Earlier this month, AMD added a bunch of new FirePro products to its lineup of workstation graphics cards. All of them employ 28-nm GPUs based on the Graphics Core Next architecture. Now, the very same chips have made their way into FirePro cards destined for servers. Introducing the FirePro S7000 (left) and S9000 (right):
Why might one need a substantial amount of graphics horsepower in a server? Virtualization, for one. The new FirePro cards support RemoteFX GPU virtualization, which allows multiple clients to share the resources of a single graphics chip. VMware or Citrix software can also be used to virtualize the FirePros, but they're limited to pass-through modes that allow only one client per GPU. More clients can be supported by adding FirePro cards, of course.
When it briefed us on the new FirePro cards, AMD noted that GPU virtualization is currently handled by the hypervisor software. AMD's current GPUs don't support hardware-based resource sharing. However, it sounds like that might change in a future generation.
Crunching compute workloads is also part of the job description for the new FirePro cards. For highly parallelized workloads, GPUs can deliver substantially better performance than the fastest CPUs. AMD mentions workstation graphics applications, too, but I suspect FirePro W-series cards are more appropriate in that arena. The workstation-oriented models have a lot more output options than the server cards, which are equipped with only a single DisplayPort out each.
|GPU||Core clock||ALUs||Memory bus||Memory||TDP||Price|
|FirePro S7000||Pitcairn||700MHz||1280||256-bit||4GB GDDR5||150W||$1,249|
|FirePro S9000||Tahiti||925MHz||1792||384-bit||6GB GDDR5||225W||$2,499|
At the high end, the FirePro S9000 employs the same Tahiti GPU that underpins the Radeon HD 7900 series. The chip runs at 925MHz, and it's backed by 6GB of GDDR5 RAM. Despite the card's 225W thermal envelope, the dual-slot cooler has no fans. At least 20 CFM of chassis airflow needs to be directed toward the passive heatsink to prevent the GPU from overheating.
The FirePro S7000 is a single-slot solution with a less exotic Pitcairn GPU pulled from the Radeon HD 7800 series. Pitcairn has fewer ALUs than Tahiti, and in the S7000, it's clocked at only 700MHz. As evidenced by its tighter 150W TDP, the S7000's cooling requirements are more relaxed. The heatsink is smaller, and it requires just 10 CFM of airflow.
On the memory front, the S7000 gets by with 4GB of GDDR5. The card doesn't support ECC, a feature that's limited to the S9000. AMD says enabling error correction cuts memory performance by about 6%.
Like their desktop counterparts, both FirePros employ ZeroCore power to reduce power consumption at idle.
|Peak single precision||Peak double precision||Memory bandwidth|
|FirePro S7000||2.4 TFLOPS||152 GFLOPS||154GB/s|
|FirePro S9000||3.2 TFLOPS||806 GFLOPS||176GB/s|
Obviously, the S9000 is the faster of the two. But its performance specifications don't match the 2X difference in price. The S9000's peak single-precision throughput is only 33% higher than the S7000, and its memory bandwidth is just 14% higher. That said, the S9000 has more than five times the double-precision throughput of its little brother. You can thank the Tahiti GPU, which has much better double-precision performance than Pitcairn.
Both FirePro cards are expected to start shipping at the end of September. Servers typically have longer upgrade cycles than the average enthusiast's desktop, so AMD has pledged to support these FirePro cards until at least 2015.
|Rumor: Intel Skylake-X and X299 will headline Computex 2017||33|
|Rumor: Nvidia to answer Radeon RX 550 with GeForce GT 1030||15|
|Samsung Galaxy Book tablets blend Windows 10 and Intel CPUs||13|
|Deals of the week: a mighty PSU, mid-range CPUs, and more||27|
|AMD board partners begin tricking out RX 560s and RX 550s||16|
|Dell shows off a pro-grade 4K HDR display and AIO machines||15|
|Rumor: Google to bake ad-blocking into Chrome browser||47|
|EpicGear's Defiant modular gaming keyboard reviewed||12|
|GeForce cards with faster RAM are inbound from multiple locations||19|
|Those power consumption numbers are very fermi-liar||+51|