is still low, but it is relatively higher than other single core cards.

Among the working energy consumption, the performance of TITAN X is basically in line with expectations. The power consumption is roughly the same as that of GTX 780Ti. Compared with GTX 980, there is a certain increase. After all, from the perspective of data, The TDP is 85W worse, and our measured results are almost the same as the data. At the same time, the power consumption of TITAN X is similar to that of R9 290X UBER. That is to say, NVIDIA's 4K performance is 44% higher with the same power consumption, showing the advantage of N card efficiency.

Since there is no difference between TITAN X and TITAN in terms of heat dissipation design and power consumption, the natural heating situation is also equivalent, and the performance at 32 ℃ under static conditions is good.

When processing graphics, the upper limit of the original temperature of TITAN X is 83 ℃, and then it will trigger overheating frequency reduction, but the frequency reduction will not be too much.

Finally, the noise level. Thanks to the heat dissipation design of TITAN, the performance of TITAN X at 37 dB in static state is satisfactory. In fact, TITAN X is the quietest graphics card in static state in this test.

As for the working noise, the performance of TITAN X and 780Ti is surprisingly mild again. Although the noise is much higher than that of GTX 980, we think the noise control is still excellent for this level of performance.

In general, from three aspects of power consumption, temperature and noise, TITAN X performs very impressively. Thanks to the Maxwell architecture and TITAN's thermal design, NVIDIA has squeezed out stronger performance while maintaining the same power consumption and noise level.

Overfrequency:

In terms of basic design, TITAN X has basically reached the limit of the current design. Nvidia has set the maximum power limit of 275W, as previously said, there is no room for improvement. However, since the voltage of the original factory is not high, it should be feasible to overclock by pressurizing. However, whether it is to increase TDP or voltage, In the end, it did not reach the essence of the overclocking game - it did not change the voltage but just increased the frequency setting to push the entire frequency curve higher.

Even if 8 billion transistors are inserted, The GM200 still has a good overclocking space. In general, the overclocking amplitude can reach 200MHz, the video memory overclocking amplitude can reach 800MHz, and after the overclocking, the maximum frequency of TITAN X can reach 1.45GHz. Although the setting of TDP makes it impossible to maintain such a frequency for a long time, it is gratifying to see that TITAN X has such potential at least.

Of course, the cost of overclocking is the increase of power consumption and noise, and the increase is not small. When the power consumption increases by 25W and the noise reaches 55dB, the overclocking achieves 16% performance improvement. As for whether it is worth it, let's talk about it.

Conclusion:

In general, TITAN X is naturally the strongest game graphics card on the market at present, with an overall increase of 33% compared with GTX 980. Nvidia has indeed pushed the existing design to the limit. Compared with GTX 780Ti and TITAN, TITAN X also has a performance improvement of 50%~60%, and the most important thing is that such performance improvement does not come at the cost of significantly increasing power consumption and noise.

On the other hand, technically, TITAN X represents a change in Nvidia's high-end GPU design philosophy. By increasing GM204 and shifting its performance orientation to graphics, it has created the strongest game graphics card. The 601 square millimeter GM200 is also Nvidia's largest GPU ever. At the same time, it has pushed the current 28nm process to the limit. But will Nvidia continue to use this design philosophy in the future, We should also see the effect after the event.

In terms of the market, it is not surprising that TITAN X maintains the original price of $999 for TITAN. TITAN series graphics cards are originally positioned at a high end, and the lack of competitors means that Nvidia can charge exorbitant prices.

As for multi-core graphics cards, the GTX980SLI and R9 295X2 show extraordinary strength, but to be honest, multi-core GPUs are just a way out on the premise of not improving the performance of a single core. The R9 295X2 has a high cost performance ratio, but for the stability and durability of the performance, we prefer to set a single core GPU. Of course, the premise is that the money is not bad.

Finally, if we regard TITAN X as the signal that the TITAN series graphics card is separated from the entry-level double precision computing card, the final positioning of TITAN X, which is separated from the positioning of the producer and consumer, is nothing more than a plaything. Maybe Nvidia will change in the future, and perhaps continue to introduce entry-level FP64 computing cards and high-performance graphics cards in the future, but TITAN X, That's it.

Original text and pictures from