AMD 7nm Vega graphics performance increased by 65% over Titan V? It is too expensive
This article has been reproduced by the Supernet. Other media reprints are subject to the approval of Supernet.
AMD launched its 7nm process GPU, the Vega 20, with 32GB HBM 2 memory at the Taipei Computer Show this year. It is expected to be launched in the second half of this year, mainly for professional markets such as deep learning. Considering AMD's lack of high performance for the past two years. GPU, 7nm Vega has also attracted a lot of high-end players, thinking about AMD if it is a 7nm Vega high-end game.
For the 7nm Vega graphics card, we still know very little information, especially its performance, the latest analysis 7nm Vega's performance has increased by 65%, floating point performance of 20.9TFLOPS, easy to kill NVIDIA's Titan V and Tesla V100 graphics card, but the cost of 7nm process is too expensive, can not drop.
For the 7nm Vega core, previous reports generally believe that it is actually a 7nm improved version of the existing Vega core, the core size has not changed, still 64 sets of CU units.
However, the Ascii website in Japan has made a unique analysis recently, and calculated that the number of CU units of the 7nm Vega GPU has increased significantly. Together with the 7nm process, it has a performance advantage, and its overall performance is improved by 65%.
The calculation process of the original text is quite complicated. Here is a brief description of their calculation process:
· First, calculate the core area of 7nm Vega. The size of HBM 2 memory is 7.75x11.87mm. The area of 7nm Vega is about 358.5mm2 (15x23.9mm). The rounding is calculated as 360mm2. The core area of the existing RX Vega 64 is 510mm2, so the core area of the 7nm Vega is reduced to about 70% of the current level.
After calculating the core area, they reversed the core scale based on this indicator, which was not mentioned by the people who calculated the core area before. Globalfoundries previously announced that the transistor density of the 7nm LPP process is 2.8 times that of the 14nm LPP process, except that AMD does not use the official library of GF, but uses its own cell library, which is twice the density of the 14nm process. Therefore, if the 7nm Vega is manufactured using the 14nm LPP process, the core area will be 720mm2.
· Estimated here will be a bit ambiguous. The purpose of using this counterexample is to prove that the core size of 7nm Vega is increased. Otherwise, the core area should not be reduced to only 70%. They think there are at least 88 groups inside 7nm Vega. CU unit, which is 37.5% more than the existing 64 units.
· Then, they reported the 7nm process performance data with GF - 40% increase in performance under the same power consumption to calculate the possible changes in GPU frequency. This process is also a bit complicated, because AMD's 7nm Vega is not only focused on performance improvement, but also Will optimize power consumption.
According to their estimates, 7nm Vega due to the significant increase in HBM2 memory, TDP power consumption may increase from RX Vega's 295W to more than 400W, but the trade-off, they believe that in the case of 20% increase in frequency, power consumption can be reduced by 40%.
In short, according to their estimation, the performance improvement of 7nm Vega is about 65%. The result is achieved by the core scale of 1.375x and the frequency of 1.2x, which is exactly 1.375x1.2=1.65, and the performance is improved by 65%.
As a result, the floating point performance of 7nm Vega will be increased from 12.66TFLOPS of RX Vega 64 to 20.9TFLOPS. This performance is very remarkable. At present, NVIDIA's strongest computing card Tesla V100 NVlink version has a floating point performance of 15.7TLOPS, Titan. The floating point performance of V graphics is 12.2 to 14.9 TFLOPS (acceleration frequency), so the performance of 7nm Vega graphics card will easily exceed NVIDIA's current computing cards and graphics cards.
Ascii's calculation process is very eye-catching. It seems that the performance of 7nm Vega has been underestimated before, but their calculation process is logically clear and may have a large loophole.The current 7nm Vega graphics card is generally considered to be produced by TSMC's 7nm process, and the calculated standard GF data may bring a large error.
So, the calculation here can only be a reference, which is different from the previous situation. If it is really guessed by Ascii, it can in fact prove that the 7nm Navi core is unlikely to be a high-performance core. Otherwise, AMD is a bit wasteful to do this, and at the same time developing two high-performance cores is a bit out of AMD's style.
Regardless of the performance of the 7nm Vega calculated by Ascii, it is right that one thing they said should be correct, that is the cost of 7nm Vega. The third part of the original article talked about this problem.
AMD chose the professional market instead of the consumer market to launch the 7nm Vega because the cost could not be reduced. The wafer price of the 7nm LPP process is at least 10,000 US dollars (the entire 12-inch wafer), which is twice as expensive as the 14nm PPP wafer.
Although the 7nm process has significantly reduced the core area by about 30%, the cost of the 7nm chip will still increase by more than 40% in consideration of the increase in manufacturing cost. In addition, the 7nm Vega uses 32GB of HBM 2 memory, which is the existing RX Vega. 4 times, the cost will increase greatly. These two factors add up, 7nm Vega can not be used in the consumer market in a short time.
Another factor is the capacity of GF. Currently, their 14nm process chips are produced at the Fab 8 plant in New York. The next 7nm chips will also be produced here. GF is also considering increasing the 7nm capacity, but they are short of money. GF had mentioned it before, and they had no complaints about AMD's allocation of 7nm orders to TSMC. They also adjusted the process parameters to close to TSMC's 7nm process, so that AMD could better use two different 7nm processes.
With so many factors combined, AMD is not much choice now, 7nm Vega can only be prioritized to the high-value professional market, and the game market is delayed.