loading
loading
The NVIDIA RTX PRO 4500 Blackwell Server Edition is the card NVIDIA built for the servers most enterprises already own. It is a single-slot, passively cooled, 165W board with 32GB of GDDR7 and the full Blackwell feature set: fifth-generation Tensor Cores with FP4, fourth-generation RT cores, three encoders and three decoders, and Multi-Instance GPU support that can split it into two isolated 16GB partitions. It replaces the L4 as the entry point to NVIDIA’s RTX PRO server lineup, and it fits in a 2U PowerEdge, ProLiant, or ThinkSystem without a power or cooling concern. NVIDIA’s pitch for the card is right-sizing: most enterprises are not building AI factories; they run air-cooled, brownfield data centers with hard limits on rack power, and they need a first accelerated computing step that does not require a new server class. Or they’re working at the edge, where the RTX PRO 4500 is better aligned with the workloads and power constraints that typically determine what can be deployed. The RTX PRO 6000 Blackwell covers the top of that market with 96GB and up to 600W. The 4500 keeps the same architecture and cuts everything else roughly in half, so customers with smaller models, many smaller tasks, or a per-dollar-per-watt scale-out budget get Blackwell without paying for headroom they won’t use. The workloads for this class of GPU are well understood: small language model inference, vision AI and video analytics, vector search and embedding generation, transcoding, and virtual workstations under vGPU 20. It claims over 5x the L4 for small and mid-size model inference and up to 10x on Nemotron models, and its launch partners span NTT DATA for edge video, Pacific Biosciences for sequencing, and Nokia for AI-RAN base stations. AWS built its new G7 instances on it as a replacement for the G6. We already tested the card against the L4 inside an HPE ProLiant DL145 Gen11 in our RTX PRO 4500 edge review, where it delivered 2.7x the L4’s throughput. This review answers how close it gets to the 350W, dual-slot L40S that has been the default mainstream inference card for two years, and what the RTX PRO 6000 above it adds in extra memory and power. To find out, we put the 4500 in a Dell PowerEdge R770 against an L40S and L4, added the RTX PRO 6000 Blackwell data from our Dell Pro Max Tower T2 at both its 600W default and a 450W cap matching what the R770 allows a GPU slot, and ran vLLM online serving across four models at two context lengths. As a brief preview, at typical chat and agent context lengths, the 4500 keeps pace with the L40S on FP8 and BF16 models at less than half the power; at 8K input, the 32GB frame buffer shows its limits; and NVFP4, the precision NVIDIA leads its Blackwell pitch with, performs in line with the rest of the lineup on this vLLM build. Key Takeaways - L40S throughput at 165W in one slot: At 512-token input and output, the 4500 matched or edged the L40S on every FP8 and BF16 model we ran, 1,107 vs. 995 tokens/s on Ministral 3 14B FP8 and 2,006 vs. 1,988 on Ministral 3 3B, from a card rated for 47 percent of the L40S’s power and half its slots. - Roughly 3x the L4 it replaces: 2.8x to 3.6x the L4’s output throughput at short context, at 2.3x the L4’s board power. - 32GB has a ceiling: At 8K input the 4500 trailed the L40S by 17 to 33 percent and its curves flattened at concurrency 8 where the L40S kept climbing to 16. The flattening is KV cache running out, and 32GB sets that limit. - NVFP4 is a software story, on every card: On Qwen3.8-27B NVFP4 the 4500 produced 320 tokens/s at concurrency 32, the L40S 480 with no FP4 hardware at all, and the RTX PRO 6000 621. The NVFP4 path in this vLLM build does not scale with memory bandwidth on any of them, while the MAMF sweep shows the 4500’s FP4 hardware reaching 568.6 TFLOPS against 280.1 in FP8. - What the bigger card buys you: Capped to 450W, the ceiling a Dell R770 GPU slot allows, the RTX PRO 6000 ran about 2x the 4500 at short context and 2.5x to 3.2x at 8K. Its full 600W rating added 1 to 2 percent at short context and 4 to 9 percent at long context on top of that, so most of what the 6000 delivers over the 4500 is memory. NVIDIA RTX PRO 4500 Blackwell Server Edition