facebook rss twitter

Review: NVIDIA (ZOTAC) GeForce 9800 GX2 - the champ is back!

by Tarinder Sandhu on 20 March 2008, 08:47

Tags: NVIDIA (NASDAQ:NVDA), ZOTAC

Quick Link: HEXUS.net/qamag

Add to My Vault: x

GeForce 9800 GX2 - sounds really new?

Spec Given the nomenclature, GeForce 9800 GX2, it would be reasonable to assume that the newest GPU would be a technological evolution from the slew of 8-series cards currently plying the mid-to-high-end market.

Rather, the 9800 GX2 amalgamates two GeForce 8800 GTS 512 SKU on to one board, pre-SLI's them, and, really, that's about all.

We'll show you exactly what we mean by referencing the specification-comparison table.

Graphics cards NVIDIA GeForce 9800 GX2 1024 NVIDIA GeForce 8800 GT 512 NVIDIA GeForce 8800 GTS 512 NVIDIA GeForce 8800 Ultra 768 ATI Radeon HD 3870 X2 1024 ATI Radeon HD 3870 512
PCIe PCIe 2.0 PCIe 1.x PCIe 2.0
GPU clock 600MHz 600MHz 650MHz 612MHz 825MHz 775MHz
Shader clock 1,500MHz 1,500MHz 1,625MHz 1,500MHz 825MHz 775MHz
Memory clock (effective) 2,000MHz 1,800MHz 1,940MHz 2,160MHz 1,802MHz 2,250MHz
Memory interface, size, and implementation 2x 256-bit, 1,024MiB, GDDR3 256-bit, 512MiB, GDDR3 384-bit, 768MiB, GDDR3 2x 256-bit, 1,024MiB, GDDR3 256-bit, 512MiB, GDDR4
Memory bandwidth 128GB/sec (card) 57.60GB/sec 62.1GB/sec 103.68GB/sec 115.328GB/sec (card) 72.8GB/sec
Manufacturing process TSMC, 65nm TSMC, 90nm TSMC, 55nm
Transistor count 1,508M 754M 681M 1,300M 666M
Die size 2x 296mm² 296mm² 484mm² 2x 192mm² 192mm²
DirectX Shader Model 4.0 4.1
Vertex, fragment, geometry shading (shared) 256 FP32 scalar ALUs, MADD dual-issue + MUL (unified) 112 FP32 scalar ALUs, MADD dual-issue + MUL (unified) 128 FP32 scalar ALUs, MADD dual-issue + MUL (unified) 640 FP32 scalar ALUs, MADD dual-issue (unified) 320 FP32 scalar ALUs, MADD dual-issue (unified)
Peak GFLOP/s 1,152 504 624 576 1,056 496
Data sampling and filtering 128ppc address and 128ppc bilinear INT8 filtering, max 16xAF 56ppc address and 56ppc bilinear INT8 filtering, max 16xAF 64ppc address and 64ppc bilinear INT8 filtering, max 16xAF 32ppc address and 64ppc bilinear INT8 filtering, max 16xAF 16ppc address and 32ppc bilinear INT8 filtering, max 16xAF
Peak GTexel/s (bilinear) 76.8 33.6 41.6 19.584 26.4 12.4
ROPs 32 16 16 24 16
Outputs 2 x dual-link DVI w/HDCP, HDMI 2 x dual-link DVI w/HDCP, mini-DIN 2 x dual-link DVI w/HDCP (discrete ASIC), mini-DIN 2 x dual-link DVI (HDMI) w/HDCP, mini-DIN (VIVO)
Hardware-assisted video-decoding engine NVIDIA's PureVideo HD - full H.264 decode and partial VC-1 decode AMD UVD - full H.264 and VC-1 decode
Reference cooler dual-slot single-slot dual-slot dual-slot
Retail price (default-clocked model) £399 £139 £189 £399 £269 £129


A mishmash of two G92 SKUs - 8800 GT 512 and 8800 GTS 512

Built on a 65nm fabrication process, the DX10-compliant, PCIe 2.0-based 9800 GX2 is literally two PCBs bundled together. Each PCB carries an identical G92-class GPU which incorporates 128 stream processors, made up of eight blocks of 16 units. That, then, gives it the same per-PCB SP array as the GeForce 8800 GTS 512, 8800 GTX, and Ultra SKUs.

