NVIDIA GB10 vs NVIDIA Tesla V100 PCIe 32 GB Comparison

NVIDIA
GEFORCE

NVIDIA GB10

CORE STATE GB20B
VRAM 128 GB
CLOCK SPEED 2418 MHz
TDP 140 W
BUS WIDTH 256 bit
ARCHITECTURE Blackwell 2.0
nm
PROCESS 5 nm
LAUNCH DATE 2025
VS
NVIDIA
GEFORCE

Tesla V100 PCIe 32 GB

CORE STATE GV100
VRAM 32 GB
CLOCK SPEED 1380 MHz
TDP 250 W
BUS WIDTH 4096 bit
ARCHITECTURE Volta
nm
PROCESS 12 nm
LAUNCH DATE 2018

PERFORMANCE BENCHMARKS

geekbench_opencl
120,137
168,763
geekbench_vulkan
114,648
131,847

Analysis: NVIDIA GB10 vs NVIDIA Tesla V100 PCIe 32 GB

NVIDIA Tesla V100 PCIe 32 GB and NVIDIA GB10 sit at opposite ends of the GPU spectrum, one a data-center veteran from the Volta era and the other a current-generation Blackwell part aimed at compact server and edge workloads. In direct benchmark comparisons, the V100 takes both available tests, but the GB10 counters with a dramatically newer architecture, higher clocks, and a memory capacity four times larger. The data shows a split personality: raw compute throughput favors the old card, while efficiency, feature set, and memory footprint favor the new one.

Head-to-Head Benchmarks

The Geekbench OpenCL test delivers a decisive win for the Tesla V100 PCIe 32 GB. It scores 168,763 against the GB10's 120,137, a 40.5% advantage. That is not a marginal gap; it is a full-class separation in raw compute performance. The V100's FP32 output of 14.13 TFLOPS, combined with its 897.0 GB/s memory bandwidth, clearly outperforms the GB10's 29.71 TFLOPS but far narrower 273.2 GB/s pipe. In OpenCL workloads that stress memory throughput, the V100's 4096-bit HBM2 interface is simply in another league.

The Vulkan test tells a similar but closer story. The V100 scores 131,847 versus the GB10's 114,648, a 15% margin. This smaller delta suggests that Vulkan titles or compute workloads lean more on shader and geometry throughput, where the GB10's higher clock speeds (1665 MHz base, 2418 MHz boost) partially compensate for its narrower memory subsystem. Still, the V100 wins both head-to-head tests, giving it a clean 2–0 record in this comparison. Its average benchmark score of 150,305 versus the GB10's 117,393 reinforces that overall, the older card remains the stronger compute performer.

Context from the rivals list sharpens the picture. The V100 sits at the 96th percentile of all GPUs, with its nearest rival being the AMD Radeon Pro W6800X at 160,671 (6.5% higher) and the NVIDIA A100 PCIe 40 GB at 162,504 (7.5% higher). The GB10, at the 95th percentile, trails its closest competitor, the NVIDIA RTX 4000 SFF Ada Generation, by just 0.3%, effectively a tie. Notably, the GB10's nearest rivals include the Tesla V100 SXM2 16 GB, which it beats by 2.6%. So while the V100 PCIe 32 GB dominates this specific head-to-head, the GB10 is competitive with a previous-generation V100 variant.

Architecture Differences

The architectural gap between these two is generational, literally. The V100 uses the GV100 chip on TSMC's 12 nm process, packing 21,100 million transistors into an 815 mm² die. The GB10 uses the GB20B chip on TSMC's 5 nm node, with a die size of 382 mm² — less than half the area, though transistor count is listed as unknown. The V100's transistor density comes to 25.9M per mm², while the GB10's density is not specified. The GB10 is built on Blackwell 2.0 architecture versus the V100's Volta, and that brings fundamental changes.

Clock speeds are a clear point of differentiation. The GB10 runs at 1665 MHz base and 2418 MHz boost, versus the V100's 1230 MHz base and 1380 MHz boost. That is a 36% higher base clock and a 75% higher boost clock. The GB10 also has more shading units (6144 vs 5120), more TMUs (384 vs 320), and it introduces 48 RT cores and 384 tensor cores. The V100 has 640 tensor cores but no RT cores at all. The V100 counters with far more ROPs (128 vs 48) and a much wider memory bus (4096-bit vs 256-bit).

