NVIDIA GB10

# NVIDIA GB10

The NVIDIA GB10 occupies a unique position in the hardware landscape: it is a server-class Blackwell 2.0 part with a 50th percentile ranking among all GPUs, meaning it sits exactly at the median of the performance distribution. Its benchmark data is sparse — an average score of 0 with no listed rivals — so the analysis here relies on architectural specifications and memory characteristics rather than comparative frame rates. The GB10 is not a typical consumer graphics card; it is an integrated graphics processor (IGP) with a 140 W TDP, designed for a specific server workload profile where raw compute throughput and massive memory capacity outweigh traditional gaming metrics.

Who Should Consider It

The GB10 is for compute-first environments where the 128 GB LPDDR5X memory pool is the primary asset. With 6144 shading units and 29.71 TFLOPS of FP32 performance, this part can handle substantial parallel workloads, but its 48 ROPs and 116.1 GPixel/s pixel rate suggest it is not optimized for high-refresh 4K gaming. For resolution-specific guidance: at 1080p, the 29.71 TFLOPS compute should drive demanding scenes adequately, but the low ROP count relative to shading units means pixel fill could become a bottleneck in geometry-heavy titles. At 1440p, the 273.2 GB/s memory bandwidth is sufficient for most textures, yet the 48 ROPs will likely limit frame rates in games that rely on heavy post-processing. At 4K, the 256-bit bus width and 273.2 GB/s bandwidth are modest for the pixel workload, and the 116.1 GPixel/s fill rate is a hard ceiling — expect playable but not competitive performance. The 140 W TDP and IGP slot width mean this is a low-power embedded solution, likely for servers where space and thermal envelopes are constrained. If your workload involves large datasets that fit within 128 GB, this is the target audience. If you need high-refresh gaming or ray-traced visuals, look elsewhere.

Memory Subsystem

The GB10 ships with 128 GB of LPDDR5X memory on a 256-bit bus, yielding a total bandwidth of 273.2 GB/s. The memory clock runs at 1067 MHz, which translates to 8.5 Gbps effective. This configuration is unusual: 128 GB is an enormous capacity for a GPU, typically reserved for compute or AI inference tasks where model weights and activations exceed the 24 GB or 48 GB found on many server accelerators. However, the 273.2 GB/s bandwidth is not exceptionally high — many consumer parts with 384-bit buses exceed this figure. The trade-off is clear: capacity over speed. For high-resolution rendering, 273.2 GB/s is adequate for 1440p textures but will struggle with 4K texture streaming in open-world titles that demand rapid data movement. The 256-bit bus width is half of what you might find on a flagship gaming GPU, but the LPDDR5X type offers lower power consumption, aligning with the 140 W TDP. In practice, this memory subsystem favors batch processing and large buffer allocations over latency-sensitive gaming. The 128 GB pool allows entire datasets to reside on-chip, eliminating PCIe transfers — a decisive advantage for server workloads that repeatedly access the same data.

Ray Tracing and Feature Set

The GB10 includes 48 ray tracing cores and 384 tensor cores, both built on the Blackwell 2.0 architecture. The RT cores are present but the pixel rate of 116.1 GPixel/s and 48 ROPs suggest ray-traced workloads will be compute-bound rather than fill-bound. With 29.71 TFLOPS of FP32 and an equal 29.71 TFLOPS of FP16 (1:1 ratio), the tensor cores can accelerate AI-based denoising and DLSS-style features, though no specific API support (DirectX, OpenGL, Vulkan) is listed in the data. The absence of API details means software compatibility is unverified; however, the Blackwell 2.0 generation typically supports modern graphics APIs, and the PCIe 5.0 x16 interface ensures high-bandwidth communication with the host. The single HDMI display output limits multi-monitor setups, reinforcing that this is not a gaming-oriented product. The 384 tensor cores are the standout feature: they provide 29.71 TFLOPS of FP16 compute, which is directly applicable to neural network inference and training. For ray tracing, the 48 RT cores will handle basic intersection tests, but the 273.2 GB/s bandwidth and 48 ROPs will cap the effectiveness of complex BVH traversal and shading. In short, the feature set is server-grade compute with ray tracing as a secondary capability.

FAQ

Q: Is the GB10 suitable for 4K gaming?

A: No, likely not. The 116.1 GPixel/s pixel rate and 48 ROPs are limiting factors, and the 273.2 GB/s bandwidth is modest for 4K textures. The 128 GB memory capacity is excessive for gaming but irrelevant to performance.