Specifications Two cards carry the RTX PRO 4500 Blackwell name, and they’re similar but not the same product. The Workstation Edition, announced at GTC 2025, is a 200W, dual-slot, actively cooled card with four DisplayPort outputs and 896GB/s of memory bandwidth. The Server Edition we are reviewing here arrived at GTC 2026 as a 165W, single-slot, passive, headless board at 800GB/s, and it trades the display outputs and some clock headroom for the data center feature set: MIG, confidential computing, a third NVENC and NVDEC, and vGPU 20 support. Underneath, they have the same 10,496-core GPU and 32GB of GDDR7 with ECC. NVIDIA’s data put the die at 45.6 billion transistors on 378 mm² of TSMC 4nm custom silicon with 82 fourth-generation RT cores, which works out to 82 SMs and 328 Tensor Cores. Those are GB203 numbers, the same die family as the RTX 5080, lightly cut. The Tensor figures below are NVIDIA’s peaks with sparsity; dense is half, which NVIDIA’s own peakTOPS binary confirms at 805 TFLOPS FP8 and 403 TFLOPS FP16 measured on this card. Design and Build The 4500 Server Edition is a full-height, full-length board, 4.4 inches tall and 10.5 inches long, and it fills one slot. That is an important change from the L4, which was a half-height, half-length card. The L4 is still the comparison it benchmarks against, but the form factor moved to FHFL to make room for the memory, the 16-pin power input, and a heatsink that can shed 165W with chassis airflow alone. That form factor is the one real break from the L4, which fit any HHHL slot and drew its 72W from the slot, so it went into 1U servers and small edge boxes. The 4500 is full height, full length, and needs the 16-pin cable, and the 1U mainstream servers we work with, Dell’s current 1U PowerEdge line among them, top out at HHHL or FHHL cards without that connector. In practice, this is a 2U card, or one for the few whitebox 1U chassis built around FHFL slots, so a shop that bought L4s for 1U or edge boxes does not get a drop-in replacement. Stacked edge-on with the L4 above it and the L40S below, the 4500 needs one slot where the L40S needs two, which is why a 2U server that takes four double-wide cards can take eight of these, and how AWS’s G7 instances run up to eight of them per host. Power is delivered through a single PCIe CEM5 16-pin connector, the same connector the RTX PRO 6000 uses, with no auxiliary 8-pin option. Cooling is entirely passive, so the card depends on the server’s fans and will throttle in a chassis that cannot move air through it, which is normal for the class, and all other GPUs have the same requirement. The card has no display outputs, so graphics work, whether vGPU virtual workstations or rendering, goes out over the network, and the bracket is vented. The three ninth-generation NVENC and three sixth-generation NVDEC engines add 4:2:2 H.264 and HEVC encode and decode over the L4’s generation, and NVIDIA quotes 2x the H.264 decode throughput per NVDEC engine versus the L4. The bracket shot below makes the split visible: the L4 and the 4500 carry vents and nothing else, while the L40S and the RTX PRO 6000 Workstation Edition each carry four DisplayPort outputs. Where the RTX PRO 4500 Blackwell Server Edition Fits NVIDIA’s framing is that the card’s “central value is the combination of AI, compute, graphics, and video acceleration in a 165 W, passive, single-slot design,” built for servers “where rack density, power, and cooling are primary deployment constraints.” It sits in what NVIDIA calls the mainstream data center and edge segment of the RTX PRO server portfolio, below the RTX PRO 6000 Blackwell Server Edition and above nothing; it’s the entry point of the Blackwell server line. Against the L4, NVIDIA cites up to 10x on small model inference with its Nemotron Nano models (NVFP4 on the 4500 against FP8 on the L4), up to 4x on video summarization with Cosmos Reason 2, 1.9x on SPECviewperf 15 at 4K, and 1.8x streaming density. Against CPUs, it cites up to 50x on vector database work with cuVS CAGRA in Milvus and up to 100x on vision AI, both measured on eight-card 4500 servers against dual-socket EPYC 9654 systems. Dell’s server team told us it sees the 4500 in