Memory architecture could not differ more. The V100 uses 32 GB of HBM2 across a 4096-bit bus, delivering 897.0 GB/s bandwidth. The GB10 uses 128 GB of LPDDR5X on a 256-bit bus, yielding only 273.2 GB/s. That is a 3.3x bandwidth advantage for the V100, but a 4x capacity advantage for the GB10. The V100's memory runs at 876 MHz (1752 Mbps effective), while the GB10's runs at 1067 MHz (8.5 Gbps effective). The GB10's FP16 throughput matches its FP32 at 29.71 TFLOPS (1:1 ratio), whereas the V100's FP16 is exactly double its FP32 at 28.26 TFLOPS (2:1 ratio). The V100 also supports DirectX 12 (12_1), OpenGL 4.6, and Vulkan 1.4; the GB10 lists N/A for all three APIs.

Where Each One Wins

The V100 PCIe 32 GB wins wherever memory bandwidth or raw FP32 compute is the bottleneck. Its 897.0 GB/s bandwidth is nearly 3.3x the GB10's, which makes it the clear choice for large matrix operations, dense linear algebra, or any workload that streams data through memory faster than the compute units can consume it. Its 14.13 TFLOPS FP32 output, while lower than the GB10's 29.71 TFLOPS in peak terms, still produced a 40.5% higher OpenCL score — indicating that real-world throughput depends heavily on memory feeding the shaders. The V100 also matches or exceeds the GB10 in texture rate (441.6 GTexel/s vs 928.5 GTexel/s) only in the sense that its pixel rate is higher (176.6 GPixel/s vs 116.1 GPixel/s), which benefits rasterization-heavy tasks.

The GB10 wins on capacity, efficiency, and modern features. Its 128 GB of memory dwarfs the V100's 32 GB, making it the only option here for datasets that exceed 32 GB — think large language model weights, massive simulation grids, or multi-instance virtualization. Its 140 W TDP versus the V100's 250 W means it draws less power and needs no external power connectors (the V100 requires 2x 8-pin), and its IGP slot width means it fits in compact systems where the V100's dual-slot footprint cannot. The GB10's 48 RT cores enable ray tracing workloads that the V100 cannot touch, and its PCIe 5.0 x16 interface doubles the V100's PCIe 3.0 bandwidth for host transfers. The GB10 also has a display output (1x HDMI) while the V100 has none.

The Verdict

Choose the Tesla V100 PCIe 32 GB if your workloads are memory-bandwidth-bound and you need maximum compute per dollar of board space. Benchmark results are unambiguous: it beats the GB10 by 40.5% in OpenCL and 15% in Vulkan. Its 897.0 GB/s bandwidth and 4096-bit bus are unmatched by the GB10's 273.2 GB/s, and its 96th percentile ranking versus the GB10's 95th percentile shows it still outperforms most GPUs on the market. It is end-of-life, but that does not change what the data shows: for pure compute throughput, it wins.

Choose the GB10 if you need memory capacity, power efficiency, or modern architecture features. Its 128 GB of memory is four times larger, its 140 W TDP is 44% lower, and its Blackwell 2.0 architecture brings RT cores and 1:1 FP16/FP32 ratio that the Volta card lacks. It is active production, has a launch MSRP of 3,999 USD, and fits in an IGP form factor with no power connectors. The GB10's average benchmark score of 117,393 is lower, but its 95th percentile standing shows it remains competitive with current workstation cards like the RTX 4000 SFF Ada Generation (0.3% delta). If your application fits within its memory bandwidth limits, the GB10's modern feature set and efficiency make it the sensible long-term buy.

FAQ

Q: Which GPU has the higher average benchmark score?

A: The NVIDIA Tesla V100 PCIe 32 GB, with an average score of 150,305, versus 117,393 for the NVIDIA GB10.

Q: How much faster is the V100 in OpenCL?

A: The V100 scores 168,763 versus the GB10's 120,137, a 40.5% lead.

Q: Does the GB10 support ray tracing?

A: Yes, the GB10 has 48 RT cores. The V100 has none.

Q: What memory capacities and types do they use?

A: The V100 uses 32 GB of HBM2 on a 4096-bit bus. The GB10 uses 128 GB of LPDDR5X on a 256-bit bus.

Q: Which card has a higher boost clock?

A: The GB10, at 2418 MHz, versus the V100's 1380 MHz.

Q: What is the power draw difference?

A: The V100 has a 250 W TDP and requires 2x 8-pin power connectors. The GB10 has a 140 W TDP and uses no power connectors.

