Why you can trust Tom’s Hardware
Our expert reviewers spend hours testing and comparing products and services so you can choose the best for you. Find out more about how we test.
Introducing the Nvidia GeForce RTX 5070 Founders Edition
The Nvidia GeForce RTX 5070 Founders Edition has a big hole to fill in the graphics card market. As the first true mainstream offering for the Blackwell RTX 50-series GPUs, it takes over from the discounted RTX 4070 Founders Edition with the same nominal $549 base MSRP. It also has the same 12GB of VRAM and nearly the same number of streaming multiprocessors (SMs) — 48 versus 46 — but with the new Blackwell features. On paper, getting a faster GPU for less money with new features should make this one of the best graphics cards, but we have some concerns.
The biggest problem will no doubt be retail availability and pricing, and we’ve seen every GPU launch of the past few months sell out almost instantly. From Intel’s $249 Arc B580 to the $1,999 RTX 5090, with the RTX 5080 and RTX 5070 Ti filling in the middle, MSRPs have been effectively non-existent. We don’t expect the 5070 to buck that trend, and it’s all starting to feel a lot like 2021 — just with AI-induced GPU shortages rather than cryptocurrency mining shortages. When will it end? That’s a difficult question to answer.
Nvidia posted record earning of $130 billion for the 2025 fiscal year that just ended, more than double its 2024 earnings. Nearly all of the gains came from its AI and data center business, which accounted for 88% of gross revenue. Gaming was a very distant second place at just 8.7% of the total revenue. Nvidia has been saying it’s no longer primarily a gaming company for a while now, and nowhere is that more apparent than in the financials.
With massive demand coming from the AI sector, and with limited 5nm-class wafers from TSMC, the simple economics show that it’s far more profitable to make data center and AI products right now rather than consumer GPUs. It’s not that Nvidia won’t order any consumer GPUs, but it’s unlikely to be anywhere near sufficient to meet the demand. And in fact, right now virtually every graphics card of the past two years is either sold out or severely overpriced relative to the launch MSRP — with the only exceptions being the RTX 4060, AMD’s RX 7600 (the RX 7600 XT currently starts at $430, $100 more than its original MSRP), and Intel’s Arc B570.
The prospects for reasonably priced GPUs look grim, in other words. It could be many months before anything gets close to MSRP — and that goes for AMD’s RX 9070 XT and RX 9070 that are slated for review tomorrow. We expect those to be just as hard to acquire at MSRP as the RTX 5070, which will officially go on sale tomorrow. But maybe our pessimism will prove misplaced! For now, all we can do is look at the performance and features on tap, and hope that supply will catch up to demand sooner rather than later.
We’ve been kept busy during the past two months testing and retesting graphics cards. The fourth Nvidia GPU launch of the year and sixth new graphics card since December hasn’t given us time to catch our collective breath, never mind getting all the other prior generation GPUs we’d like to test filed through our new test suite.
Last month we also took a closer look at DLSS 4 and MFG, using the 5080 and 5090, which will have to suffice for now — time constraints didn’t allow us to cover the same tests on the RTX 5070 Ti or the 5070, or the 9070 XT and 9070 for that matter. But we’ll get around to those hopefully by next week and update the appropriate review pages.
Until then, the TLDR remains the same: MFG is a great way to inflate benchmark scores, and in the right scenarios it can feel better than framegen or non-framegen even if it has slightly higher input latencies. But the benchmark numbers tend to be much higher compared to how games actually feel. It’s not bad as such, but subjectively MFG4X might feel more like 30~40 percent faster than the non-MFG performance, rather than the 200% improvement benchmarks can show. It will look smoother even while typically delivering the same or lower levels of responsiveness.
For additional information about Nvidia’s Blackwell RTX GPUs, check the links in the boxout. The RTX 5070 Founders Edition represents the reference clocks and design from Nvidia, which will likely be just as fast as most of the non-reference card models from AIB partners. It might also be slightly more affordable, assuming you can find any in stock. But as usual, let’s start with the specs table to see how it compares to the prior generation.