mainstream racks, since three 32GB 4500s cost near the same as one 96GB 6000, and at that point the bigger card usually wins. This is a card for a server that would otherwise have no GPU, or an L4, and is a suboptimal fit for a server that was designed to get one or more 6000s. Test Setup We installed the 4500, the L40S, and the L4 in our Dell PowerEdge R770, configured with two Intel Xeon 6787P processors (86 cores each) and 2TB of DDR5-6400 across 32 DIMMs, running NVIDIA driver 595.71.05. The test runs used the NGC vLLM 26.08-py3 container, which carries vLLM 0.27.1 on CUDA 13.4.1, so the software stack is identical across cards. Each card was tested individually at its default power limit, and the R770 supplies each card its full rating: Dell lists the L4 at 72W, the 4500 at 165W, and the L40S at 350W in that chassis. The RTX PRO 6000 Blackwell is the workstation card in our Dell Pro Max Tower T2, on the same container and driver. NVIDIA rates the workstation and server editions identically at 600W, and they profile the same in our testing, so it stands in for the server card. Incidentally, we do have a review of the RTX PRO 6000 Server cards in an XE7740 if you want to dig into how these cards can scale. We ran the 6000 twice: once at its 600W default, and once with a verified 450W cap applied through nvidia-smi, because 450W is the most the R770 will deliver to a 6000 in a GPU slot, and it is the figure a buyer choosing between a 4500 and a 6000 for a typical 2U server should compare against. The host platform differs for the 6000, but at these GPU-bound serving loads, the CPU and PCIe topology move the result by far less than the power cap does. We used vLLM’s online serving benchmark at concurrency levels of 1, 2, 4, 8, 16, and 32 with two request shapes: 512 input tokens and 512 output tokens, which approximates chat and short agent turns, and 8,192 input tokens with 1,024 output, which approximates RAG and document work where prefill dominates and KV cache fills. The models were chosen to fit the smallest card in the group where possible, so the L4’s 24GB sets the floor: Ministral 3 3B Reasoning and Ministral 3 8B Reasoning in BF16, Ministral 3 14B Instruct in FP8, and Qwen3.8-27B in NVFP4, which fits the 32GB cards and up but not the L4. We report output token throughput at each concurrency level and cite total throughput, including prefill tokens, where it changes the picture. What is not here: larger models that only the 48GB and larger cards can hold, MIG partitioning, vGPU, video transcode, metered power, the two-card 4500 configuration, and the datacenter cards above this class. Maximum Achievable Matmul FLOPS (MAMF) MAMF is a sustained dense GEMM test that reports the best matmul throughput a card reaches in each precision, making it the easiest way to check whether the FP4 hardware works once the serving stack is out of the picture. The 4500 reached 135.9 TFLOPS in BF16, 280.1 in FP8, and 568.6 in NVFP4, with each step down in precision close to doubling throughput. Against NVIDIA’s dense peaks of 202, 406, and 811 TFLOPS, that is 67% to 70% of peak at every precision, and it lines up with NVIDIA’s own cuBLAS figures for the card of 133, 275, and 534. The RTX PRO 6000 scaled similarly, reaching 398.9, 773.9, and 1,480.9 TFLOPS at 600W, 2.6x to 2.9x the 4500 across the three precisions, which matches what the die size and power budget predict. At the 450W cap, it reached 368.6, 772.4, and 1,413.1: FP8 did not move, BF16 lost 8 percent, and NVFP4 lost 5 percent, which says the 6000’s FP8 tensor path sits under 450W on this test, and the other two precisions only brush the ceiling. The L40S reached 220.0 and 368.6 TFLOPS in BF16 and FP8, and the L4 reached 50.1 and 115.8; neither Ada card has FP4 Tensor hardware. The useful comparison is the 4500’s NVFP4 result against everyone else’s FP8: 568.6 TFLOPS is 1.5x the L40S’s FP8 ceiling and 74% of the capped 6000’s, from a 165W card. That is the FP4 advantage, and it exists at the GEMM level. The vLLM serving results that follow show that it has not reached the serving stack in this build, and the MAMF result suggests the serving numbers will continue to grow as the NVFP4 kernels catch