Updated (1/11/2019, 6 pm EST): An AMD spokesperson informed us that the widespread GPU configuration for Radeon VII is incorrect. The Radeon VII does not have 128 ROP. Like Vega 64, Radeon VII is a 64 ROP card. The reports on a 128-POR design were prepared by several reputable sites and journalists present at the CES. It is not clear why this information was reported incorrectly or how the error was introduced.
ExtremeTech regrets the error. All discussions and speculation regarding a 128-rule model in the story below should be considered incorrect. Our story on the correct configuration of the 64 ROPs can be read here.
Original story below:
The announcement of AMD's Radeon VII earlier this week was a surprise. The company has already evoked its Radeon Instinct hardware family and unveiled its MI50 and MI60 statistics at its Next Horizons event, without suggesting that it was considering launching a new high-end Vega GPU. The company has now shared a little more information on the map, including its own performance data.
Radeon VII is based on Vega 20, the 7 nm matrix removal built at TSMC. The chip size is 331 mm squared for 13.2 billion transistors, compared to 487 mm2 transistors and 12.5 billion for Vega 64. The overall improvement in density appears to be 1.56x and AMD claims that its Performance per chip zone has been improved by 1.8x according to THG. Although not insignificant – matrix performance has implications for die size and thus manufacturing cost – it is not very relevant to the question of how well Radeon VII will compete with RTX 2080, its future rival.
The graphics processor has 60 CPUs, which corresponds to 3,840 graphics processor cores. The base clock is 1.45 GHz, with a boost frequency of 1.8 GHz, offering 13.8 TFLOP single-precision performance. The clock boost is an improvement of 1.16x compared to the maximum frequency of the Vega 64, just like the basic clock.
GCN has never been a high-speed architecture, despite the various revisions and improvements of AMD. The shift of 7 nm does not increase the number of clocks as much as the transition between 28 nm and 14 nm, although the difference is small: Vega 64 has been multiplied by 1.12 times faster than Fury X, while that Vega VII increased by 1.16x compared to Vega 64.
The main configuration is 3840: 240: 128. In terms of active cores, it's halfway between the Vega 56 and the Vega 64, but the output count has doubled to 128. The amount of memory embedded has doubled to 16 GB of HBM2 and 1 TB / s of memory bandwidth. an effective memory clock at 2 GHz, up from the effective clock at 1890 MHz on Vega 64. The fill rate of the Radeon VII should be 230 GPixels / second, more than double that of Vega 64.
A logical question, based on the overall improvements of Radeon VII, is why AMD does not plan better performance for the card. On paper, the Radeon VII has enormous resources. Its maximum pixel throughput blows RTX 2080 Ti out of the water, as is its memory bandwidth. But the only point of reference where these changes really shone is Luxmark, where the Radeon VII is 1.64 times faster than the Vega 64. AMD also provided additional baseline data. As always, the reference data provided by the manufacturer must be taken with a specific grain of salt. In the graph below created by Overclock3D), the top bar (gray) represents the percentage improvement from Vega 64 to Radeon VII, the second bar (red) indicates the performance of Radeon VII, while the third (white) represents Vega 64..
Based on these figures, the average performance improvement from Vega 64 to Radeon VII is 28%. If you remove the two outliers (+ 68% in Fallout 76 and + 7.46% in Hitman), the average improvement is 27.25%. In any case, we find ourselves in almost the same place. The gain is slightly higher than AMD's 1.25-fold flat improvement at the initial unveiling.
Based on these data, we can say a few words about the Radeon VII that we did not know before. The configuration of the graphics processor seems honestly a bit unbalanced, in that it combines a considerable increase in memory bandwidth and bit rate in pixels without changing the total number of cores. Clearly, AMD has added these resources for a reason. Since 7nm Vega is intended for both HPC / AI markets, it makes sense that they were added to meet the AI / ML space requirements.
Questions have been raised about the overall energy efficiency of GPUs and improvements made by AMD from 7 nm, but it may be difficult to separate them from changes in the design of the device. GPU. Increasing the clock of the RAM and doubling the amount of built-in memory will increase the power consumption. HBM was much more energy efficient than GDDR5, we did not see any data for HBM2 over GDDR6. All things being equal, a GPU with 128 ROPs will pull more power than the same core with only 64 ROPs. The two 8-pin power connectors and the three-fan chiller design make THG think that the power consumption is probably around 295W. This seems quite possible.
The Radeon VII will be available February 7 to be ordered by AMD for $ 699. AMD is targeting the RTX 2080 with this launch, which is also a $ 699 GPU. At launch, the Radeon VII will be delivered with a lot of games: Resident Evil 2, Devil can cry 5, and Tom Clancy's The Division 2.