Updated (1/11/2019, 6 PM EST): An AMD spokesperson has informed us that the widely reported GPU configuration for the Radeon VII is incorrect. The Radeon VII does not have 128 ROPs. Like Vega 64, Radeon VII is a 64 ROP card. The reports of a 128 ROP design were made by multiple well-regarded sites with reporters on-site at CES. It’s not clear why this information was reported incorrectly or how the error was introduced.
ExtremeTech regrets the error. All discussion and speculation of a 128 ROP design in the story below should be considered incorrect. Our story on the correct 64 ROP configuration can be read here.
Original Story Below:
AMD’s Radeon VII announcement earlier this week came as something of a surprise. The company has previously discussed its Radeon Instinct family of hardware and disclosed stats on the MI50 and MI60 at its Next Horizons event but gave no hint that it intended to launch a new high-end Vega GPU. Now the company has shared a bit more information on the card, including its own performance data.
Radeon VII is based on Vega 20, the 7nm die shrink built at TSMC. Die size is 331mm sq at 13.2B transistors, compared with 487mm sq and 12.5B transistors for Vega 64. The overall density improvement appears to be 1.56x and AMD claims its performance per die area has improved by 1.8x according to THG. While not unimportant — performance per die area has implications for die size and therefore manufacturing cost — it’s not terribly relevant to the question of how well Radeon VII will compete against the RTX 2080, its intended rival.
The GPU has 60 CUs, which translates to 3,840 GPU cores. The base clock is 1.45GHz with a boost frequency of 1.8GHz, offering 13.8TFLOPs of single-precision performance. The boost clock is a 1.16x improvement over the top-end frequency of the Vega 64, as is the base clock.
GCN has never been a high-clock architecture, despite AMD’s various revisions and improvements. The 7nm shift doesn’t increase clocks by as much as the 28nm – 14nm transition did, though the gap is small: Vega 64 clocked 1.19x faster than Fury X, while Vega VII clocks 1.16x higher than Vega 64.
The core configuration is 3840:240:128. In terms of active cores, that’s halfway between the Vega 56 and the Vega 64, but the number of render outputs has doubled, up to 128. The amount of memory onboard has doubled to 16GB of HBM2 and the 1TB/s of memory bandwidth requires an effective 2GHz memory clock, up from the effective 1890MHz clock on Vega 64. The Radeon VII’s fill rate should be a massive 230GPixels/second, more than double Vega 64’s fill rate.
One logical question, based on Radeon VII’s overall improvements, is why AMD isn’t forecasting better performance for the card. On paper, the Radeon VII has enormous resources. Its peak pixel throughput blows the RTX 2080 Ti out of the water, as does its memory bandwidth. But the only benchmark where these changes really shine is Luxmark, where the Radeon VII is 1.64x faster than the Vega 64.AMD has also provided additional benchmark data. As always, manufacturer-provided benchmark data should be taken with a decided grain of salt. In the chart below (graph created by Overclock3D), the top bar (gray) represents the improvement in percentage terms from Vega 64 to Radeon VII. The second bar (red) is Radeon VII performance, while the third (white) is Vega 64.
The average performance improvement from Vega 64 to Radeon VII based on these figures is 28 percent. If you remove the two outliers (+68 percent in Fallout 76 and +7.46 percent in Hitman), the average improvement is 27.25 percent. Either way, we end up in about the same place. The gain is very slightly larger than the flat 1.25x improvement AMD projected at the initial unveil.
Based on this data, we can say a few things about the Radeon VII that we didn’t know previously. The GPU configuration honestly seems a little lopsided, in that it combines a huge increase in memory bandwidth and pixel throughput with no change in the total number of cores. Clearly, AMD added those resources for a reason. Given that 7nm Vega is intended for both HPC / AI markets, it makes sense to think they were added to address the AI/ML space.
There have been questions about the GPUs overall power efficiency and what improvements AMD saw from 7nm, but it may be tricky to separate those from the changes to the GPU design. Increasing the RAM clock and doubling the amount of onboard memory will both increase power consumption. HBM was much more power-efficient than GDDR5 we haven’t seen data for HBM2 versus GDDR6. All else equal, a GPU with 128 ROPs is going to draw more power than the same core with just 64 ROPs. The two 8-pin power connectors and triple-fan cooler design make THG think the power draw is likely in the 295W range. That seems quite possible.
The Radeon VII will be available on February 7 for order from AMD at a cost of $699. AMD is targeting the RTX 2080 with this launch, which is also a $699 GPU. At launch, the Radeon VII will ship with a bundle of games: Resident Evil 2, Devil May Cry 5, and Tom Clancy’s The Division 2.
- AMD Announces Radeon VII: 7nm Vega Coming to the Consumer Market
- Nvidia CEO Sneers at AMD Radeon VII, Calls It ‘Lousy’
- AMD Announces Radeon Instinct MI60: Vega Arrives on 7nm