up. vLLM Online Serving: Short Context (512/512) Short context is where the 4500 makes its case: on the three FP8 and BF16 models, its output throughput at concurrency 32 was 1,107 tokens/s on Ministral 3 14B against 995 for the L40S, 1,017 against 1,005 on Ministral 3 8B, and 2,006 against 1,988 on Ministral 3 3B. Every one of those is a Blackwell card at 165W matching or beating an Ada card at 350W, and the curves sit on top of each other from concurrency 1 through 32, so this is not a peak-only result. The L4 landed at 364, 279, and 704 on the same models, which puts the 4500 at 2.8x to 3.6x its predecessor. The RTX PRO 6000 shows what the rest of the Blackwell budget delivers. Capped at 450W, it reached 2,161, 2,042, and 4,069 tokens/s on the same three models, 1.95x to 2.03x the 4500; at its full 600W, it reached 2,283, 2,139, and 4,155, which is 2.06x to 2.1x. The two 6000 curves sit on top of each other through concurrency 16 on every model and separate by 2 to 5 percent only at 32, so at short context the last 150W adds little. Against the 4500, that is double the throughput from a card with 3x the memory and, in the R770, 2.7x the slot power. Total throughput, which counts the 512 prefill tokens per request as well, follows the same order: 2,215 tokens/s for the 4500 on the 14B model against 1,990 for the L40S and 727 for the L4, with the 6000 at 4,323 capped and 4,566 at 600W. The gap to the L4 is widest on the 8B model at 3.6x and narrowest on the 3B model at 2.8x, and the gap to the L40S never exceeds 12 percent in the 4500’s favor. vLLM Online Serving: Long Context (8K/1K) At 8,192 input tokens, the picture changes, and the reason is visible in the shape of the curves. On Ministral 3 14B FP8, the 4500 tracks the L40S through concurrency 8, then flattens at about 180 tokens/s, while the L40S continues to 230 at concurrency 16 and holds there. The 8B model does the same: 185 tokens/s for the 4500 at 32 against 277 for the L40S, with the 4500 plateauing at 8 and the L40S at 16. On the 3B model, both cards plateau at 16, 384 against 475. The reason is KV cache capacity: every concurrent 8K request needs its context resident on the card, and once the weights are loaded, a 32GB frame buffer holds fewer of them than a 48GB one. When vLLM runs out of KV space, it stops adding useful concurrency, and the throughput line goes flat. The 4500 gives up 17 to 33 percent to the L40S at concurrency 32 on these models for that reason alone; the compute side is not the bottleneck, as the short-context results show. The L4, with 24GB and 300GB/s, is flat from concurrency 4 on every model, at 34 to 130 tokens/s. The 6000’s 96GB never runs short in this test, and its curves keep climbing to 32 on every model. At the 450W cap, it reached 552, 589, and 973 tokens/s, 2.5x to 3.2x the 4500; at 600W, it reached 608, 639, and 1,033, or 2.7x to 3.5x. This is the one place the power cap shows up: long-context prefill keeps the 6000 compute-bound longer, so the capped card gives up 4 to 9 percent at concurrency 16 and 32, versus 2 to 5 at short context. Either way, the ratio to the 4500 widens from 2x to roughly 3x, and memory is the reason more than power. If your workload is document Q&A, RAG over long contexts, or agents that carry large tool outputs, memory is the predominant spec, and the 4500’s 32GB is a concurrency constraint. If it’s chat, classification, or short agent turns, it’s not. Total throughput at 8K/1K is dominated by prefill and repeats the same ordering: 1,713 tokens/s for the 4500 on the 14B model against 2,070 for the L40S and 381 for the L4, with the 6000 at 4,971 capped and 5,474 at 600W. NVFP4: Where the Software Sets the Ceiling Qwen3.8-27B in NVFP4 is the one model in the group that uses the FP4 Tensor hardware Blackwell leads with, and the result is the flattest set of curves in the review. At concurrency 32 in the short-context test, the 4500 produced 320 output tokens/s; the L40S, 480, and RTX PRO 6000 hit 621, capped, and 616 at 600W. At concurrency 1, every card landed between 17 and 22 tokens/s on a 27B model, including the 6000 with 1.6TB/s of GDDR7, where the same card delivers 99 tokens/s single-stream on the 14B FP8 model. The