| Graphics Card | RTX 5070 | RTX 4070 | RTX 3070 | RX 9070 |
|---|---|---|---|---|
| Architecture | GB205 | AD104 | GA104 | Navi 48 |
| Process Technology | TSMC 4N | TSMC 4N | Samsung 8N | TSMC N4P |
| Transistors (Billion) | 31 | 32 | 17.4 | 53.9 |
| Die size (mm^2) | 263 | 294.5 | 392.5 | 356.5 |
| SMs / CUs | 48 | 46 | 46 | 56 |
| GPU Shaders (ALUs) | 6144 | 5888 | 5888 | 3584 |
| Tensor / AI Cores | 192 | 184 | 184 | 112 |
| Ray Tracing Cores | 48 | 46 | 46 | 56 |
| Boost Clock (MHz) | 2512 | 2475 | 1725 | 2520 |
| VRAM Speed (Gbps) | 28 | 21 | 14 | 20 |
| VRAM (GB) | 12 | 12 | 8 | 16 |
| VRAM Bus Width | 192 | 192 | 256 | 256 |
| L2 / Infinity Cache | 48 | 36 | 4 | 64 |
| Render Output Units | 80 | 64 | 96 | 128 |
| Texture Mapping Units | 192 | 184 | 184 | 224 |
| TFLOPS FP32 (Boost) | 30.9 | 29.1 | 20.3 | 36.1 |
| TFLOPS FP16 (FP4/FP8/INT4 TOPS) | 247 (988) | 233 (466) | 163 | 289 (1156) |
| Bandwidth (GB/s) | 672 | 504 | 448 | 640 |
| TBP (watts) | 250 | 200 | 220 | 220 |
| Launch Date | Feb 2025 | Apr 2023 | Oct 2020 | Mar 2025 |
| Launch Price | $549 | $599 | $499 | $549 |
The paper specifications don’t necessarily tell the full story. For example, the Blackwell architectgure doubles the ray/triangle intersections per clock for the RT cores, the tensor cores support new number formats like FP4, and the CUDA cores all support FP32 and INT32 operations (only half of the CUDA cores in the RTX 40- and 30-series GPUs supported INT32 operations). That leads to what might appear at first to be little to no change in performance potential.
RTX 5070 has peak theoretical FP32 compute of 30.9 TeraFLOPS, compared to 29.1 TeraFLOPS on the RTX 4070 — a mere 6.2% increase. TGP (Total Graphics Power) has increased from 200W to 250W, however, along with memory getting a sizeable 33% bump in bandwidth thanks to the move to GDDR7 memory. So in theory, the 5070 should be somewhere between 6% and 33% faster than its direct predecessor for ‘normal’ workloads (i.e. things that don’t leveral the FP4 support or MFG). In practice, the gains are on the higher end of that range for most games.
Die size and transistor counts are interesting as well, mostly because the previous generation AD104 GPU was used in the RTX 4070 Ti and had up to 60 SMs available, even though only 46 were enabled in the 4070. The GB205 die only has up to 50 SMs, however, with 48 enabled in the RTX 5070. That’s what makes the new chip smaller and also gives it fewer transistors — both chips are made on the same TSMC 4N node.
AI compute does potentially favor the RTX 5070 a lot, but only if we include FP4 support. It has up to 988 TFLOPS of FP4 compute (which Nvidia classifies as “TOPS” even though that’s normally only used for integer calculations), more than double the 4070’s 466 TFLOPS of FP8. But for FP8 compute, it’s the same 6.2% difference as the graphics FP32 compute. Clock speeds on paper are only slightly higher with the 5070 compared to the 4070, but we’ll need to look at real-world clocks as Blackwell and Ada GPUs tend to run at much higher clocks than the stated boost clocks.
The RTX 5070 offers a much bigger improvement over the older RTX 3070, naturally, with about 50% more theoretical compute and up to 6X more AI compute (comparing FP4 to FP16, with sparsity in both cases). But AMD’s upcoming RX 9070, which we’ll review tomorrow, looks set to deliver some serious competition. Check back in 24 hours and we’ll have the full review for AMD’s 9070 and 9070 XT.
Again, before we even get to the benchmarks, there are a couple of elephants in the corner.
First is retail availability and pricing. We have every reason to expect the RTX 5070 cards will sell out quickly tomorrow when they go on sale, and that many models will end up at significantly higher prices than the ostensible $549 MSRP. After all, the cheapest graphics cards are pretty much stupidly expensive — and that goes for used cards on places like eBay as well, where the RTX 4070 price in the past 30 days has averaged over $650. Will a card that’s newer, faster, and has more features cost less than the previous generation? Not a chance.
The other item to remember is the impending AMD Radeon RX 9070 and RX 9070 XT launch, which will be one day after the RTX 5070 — meaning, MSRP-priced reviews go up tomorrow, and the cards go on sale starting March 6. The RX 9070 competes directly with the RTX 5070 on price, or at least MSRP. Traditionally, AMD GPUs also don’t command quite as much demand as Nvidia GPUs. But the RX 9070 XT for $50 more looks like it will potentially compete with the RTX 5070 Ti, or alternatively it should easily beat the RTX 5070 for a relatively minor price increase.
But AMD GPU availability right now isn’t any better than Nvidia GPUs. Everything from the RX 7600 XT and above is horribly overpriced, and the previous generation RX 7900 GRE that was intended to compete with the RTX 5070 at the $549 price point now sells for over $900, with the average eBay price for used GPUs over the past 30 days sitting at $711. Newer, faster, and better RX 9070-class GPUs will inevitably sell out and end up going for much more than $549 or $599.
