AMD Reveals Navi and Its RDNA Architecture
AMD’S LAST MAJOR REVAMP of its GPU architecture was in 2012, with the HD 7970 and related first-gen GCN graphics cards. Everything since then, up through the Vega 20 GPU at the heart of the Radeon VII, has used a variation of GCN. At E3, and with the official unveiling of the Radeon RX 5700 XT, AMD finally gave us the skinny on its plans for the next generation of GPUs, code-named Navi, and using a new RDNA architecture.
Perhaps the biggest change from GCN is that AMD has reworked the way instructions are grouped. In GCN, AMD used a Wave64 work-item that combined work for 64 GPU threads. Each Wave64 would end up being split into four chunks of work that would be sent to a SIMD16 structure, requiring four cycles to execute the work-item. If the Wave64 didn’t fully utilize 64 threads, no matter: It still took four cycles. A shared scalar unit would handle interleaving the workload on to the SIMD unit.
For RDNA, AMD GPUs now have a base Wave32 work-item, and it feeds into a SIMD32 vector unit. There are also two scalar units for scheduling the workloads, which increases throughput to one Wave32 per cycle. Not only do heavy workloads end up executing faster, but lighter workloads— where only 16 threads of work are needed, for example—now complete in one cycle instead of four. Basically, everything becomes more efficient.
AMD also adds a new L1 cache to the GPU caching hierarchy, joining the existing L0 and L2 caches, and RDNA features dual compute units with some shared resources to keep things plugging along—AMD calls this a Workgroup Processor. Wider data paths connect the pieces together, and the full graphics pipeline can now work directly with compressed color data.
The net benefit is that RDNA should deliver roughly 25 percent better performance per clock— or IPC (instructions per clock). And while I haven’t been able to run benchmarks yet, AMD claims the 9.75 Tflops RX 5700 XT should perform slightly better than the 12.67 Tflops RX Vega 64. It also manages this with 40 CUs compared to Vega’s 64 CUs. But it’s not just about raw performance. The RX 5700 XT will also have a TBP (typical board power) of 225W, compared to Vega 64’s 295W.
That’s an excellent improvement if true, and AMD says performance per watt is 50 percent better than the last-gen GCN architecture. It needs to be a major improvement, as Nvidia has been killing AMD on GPU efficiency for seven years.
What’s alarming is that RDNA appears to continue a lot of the GCN legacy. Navi is manufactured at 7nm, yet performance and efficiency look to be at the same level as Nvidia’s three-yearold 16nm Pascal architecture. The RX 5700 should equal the performance of Nvidia’s RTX 2070, but without any ray tracing or deep learning enhancements. (Ray tracing will come in next year’s Navi 20 and RDNA 2.) It looks like Pascal rather than Turing. And the 12nm RTX 2070 is a 175W TBP part, while the 5700 XT is 225W TBP.
RDNA and Navi are moving in the right direction, but Nvidia has a response—it will be launching the RTX 2060 Super and RTX 2070 Super (and later the 2080 Super). Performance should be up to 25 percent faster than existing cards, so the target AMD was aiming for has moved. I’ll look at new GPUs from both companies next month.