L40S has no FP4 Tensor hardware, so vLLM runs an NVFP4 checkpoint by dequantizing the weights and computing in higher precision, a memory-bandwidth-bound path, and it beats the 4500 by 1.5x. The 6000, which has the same FP4 hardware as the 4500 and 2x its bandwidth, lands at the same roughly 2x over the 4500 that separates the two cards everywhere else, and its 450W and 600W runs are within 1 percent of each other at every concurrency, so the kernel is not power-bound either. The pattern across all four series, with single-stream throughput that does not move with memory bandwidth and a concurrency-32 ceiling of 480 to 621 regardless of the GPU underneath, points to the NVFP4 kernel path in this vLLM build as the limit on every card. At 8K/1K, the 4500 reached 153 tokens/s at concurrency 32, compared with 190 for the L40S and 402 and 425 for the 6000 capped and uncapped. Those are the same KV-cache-limited shapes the FP8 and BF16 models showed, with the 6000’s 96GB again being the reason it keeps climbing while the others flatten. NVIDIA has FP4 on this card at 1,221 tokens/s on Llama 3.1 8B at 2K/256 and 905 tokens/s on Nemotron 3 Nano 30B A3B under TensorRT-LLM, so the hardware delivers under NVIDIA’s stack. What the data says today is that NVFP4 on this card is a vLLM software story shared by every GPU we tested, and buyers counting on FP4 throughput should test their serving stack before they count on it. Throughput per Rated Watt We did not meter power in this round, so the efficiency math available this time is against each card’s tested power limit. On Ministral 3 14B FP8 at concurrency 32, the 4500 delivers 6.7 output tokens/s per rated watt, against 2.8 for the L40S, 5.1 for the L4, 4.8 for the RTX PRO 6000 at its 450W cap, and 3.8 for the 6000 at 600W. On the 3B model, the ratio holds: 12.2 for the 4500, 5.7 for the L40S, 9.8 for the L4, 9.0 for the capped 6000, and 6.9 for the 6000 at 600W. The 4500 is the most efficient card in the group by this measure at short context, and the capped 6000 is the only one that gets close. Pulling the 6000 down to 450W costs it 2 to 5 percent of throughput and returns 25 percent of its power budget, which is why its efficiency figure jumps by a quarter. The L4 stays competitive on this metric because 72W is a very small denominator, but it needs three times as many slots and servers to reach the same throughput, which is the trade the 4500 was built to end. At 8K/1K, the order changes: the 4500 lands at 1.2 tokens/s per rated watt on the 14B model, the capped 6000 at 1.2, and the 600W 6000 at 1.0, with the L40S at 0.7. Once the 4500 runs out of KV cache, its throughput stops growing while its power rating does not, so at long context the 6000 at 450W matches it on efficiency and delivers 2.9x the work per slot. Conclusion The RTX PRO 4500 Blackwell Server Edition does exactly what NVIDIA says it does: it puts Blackwell into a slot, a power budget, and a chassis class that previously topped out at an L4. On the models most mainstream servers will run, small and mid-size instruct, and reasoning models at chat-length context, it delivers L40S throughput from a single-slot card rated at less than half the L40S’s power, and three times what the L4 could manage. There are, of course, limits. Starting with memory: 32GB is enough weight capacity for the models this card targets, but it caps concurrency at long context, where the L40S’s 48GB and especially the 6000’s 96GB pull away, and where the 6000 runs roughly 3x the 4500 at the same request shape even with its power held to what the same server would allow. NVFP4, the headline Blackwell precision, performs in line with cards that have no FP4 hardware on this vLLM build, even though the MAMF sweep shows the FP4 hardware delivering 2x the card’s FP8 rate at the GEMM level, so the FP4 advantage is one to verify on your own stack, but there’s clearly more upside coming via software. Anyone with 2U servers, or a 1U chassis that takes an FHFL card, that needs to add inference, vision AI, transcoding, or virtual workstation capacity without touching power or cooling will find the RTX PRO 4500 to be an excellent fit. Further, anyone still running L4s who has outgrown 24GB should be enthusiastic about the 4500. For organizations whose workloads are long-context or multi-tenant enough to need 48GB or more per card, the RTX PRO 6000 is the better option. NVIDIA has done well to designate two clear swim lanes for these cards, and the 4500 fills the edge AI gap between the L4 and the 6000 while drawing just 165W. Leaderboard: The RTX PRO 4500 Blackwell Server Edition holds Best Single-Slot GPU on our Best Enterprise GPUs leaderboard, where the L40S and RTX PRO 6000 data from this review also appear.
