AMD's soon-to-be-released Radeon VII graphics card will come with FP64 (double-precision floating-point) disabled, Sasa Marinkovic, Director of Product Marketing at AMD, confirmed to Techgage in a report this week.
One of the first things you may notice about the AMD Radeon VII is its resemblance to the existing Radeon Instinct MI50. The former is a gaming graphics card, while the latter is marketed towards data centers and enterprises. However, the two are both based on the Graphics Core Next (GCN) 5.0 microarchitecture and employ the Vega 20 silicon. Not to mention that they are also product of TSMC's 7nm FinFET manufacturing process. But the similarities don't stop there.
|Row 0 - Cell 0||AMD Radeon VII||AMD Radeon Instinct MI50||AMD Radeon RX Vega 64|
|Architecture (GPU)||Graphics Core Next 5.0 (Vega 20)||Graphics Core Next 5.0 (Vega 20)||Graphics Core Next 5.0 (Vega 10)|
|Peak FP32 Compute||13.8 TFLOPS||13.4 TFLOPs||12.5 TFLOPS|
|Base Clock Rate||1450 MHz||1200 MHz||1200 MHz|
|GPU Boost Rate||1800 MHz||1746 MHz||1536 MHz|
|Memory Capacity||16GB HBM2||16GB HBM2||8GB HBM2|
|Memory Clock||2000 Mbps*||2000 Mbps||1890 Mbps|
|Memory Bandwidth||1024 GB/s||1024 GB/s||483.8 GB/s|
|Transistor Count||13.2 billion||13.2 billion||12.5 billion|
|Die Size||331 mm²||331 mm²||487 mm²|
AMD equipped both the Radeon VII and Radeon Instinct MI50 with 3,840 stream processors, 240 TMUs (texture mapping units), 64 ROPs (render output units) and 16GB of HBM2 memory across a 4096-bit memory interface. Being a gaming-focused product, the Radeon VII comes with a higher base clock and boost clock. Therefore, it seems the chipmaker had to take something away from the Radeon VII so it doesn't negatively impact Radeon Instinct MI50 sales. That's where FP64 comes in.
The Radeon Instinct MI50 can deliver up to 6.7 TFLOPs of FP64 peak performance. By gimping the Radeon VII's FP64 performance, AMD effectively creates a huge gap between the two models. Despite all this, the Radeon VII should be capable of pumping out 1,680 GFLOPs of FP64 performance and stay ahead of the Nvidia Titan RTX (509 GFLOPs) and GeForce RTX 2080 Ti (420 GFLOPs).
In a likely move to further segment the Radeon VII from the Radeon Instinct MI50, AMD has disabled PCIe 4.0 on the former. The reasoning is that AI workloads in the data center can saturate the PCIe 3.0 interface with relative ease, and limiting the available bandwidth would dissuade data center operators from repurposing cheaper consumer gear for those types of workloads. So, AMD has left the feature active only for the Radeon Instinct MI50.
For the average consumer, is it a dealbreaker that the Radeon VII's FP64 performance isn't on par with the Radeon Instinct MI50? Not really, since gamers can surely live without it. But there are professional users who are also gamers, and while they would have loved to have a graphics card that brings both premium FP64 and gaming performance, AMD isn't letting that happen with the Radeon VII.
I kind of understand AMD on this, if they didn't, the MI50 Instinct would have been outsold by the VII on price alone by lower quality silicon.
Still, I believe AMD could have left it there if they could. It would have been a superb selling point for professional.
Personally, FP64 was the primary driver of my interest in the card but now I couldn't care less about it. I knew many people that were excited also but now not one of them plan to get one. The extra ram is simply not enough to make it a compelling purchase.
Market it like we know our customers need to work but they also like to relax. Now you can do both with one card! The Radeon VII pro provides a 4k gaming experience and a similar compute value compared to a Titan V for a third of the price. Buy two and double your compute and it will still cost you less!
Maybe AMD can release drivers/software for $300 that you guys can purchase?
Miners don't even use the double precision feature.
If every 100,000th calculation was wrong for a miner they wouldn't care. (Although they should probably dial back their overclock slightly.)
They would just keep going with more calculations.
Double precision is literally doing every calculation twice to make sure the math was done right.
The chart on
illustrates this beautifully.
You're right that miners probably don't care about FP64 performance though. But that's because most popular algorithms for GPU mining these days are more memory-bound than compute-bound.