Q: How much power does the GB10 require?

A: The TDP is 140 W, and the suggested PSU is 300 W. It uses no external power connectors, drawing power entirely from the PCIe 5.0 x16 slot.

Q: What is the memory bandwidth and capacity?

A: The GB10 has 128 GB of LPDDR5X memory on a 256-bit bus, with a bandwidth of 273.2 GB/s and an effective memory clock of 8.5 Gbps.

Q: Does the GB10 support ray tracing?

A: Yes, it has 48 ray tracing cores. However, performance is constrained by the low ROP count and pixel fill rate, so ray-traced workloads are likely compute-limited.

Q: What is the launch MSRP?

A: The launch MSRP is 3,999 USD.

Q: What is the form factor and interface?

A: The GB10 is an IGP (integrated graphics processor) with a slot width of IGP, dimensions of 150 mm by 51 mm by 150 mm, and a PCIe 5.0 x16 bus interface.

How It Compares

The GB10 has no nearest rivals listed in the benchmark data, and its average benchmark score is 0, so direct performance comparisons are unavailable. Its percentile rank of 50 places it at the median of all GPUs, but this is based on incomplete data. Given the specifications, it sits between consumer gaming cards and dedicated compute accelerators. Against a typical mid-range gaming GPU with 8-16 GB VRAM and 256-bit bus, the GB10 offers 8-16 times the memory capacity but comparable bandwidth. Against high-end server GPUs with HBM memory and 1 TB/s+ bandwidth, the GB10 falls short on speed but compensates with LPDDR5X power efficiency. The 48 ROPs are notably low — a modern gaming GPU often has 96-192 ROPs — which will hurt rasterization performance. The 384 tensor cores and 29.71 TFLOPS FP16 are competitive with mid-tier AI accelerators, but the lack of benchmark scores makes quantitative comparison impossible. The 140 W TDP is a fraction of typical server GPUs, which often exceed 300 W, making the GB10 a low-power alternative for inference workloads with strict thermal budgets.

Power and Cooling

The GB10 has a TDP of 140 W, a remarkably low figure for a GPU with 6144 shading units and 128 GB of memory. The suggested power supply is 300 W, which is modest and indicates that the IGP draws power solely from the PCIe 5.0 x16 slot — there are no power connectors on the card. This simplifies installation in server chassis where auxiliary power cables are unavailable or undesirable. The IGP slot width means it occupies no additional PCIe slots beyond the single x16 slot, and the dimensions of 150 mm by 51 mm by 150 mm (5.9 inches by 2 inches by 5.9 inches) are compact, fitting in most server enclosures. Cooling requirements are minimal given the 140 W TDP; a passive heatsink or low-profile active cooler should suffice in a well-ventilated server case. The 5 nm process node from TSMC contributes to the low power draw, as does the LPDDR5X memory, which is inherently more power-efficient than GDDR6 or HBM. For a server with multiple GPUs, the 140 W TDP allows for dense packing without overwhelming cooling infrastructure.

Benchmark Performance

The GB10's benchmark performance is undefined: the average benchmark score is 0, and the nearestRivals array is empty. The 50th percentile ranking is the only comparative data point, but it is based on this incomplete score, so it should be interpreted cautiously. Without rival scores or deltaPct values, percentage-based comparisons are impossible. What the data does show are raw compute metrics: 29.71 TFLOPS FP32 and 29.71 TFLOPS FP16 (1:1 ratio). The FP16 performance is notable because it matches FP32, which is rare — many GPUs halve FP16 throughput. This suggests the GB10 is designed for mixed-precision workloads where FP16 is sufficient, such as AI inference. The texture rate of 928.5 GTexel/s is high, driven by 384 TMUs, but the 116.1 GPixel/s pixel rate is disproportionately low, indicating a compute-heavy, fill-light design. In the absence of benchmark scores, these architectural ratios tell the story: the GB10 will excel at compute and tensor operations but lag in pixel-bound tasks. The 273.2 GB/s bandwidth, while not exceptional, is sufficient to feed the 29.71 TFLOPS compute at a ratio of roughly 9.2 bytes per FLOP, which is adequate for many compute kernels. Overall, the GB10's performance profile is that of a specialized compute accelerator, not a general-purpose gaming GPU, and the lack of benchmark data means any direct performance verdict must remain provisional.