The NVIDIA RTX PRO 4500 Blackwell Server Edition is the card NVIDIA built for the servers most enterprises already own. It is a single-slot, passively cooled, 165W board with 32GB of GDDR7 and the full Blackwell feature set: fifth-generation Tensor Cores with FP4, fourth-generation RT cores, three encoders and three decoders, and Multi-Instance GPU support that can split it into two isolated 16GB partitions. It replaces the L4 as the entry point to NVIDIA’s RTX PRO server lineup, and it fits in a 2U PowerEdge, ProLiant, or ThinkSystem without a power or cooling concern. NVIDIA’s pitch for the card is right-sizing: most enterprises are not building AI factories; they run air-cooled, brownfield data centers with hard limits on rack power, and they need a first accelerated computing step that does not require a new server class. Or they’re working at the edge, where the RTX PRO 4500 is better aligned with the workloads and power constraints that typically determine what can be deployed. The RTX PRO 6000 Blackwell covers the top of that market with 96GB and up to 600W. The 4500 keeps the same architecture and cuts everything else roughly in half, so customers with smaller models, many smaller tasks, or a per-dollar-per-watt scale-out budget get Blackwell without paying for headroom they won’t use. The workloads for this class of GPU are well understood: small language model inference, vision AI and video analytics, vector search and embedding generation, transcoding, and virtual workstations under vGPU 20. It claims over 5x the L4 for small and mid-size model inference and up to 10x on Nemotron models, and its launch partners span NTT DATA for edge video, Pacific Biosciences for sequencing, and Nokia for AI-RAN base stations. AWS built its new G7 instances on it as a replacement for the G6. We already tested the card against the L4 inside an HPE ProLiant DL145 Gen11 in our RTX PRO 4500 edge review, where it delivered 2.7x the L4’s throughput. This review answers how close it gets to the 350W, dual-slot L40S that has been the default mainstream inference card for two years, and what the RTX PRO 6000 above it adds in extra memory and power. To find out, we put the 4500 in a Dell PowerEdge R770 against an L40S and L4, added the RTX PRO 6000 Blackwell data from our Dell Pro Max Tower T2 at both its 600W default and a 450W cap matching what the R770 allows a GPU slot, and ran vLLM online serving across four models at two context lengths. As a brief preview, at typical chat and agent context lengths, the 4500 keeps pace with the L40S on FP8 and BF16 models at less than half the power; at 8K input, the 32GB frame buffer shows its limits; and NVFP4, the precision NVIDIA leads its Blackwell pitch with, performs in line with the rest of the lineup on this vLLM build. Key Takeaways - L40S throughput at 165W in one slot: At 512-token input and output, the 4500 matched or edged the L40S on every FP8 and BF16 model we ran, 1,107 vs. 995 tokens/s on Ministral 3 14B FP8 and 2,006 vs. 1,988 on Ministral 3 3B, from a card rated for 47 percent of the L40S’s power and half its slots. - Roughly 3x the L4 it replaces: 2.8x to 3.6x the L4’s output throughput at short context, at 2.3x the L4’s board power. - 32GB has a ceiling: At 8K input the 4500 trailed the L40S by 17 to 33 percent and its curves flattened at concurrency 8 where the L40S kept climbing to 16. The flattening is KV cache running out, and 32GB sets that limit. - NVFP4 is a software story, on every card: On Qwen3.8-27B NVFP4 the 4500 produced 320 tokens/s at concurrency 32, the L40S 480 with no FP4 hardware at all, and the RTX PRO 6000 621. The NVFP4 path in this vLLM build does