AMD Announces FirePro S-series: S9000 and S7000

Company: AMD
Author: James Prior
Editor: Charles Oliver
Date: August 26th, 2012

AMD FirePro S-series: S9000 and S7000

AMD is ending August like they began, launching new FirePro products at a high profile computing industry event filled with vendors and customers who want to use the new products. Akin to the FirePro W5000, W7000, W8000 and W9000 launch at SIGGRAPH on August 7th, the new FirePro S7000 and S9000 cards are being introduced at VMWorld 2012.

FirePro S7000
FirePro S9000

These cards are passively cooled, clearly showing their intent to go in managed platforms mounted in racks far from prying fingers and intemperate conditions. Their server based nature is denoted in their product nomenclature, with the Workstation FirePro W replaced with an S for Server; the V prefix is dropped, as is the P suffix. That's not to say these cards can't be used for Workstation graphics, as they can - albeit with a single display output. If AMD ever manage to persuade someone to build the unicorn powered DisplayPort 1.2 MST hubs, that single DisplayPort output could power up to three displays.

The FirePro Server graphics line is comprised of the Pitcairn based S7000 and the Tahiti based W9000. The FirePro S7000 is specified identically to the FirePro W7000 and replaces the V7800P, but the FirePro S9000 is the performance twin of W8000 - not W9000 - card, and will replace the V9800P. Both cards feature PowerTune and ZeroCore Power technologies like their FirePro and Radeon brethren, but not the GeometryBoost capabilities offered by the W9000 (and Radeon HD 7970 Ghz Edition or new HD 7950 Boost BIOS).

FirePro S-series vs. FirePro series-P

The TDP of the W7000 and W8000 transfer identically, 150W and 225W respectively. Impressively, AMD gives the OEM cooling requirement as just 10CFM for the S7000 and 20CFM for the S9000. Only the Tahiti based S9000 offers ECC capabilities, at a 6% reduction in memory performance and reduces the 6GB memory to 5.25GB in capacity, like NVIDIA's comparably priced Tesla T20-based cards which reduce speed by 12.5% and capacity to 5.25GB with ECC enabled.

FirePro S7000
FirePro S9000

The price points go the other way, with the W7000's $899 cost elevated to $1249 for the S7000, and the W8000's $1599 shooting up nearly a thousand dollars to $2499. The reasons for this are two-fold, number one is the competitive landscape NVIDIA's equivalent performance offerings are more expensive and the outgoing V7800P and V9800P sell at those price points; secondly, capitalizing on the use cases these cards will see. The use case for datacenters vs. workstation is quite different, with datacenters aiming to be used 24/7 supporting worldwide operations, as opposed to workstations that get used by one person for a third of the day and then idle or powered off the rest of the time - that increase in workload is going to increase support levels and that cost has to be absorbed somehow.

Ecosystem Targets

So, what kinds of datacenters use multi-teraflop graphics cards? There are three obvious use cases: Vizualization, GPU Compute and Virtualization. Vizualization is rendering, turning vectorized descriptions of objects and interactions into realistic images humans can comphrehend, sometimes animated. GPU Compute should need no introduction now, with many applications sporting native support for offloading massively parallel operations to vector processing engines. Virtualization vendors have been pushing the Virtual Desktop Initiative (VDI) for some time, and now we're starting to see hypervisors become aware of and use graphics processors; Windows Server 2008 RemoteFX offers GPU virtualization density of up to 16 users one 1 last generation V9800P GPU, and while we don't have numbers for the new S9000 replacing the V9800P you can surely expect that density to go up. Citrix XenServer and XenDesktop support multiple users per GPU as well, with XenApp HDX 3D offering DirectX acceleration on the GPU and Xen Desktop HDX 3D Pro accelerating DirectX and OpenGL, too. On the VMware side, VMDirectPath I/O allows a single GPU to be natively used by a VM, using the normal OS driver and the potential for VMware to leverage this over many users in future. Both past generation V7800P and V9800P will support this, as well as the new S7000 and S9000 products.


The big message we've been hearing from AMD since the debut of Fusion APUs (now evolved into the HSA Foundation's platform) is move the compute to the data. For many businesses geographically dispersed offices as required by their customer's market locations, drives data into many buckets, which in turn makes both business intelligence and data safety hard to grasp. Moving to the cloud can solve this problem, but as heterogeneous compute solutions are adopted so too is the need for it in the cloud. Bringing GPU's into the cloud brings their giga- and teraflops of compute with them, reducing the bottlenecks. This model works for both big analytics and individual contributor models. Consider two offices, one on the west coast of the US and one in Asia, or perhaps India. Time zone differences prohibit both teams working together directly on the same project, so traditionally both teams would share data and check updates in and out of a central repository. This can add significant delays to productivity, and slow down bring up of new projects. Leveraging VDI, where all the compute is performed in one locale (which may be located at the cheapest infrastructure location with the best high availability/disaster recovery profile, instead of an office), the data moves and the talent works from where it resides. If you can stream HD movies, why not a smooth desktop experience?

The obvious answer to that is latency, but seeing as we can play competitively across continents in twitch games selling for $60USD, it doesn't seem too much of stretch to think the Enterprises with the kind of resources to mass deploy AMD FirePro S-series cards can solve the problem of secure low latency prioritized network connections, too.

It's also interesting to see AMD talking about the possibilities of Virtualization with VMware. AMD's own internal hypervisor of choice is VMware ESX, and you can be sure they're working with VMware closely to get support for GPU's into the datacenter especially as they're building FirePros in to APUs now. As HSA pushes forward with making the GPU a peer level device, integral and inextricably entwined with the CPU, virtualization hypervisors and features will need to follow suit. Cloud gaming, high definition media encode and stream playback with real time enhancements and steadying, all need the GPU right there with the data.

Market Opportunities

AMD's stategy for success is similar to that used for the previous generation V7800P & V9800P: a single product fits the bill for three markets, at a specific price point. AMD also have performance options to suit different budgets. The FirePro S-series are a single card solution in markets where the competition requires two, or offers much less performance, or a bigger card; meaning lower density - until more Kepler based Tesla cards appear - but until then AMD has the bang/buck play. AMD's execution of the AMD FirePro series appears to be well handled with the FirePros hitting the market under the eyes of the world at major events with fully certified drivers for key ISV's, DirectX 11.1, OpenCL 1.2 and OpenGL 4.1 (4.2 beta) support. AMD appear to have a well-rounded, well-placed product for many uses, and expect to ship at the end of September in OEM systems worldwide.

AMD Goes Red