New Opteron Servers Struggle Badly Despite Additional Cores
In the wake of Bulldozer's weak desktop launch, there was some hope that server benchmarks would redeem the chip and give AMD a fighting chance against Intel in at least one of its three core markets. Recent reviews and a comprehensive roundup of formally published results, however, indicate that while BD fares a bit better in servers, it's far from vindicated.
AMD's new 6200 Opterons (Interlagos) are often a sidewise move from the older 6100 series, which is based on Magny-Cours. Interlagos is essentially trapped by the same trio of issues that cripple Bulldozer. The shared core design hurts SMP scaling; a 16-core Interlagos scales more like a 13-14 core Magny-Cours. The clock speed increases are offset by Bulldozer's lower single-thread performance. Finally, Interlagos draws more power than its predecessor, which in turn prevents further clock speed scaling.
As Ars Technica notes, Interlagos is faster than its predecessor in some tests -- but often by much less than one might expect. In JBB2005, Interlagos is just 27 percent faster than Magny-Cours, despite 33% additional cores and a four percent clock speed advantage.
This didn't exactly happen
Bulldozer / Interlagos' lowered IPC is at the heart of the CPU's problems. If AMD had even managed to maintain single-thread performance parity with Thuban, than the chip's higher clock speeds and greater number of cores would give it at least a marginal fighting chance, particularly in highly multithreaded environments. As Ars writes, "AMD compromised single-threaded performance in order to allow Bulldozer to run more threads concurrently, and that trade-off simply hasn't been worth it... it leaves us wondering what might have happened if AMD had simply extended its old architecture. Another four cores in a Magny-Cours processor would show close to the same 33 percent gain, and would do so without compromising single-threaded performance."
The same thought has occurred to us. Bulldozer/Interlagos was supposed to save die space compared to Magny-Cours, but AMD's decision to load the chip with 16MB of L2+L3 cache per eight-core configuration largely obviated the space savings. At this point, extending Thuban does look as though it would've been the smarter option.
The problem with that argument, however, is that it wouldn't have addressed the greater issue. While it's true that a 16-core Magny-Cours and subsequent 8-core 32nm desktop chips would've been more attractive than Bulldozer, they wouldn't have improved AMD's competitive position vs. Intel. Sunnyvale would still need a new CPU architecture that was capable of competing with Sandy Bridge, and it still wouldn't have one.
There's no quick fix for Interlagos, but there's still the possibility that future iterations of the chip will prove more capable. Right now, that -- and a handful of tepid server wins -- is the best AMD has to offer.
AMD's new 6200 Opterons (Interlagos) are often a sidewise move from the older 6100 series, which is based on Magny-Cours. Interlagos is essentially trapped by the same trio of issues that cripple Bulldozer. The shared core design hurts SMP scaling; a 16-core Interlagos scales more like a 13-14 core Magny-Cours. The clock speed increases are offset by Bulldozer's lower single-thread performance. Finally, Interlagos draws more power than its predecessor, which in turn prevents further clock speed scaling.
As Ars Technica notes, Interlagos is faster than its predecessor in some tests -- but often by much less than one might expect. In JBB2005, Interlagos is just 27 percent faster than Magny-Cours, despite 33% additional cores and a four percent clock speed advantage.
This didn't exactly happen
Bulldozer / Interlagos' lowered IPC is at the heart of the CPU's problems. If AMD had even managed to maintain single-thread performance parity with Thuban, than the chip's higher clock speeds and greater number of cores would give it at least a marginal fighting chance, particularly in highly multithreaded environments. As Ars writes, "AMD compromised single-threaded performance in order to allow Bulldozer to run more threads concurrently, and that trade-off simply hasn't been worth it... it leaves us wondering what might have happened if AMD had simply extended its old architecture. Another four cores in a Magny-Cours processor would show close to the same 33 percent gain, and would do so without compromising single-threaded performance."
The same thought has occurred to us. Bulldozer/Interlagos was supposed to save die space compared to Magny-Cours, but AMD's decision to load the chip with 16MB of L2+L3 cache per eight-core configuration largely obviated the space savings. At this point, extending Thuban does look as though it would've been the smarter option.
The problem with that argument, however, is that it wouldn't have addressed the greater issue. While it's true that a 16-core Magny-Cours and subsequent 8-core 32nm desktop chips would've been more attractive than Bulldozer, they wouldn't have improved AMD's competitive position vs. Intel. Sunnyvale would still need a new CPU architecture that was capable of competing with Sandy Bridge, and it still wouldn't have one.
There's no quick fix for Interlagos, but there's still the possibility that future iterations of the chip will prove more capable. Right now, that -- and a handful of tepid server wins -- is the best AMD has to offer.