not scale with memory bandwidth on any of them, while the MAMF sweep shows the 4500’s FP4 hardware reaching 568.6 TFLOPS against 280.1 in FP8. - What the bigger card buys you: Capped to 450W, the ceiling a Dell R770 GPU slot allows, the RTX PRO 6000 ran about 2x the 4500 at short context and 2.5x to 3.2x at 8K. Its full 600W rating added 1 to 2 percent at short context and 4 to 9 percent at long context on top of that, so most of what the 6000 delivers over the 4500 is memory. NVIDIA RTX PRO 4500 Blackwell Server Edition Specifications Two cards carry the RTX PRO 4500 Blackwell name, and they’re similar but not the same product. The Workstation Edition, announced at GTC 2025, is a 200W, dual-slot, actively cooled card with four DisplayPort outputs and 896GB/s of memory bandwidth. The Server Edition we are reviewing here arrived at GTC 2026 as a 165W, single-slot, passive, headless board at 800GB/s, and it trades the display outputs and some clock headroom for the data center feature set: MIG, confidential computing, a third NVENC and NVDEC, and vGPU 20 support. Underneath, they have the same 10,496-core GPU and 32GB of GDDR7 with ECC. NVIDIA’s data put the die at 45.6 billion transistors on 378 mm² of TSMC 4nm custom silicon with 82 fourth-generation RT cores, which works out to 82 SMs and 328 Tensor Cores. Those are GB203 numbers, the same die family as the RTX 5080, lightly cut. The Tensor figures below are NVIDIA’s peaks with sparsity; dense is half, which NVIDIA’s own peakTOPS binary confirms at 805 TFLOPS FP8 and 403 TFLOPS FP16 measured on this card. Design and Build The 4500 Server Edition is a full-height, full-length board, 4.4 inches tall and 10.5 inches long, and it fills one slot. That is an important change from the L4, which was a half-height, half-length card. The L4 is still the comparison it benchmarks against, but the form factor moved to FHFL to make room for the memory, the 16-pin power input, and a heatsink that can shed 165W with chassis airflow alone. That form factor is the one real break from the L4, which fit any HHHL slot and drew its 72W from the slot, so it went into 1U servers and small edge boxes. The 4500 is full height, full length, and needs the 16-pin cable, and the 1U mainstream servers we work with, Dell’s current 1U PowerEdge line among them, top out at HHHL or FHHL cards without that connector. In practice, this is a 2U card, or one for the few whitebox 1U chassis built around FHFL slots, so a shop that bought L4s for 1U or edge boxes does not get a drop-in replacement. Stacked edge-on with the L4 above it and the L40S below, the 4500 needs one slot where the L40S needs two, which is why a 2U server that takes four double-wide cards can take eight of these, and how AWS’s G7 instances run up to eight of them per host. Power is delivered through a single PCIe CEM5 16-pin connector, the same connector the RTX PRO 6000 uses, with no auxiliary 8-pin option. Cooling is entirely passive, so the card depends on the server’s fans and will throttle in a chassis that cannot move air through it, which is normal for the class, and all other GPUs have the same requirement. The card has no display outputs, so graphics work, whether vGPU virtual workstations or rendering, goes out over the network, and the bracket is vented. The three ninth-generation NVENC and three sixth-generation NVDEC engines add 4:2:2 H.264 and HEVC encode and decode over the L4’s generation, and NVIDIA quotes 2x the H.264 decode throughput per NVDEC engine versus the L4. The bracket shot below makes the split visible: the L4 and the 4500 carry vents and nothing else, while the L40S and the RTX PRO 6000 Workstation Edition each carry four DisplayPort outputs. Where the RTX PRO 4500 Blackwell Server Edition Fits NVIDIA’s framing is that the card’s “central value is the combination of AI, compute, graphics, and video acceleration in a 165 W, passive, single-slot design,” built for servers “where rack density, power, and cooling are primary deployment constraints.” It sits in what NVIDIA calls the mainstream data center and edge segment of the RTX PRO server portfolio, below the RTX PRO 6000 Blackwell Server Edition and above nothing; it’s the entry