The PCBs, and by inference GPUs, are hooked-up via a PCIe 2.0 bridge that runs at PCIe 2.0 speeds - bettering the PCIe 1.1 for the Radeon HD 3870 X2. Each PCB also carries 512MiB of GDDR3 memory, allotted to each GPU, giving the overall card a 1GiB frame-buffer.

Clocked in at 600MHz per core and 1,500MHz for the shaders, the base clocks are similar to the 112 SP-carrying 8800 GT 512, and a little slower than the 650/1,625MHz clocks on 8800 GTS 512. Having two PCBs, which face one another, and the heat considerations that are implied as a result, we can see why NVIDIA is keeping core- and shader-clocks at GT levels. Note, though, that the GX2's speeds are almost identical to the GeForce 8800 Ultra's.

The memory interface and ROP count also matches the GT(S) 512's, being 256-bit (512MiB in size) and 16, respectively, per GPU. These, we note, are lower than the 384-bit and 24 present on the long-in-the-tooth Ultra.

Effective memory speed is boosted to 2,000MHz, above the GTS 512's 1,940MHz, and total card bandwidth aggregates to a juicy 128GB/s.

Carrying on the G92 theme in earnest, each PCB, as per the GTS 512, is able to texture-address 64 pixels per clock, giving the GeForce 9800 GX2, in twin-GPU form, massive bilinear-filtering capacity.

The physical manifestation of what amounts to two GeForce 8800 GTS 512s - albeit with lower core/shader clocks and a higher memory clock - is a dual-slot-taking card that consumes around 200W under peak load and should benchmark somewhere in the vicinity of two discrete G92 512s in SLI.

There's nothing special at play here, no new, never-seen-before mojo that NVIDIA is bringing to the table. As should be abundantly clear now, but worth repeating in case you've missed it, GeForce 9800 GX2 represents two G92s in pre-configured SLI - give or take a few percentage points.

Priced at around £399 on release, it will appeal to enthusiasts that want, arguably, the fastest (single) graphics board that money can buy. The real bonus, for the money-no-object crowd, will lie in the ability to couple a pair of 9800 GX2s, in SLI, for the rebirth of Quad-SLI, but we're not allowed to talk much about that right now (right, NVIDIA?).

Four-GPU scaling

One can surmise that a single-board GeForce 9800 GX2's performance scaling will be good. After all, it's not that difficult to scale with two GPUs - both NVIDIA and AMD (ATI) have compelling multi-GPU technologies that leverage two GPUs in an efficient manner.

The real pain the posterior will reside with scaling with to four GPUs, based on two 9800 GX2s boards in SLI. Queuing up the requisite number of frames is hugely problematic in Windows XP and highly dependent on the games' engines in Vista. NVIDIA will introduce four-way alternate frame rendering (AFR) with the 9800 GX2 SLI driver.

NVIDIA will support Quad-SLI on the nForce 680i, 780i, 780a, and 790i chipsets, too.

Other bits and pieces

The GeForce 9800 GX2 carries NVIDIA's PureVideo HD that has been software-upgraded with the release of the ForceWare 174.xx drivers.

NVIDIA has added dynamic contrast and colour enhancements for better-looking images, it says. Further, the driver supports dual-stream decode, which hardware-accelerates the decoding of two video streams running concurrently. There are only few instances where we can see this being useful - picture-in-picture commentary, for example.

In terms of high-definition codec decode, PureVideo HD provides complete H.264 acceleration and partial VC-1 and MPEG-2 - missing out on the first step - Huffman Bitstream Processing.

Borrowing something from the mid-range GeForce 9600 GT, the 9800 GX2 supports an HDMI output on the backplate, but this has to be fed via a S/PDIF cable from your motherboard - a solution that's not nearly as elegant as ATI's DMA-based approach.

The 9800 GX2 also supports NVIDIA's HybridPower, where, on a compatible chipset that features an IGP, the discrete, wattage-hungry card can be turned off when not in heavy 3D use, saving power and routing video through the IGP instead. NVIDIA unveils HybridPower with the nForce 780a chipsets, also released today.

Specification summary

Massive processing power is harnessed by, really, coupling a couple of already-fast GeForce 8800 GTS 512 GPUs in SLI.

NVIDIA may tell you that it's not all about performance, but that's exactly what this card is all about.