Specification Differences

| Specification | NVIDIA Tesla V100 PCIe 32 GB | NVIDIA GB10 |

|---|---|---|

| Architecture | Volta | Blackwell 2.0 |

| Process Node | 12 nm | 5 nm |

| Die Size | 815 mm² | 382 mm² |

| Transistors | 21,100 million | Unknown |

| Base Clock | 1230 MHz | 1665 MHz |

| Boost Clock | 1380 MHz | 2418 MHz |

| Memory Size | 32 GB | 128 GB |

| Memory Type | HBM2 | LPDDR5X |

| Memory Bus Width | 4096 bit | 256 bit |

| Memory Bandwidth | 897.0 GB/s | 273.2 GB/s |

| Shading Units | 5120 | 6144 |

| TMUs | 320 | 384 |

| ROPs | 128 | 48 |

| RT Cores | None | 48 |

| Tensor Cores | 640 | 384 |

| FP32 Performance | 14.13 TFLOPS | 29.71 TFLOPS |

| FP16 Performance | 28.26 TFLOPS (2:1) | 29.71 TFLOPS (1:1) |

| TDP | 250 W | 140 W |

| Slot Width | Dual-slot | IGP |

| Power Connectors | 2x 8-pin | None |

| Suggested PSU | 600 W | 300 W |

| Bus Interface | PCIe 3.0 x16 | PCIe 5.0 x16 |

| Display Outputs | No outputs | 1x HDMI |

| DirectX Support | 12 (12_1) | N/A |

| OpenGL Support | 4.6 | N/A |

| Vulkan Support | 1.4 | N/A |

| Production Status | End-of-life | Active |

| Release Date | 2018-03-26 | 2025-10-14 |

| Predecessor | Tesla Pascal | Server Hopper |

| Successor | Tesla Turing | Server Rubin |

| Launch MSRP | None listed | 3,999 USD |

DETAILED SPECIFICATIONS

SPECIFICATION
GB10
Tesla V100 PCIe 32 GB
Core Specs
Shading Units
6,144
5,120 -16.7%
Shaders
6,144
5,120 -16.7%
TMUs
384
320 -16.7%
ROPs
48
128 +166.7%
SM Count
48
80 +66.7%
Clocks
Base Clock
1665 MHz
1230 MHz
Boost Clock
2418 MHz
1380 MHz
Memory Clock
1067 MHz 8.5 Gbps effective
876 MHz 1752 Mbps effective
Memory
Memory Size
128 GB
32 GB
VRAM (MB)
131,072
32,768 -75.0%
Memory Type
LPDDR5X
HBM2
Memory Bus
256 bit
4096 bit
Bandwidth
273.2 GB/s
897.0 GB/s
Cache
L1 Cache
128 KB (per SM)
128 KB (per SM)
L2 Cache
50 MB
6 MB
Performance
Pixel Rate
116.1 GPixel/s
176.6 GPixel/s
Texture Rate
928.5 GTexel/s
441.6 GTexel/s
FP32 (TFLOPS)
29.71 TFLOPS
14.13 TFLOPS
FP64 (TFLOPS)
464.3 GFLOPS (1:64)
7.066 TFLOPS (1:2)
FP16 (TFLOPS)
29.71 TFLOPS (1:1)
28.26 TFLOPS (2:1)
AI/RT
RT Cores
48
—
Tensor Cores
384
640 +66.7%
Power
TDP
140 W
250 W
TDP (W)
140
250 +78.6%
Suggested PSU
300 W
600 W
Power Connectors
None
2x 8-pin
Architecture
Architecture
Blackwell 2.0
Volta
GPU Name
GB20B
GV100
Generation
Server Blackwell (Bxx)
Tesla Volta (Vxx)
Process Size
5 nm
12 nm
Transistors
unknown
21,100 million
Die Size
382 mm²
815 mm²
Foundry
TSMC
TSMC
Density
—
25.9M / mm²
API Support
DirectX
—
12 (12_1)
OpenGL
—
4.6
Vulkan
—
1.4
OpenCL
3.0
3.0
CUDA
12.1
7.0
Shader Model
—
6.8
Physical
Slot Width
IGP
Dual-slot
Length
150 mm 5.9 inches
—
Height
51 mm 2 inches
—
Outputs
1x HDMI
No outputs
Bus Interface
PCIe 5.0 x16
PCIe 3.0 x16
Other
Launch Price
3,999 USD
—
Production
Active
End-of-life
Predecessor
Server Hopper
Tesla Pascal
Successor
Server Rubin
Tesla Turing
View GB10 Details View Tesla V100 PCIe 32 GB Details