point of the Blackwell server line. Against the L4, NVIDIA cites up to 10x on small model inference with its Nemotron Nano models (NVFP4 on the 4500 against FP8 on the L4), up to 4x on video summarization with Cosmos Reason 2, 1.9x on SPECviewperf 15 at 4K, and 1.8x streaming density. Against CPUs, it cites up to 50x on vector database work with cuVS CAGRA in Milvus and up to 100x on vision AI, both measured on eight-card 4500 servers against dual-socket EPYC 9654 systems. Dell’s server team told us it sees the 4500 in mainstream racks, since three 32GB 4500s cost near the same as one 96GB 6000, and at that point the bigger card usually wins. This is a card for a server that would otherwise have no GPU, or an L4, and is a suboptimal fit for a server that was designed to get one or more 6000s. Test Setup We installed the 4500, the L40S, and the L4 in our Dell PowerEdge R770, configured with two Intel Xeon 6787P processors (86 cores each) and 2TB of DDR5-6400 across 32 DIMMs, running NVIDIA driver 595.71.05. The test runs used the NGC vLLM 26.08-py3 container, which carries vLLM 0.27.1 on CUDA 13.4.1, so the software stack is identical across cards. Each card was tested individually at its default power limit, and the R770 supplies each card its full rating: Dell lists the L4 at 72W, the 4500 at 165W, and the L40S at 350W in that chassis. The RTX PRO 6000 Blackwell is the workstation card in our Dell Pro Max Tower T2, on the same container and driver. NVIDIA rates the workstation and server editions identically at 600W, and they profile the same in our testing, so it stands in for the server card. Incidentally, we do have a review of the RTX PRO 6000 Server cards in an XE7740 if you want to dig into how these cards can scale. We ran the 6000 twice: once at its 600W default, and once with a verified 450W cap applied through nvidia-smi, because 450W is the most the R770 will deliver to a 6000 in a GPU slot, and it is the figure a buyer choosing between a 4500 and a 6000 for a typical 2U server should compare against. The host platform differs for the 6000, but at these GPU-bound serving loads, the CPU and PCIe topology move the result by far less than the power cap does. We used vLLM’s online serving benchmark at concurrency levels of 1, 2, 4, 8, 16, and 32 with two request shapes: 512 input tokens and 512 output tokens, which approximates chat and short agent turns, and 8,192 input tokens with 1,024 output, which approximates RAG and document work where prefill dominates and KV cache fills. The models were chosen to fit the smallest card in the group where possible, so the L4’s 24GB sets the floor: Ministral 3 3B Reasoning and Ministral 3 8B Reasoning in BF16, Ministral 3 14B Instruct in FP8, and Qwen3.8-27B in NVFP4, which fits the 32GB cards and up but not the L4. We report output token throughput at each concurrency level and cite total throughput, including prefill tokens, where it changes the picture. What is not here: larger models that only the 48GB and larger cards can hold, MIG partitioning, vGPU, video transcode, metered power, the two-card 4500 configuration, and the datacenter cards above this class. Maximum Achievable Matmul FLOPS (MAMF) MAMF is a sustained dense GEMM test that reports the best matmul throughput a card reaches in each precision, making it the easiest way to check whether the FP4 hardware works once the serving stack is out of the picture. The 4500 reached 135.9 TFLOPS in BF16, 280.1 in FP8, and 568.6 in NVFP4, with each step down in precision close to doubling throughput. Against NVIDIA’s dense peaks of 202, 406, and 811 TFLOPS, that is 67% to 70% of peak at every precision, and it lines up with NVIDIA’s own cuBLAS figures for the card of 133, 275, and 534. The RTX PRO 6000 scaled similarly, reaching 398.9, 773.9, and 1,480.9 TFLOPS at 600W, 2.6x to 2.9x the 4500 across the three precisions, which matches what the die size and power budget predict. At the 450W cap, it reached 368.6, 772.4, and 1,413.1: FP8 did not move, BF16 lost 8 percent, and NVFP4 lost 5 percent, which says the 6000’s FP8 te
document fetched 2026-10-24 09:58 · canonical https://storagereview.com/review/nvidia-rtx-pro-4500-blackwell-server-edition-review